[ { "_derived": { "affects_count": 6, "families_affected": { "agents-behavior": 2, "control-automation": 1, "information-legibility": 1, "logistics-networks": 2 }, "incoming": { "applied_in": [ "generated-recipe-graphs", "mechanic-discovery-search", "specification-programming-games" ] } }, "affects": [ { "effect": "accelerates", "id": "agent-pathfinding", "note": "Path cost heuristics or search-ordering functions can be evolved against a benchmark of the game's own maps for fewer node expansions at equal route quality, as FunSearch evolved priority functions." }, { "effect": "tunes", "id": "hauling-jobs", "note": "Job-to-hauler assignment is a scheduling problem like the Borg case; the dispatch heuristic can be evolved against simulated throughput and idle time on recorded colony states." }, { "effect": "tunes", "id": "vehicle-route-scheduling", "note": "Automatic load and stop-ordering heuristics resemble the bin-packing heuristics FunSearch improved, and can be evolved against delivery-rate benchmarks." }, { "effect": "tunes", "id": "utility-action-selection", "note": "Utility scoring functions can be evolved as code against designer metrics (survival, need satisfaction, behavior variety), keeping the result readable unlike a neural policy." }, { "effect": "tunes", "id": "comparative-histograms", "note": "Evolving solutions to each puzzle before launch estimates near-optimal cost, cycle or area scores, anchoring histogram axes and flagging degenerate puzzles." }, { "effect": "enables-new", "id": "player-programming", "note": "Rivals or coaches could evolve programs in the game's own language against its evaluator, so player-written code competes with machine-evolved code under the same rules." } ], "aliases": [ "AlphaEvolve", "FunSearch", "OpenEvolve", "ShinkaEvolve", "evolutionary coding agents" ], "category": "math-discovery", "deployed_in": [], "determinism_impact": "preserves", "horizon": "1-2y", "id": "alphaevolve-algorithm-discovery", "key_results": [ { "confidence": "high", "locator": "Extremal combinatorics, Cap sets", "quote": "in dimension n = 8, FunSearch found a larger cap set than what was previously known", "sources": [ "src-romeraparedes-2023-funsearch" ], "text": "FunSearch found a cap set of size 512 in dimension 8 and online bin-packing heuristics that beat first fit and best fit on benchmark distributions." }, { "confidence": "high", "locator": "Improving data center scheduling", "quote": "This solution, now in production for over a year, continuously recovers, on average, 0.7% of Google’s worldwide compute resources", "sources": [ "src-deepmind-2025-alphaevolve-blog", "src-novikov-2025-alphaevolve" ], "text": "AlphaEvolve found a 4x4 complex matrix multiplication with 48 scalar multiplications, matched about 75% and improved about 20% of over 50 open math problems, and produced a Borg scheduling heuristic that has recovered about 0.7% of Google's fleet compute in production for over a year." }, { "confidence": "high", "locator": "Abstract", "quote": "The system rediscovered the best known solutions in most of the cases and discovered improved solutions in several", "sources": [ "src-georgiev-2025-math-discovery-at-scale" ], "text": "Applied by Georgiev, Gomez-Serrano, Tao and Wagner to 67 problems, AlphaEvolve mostly rediscovered the best known constructions, improved several, and sometimes generalized finite results into formulas." }, { "confidence": "medium", "locator": "Abstract", "quote": "ShinkaEvolve discovers a new state-of-the-art circle packing solution using only 150 samples", "sources": [ "src-lange-2025-shinkaevolve", "src-sharma-2025-openevolve" ], "text": "Open-source successors lowered the barrier: ShinkaEvolve reported a new state-of-the-art circle packing with about 150 samples, and OpenEvolve runs the loop with any OpenAI-compatible or local model under Apache-2.0." } ], "limitations": [ "Requires a fast, automatic, trustworthy evaluator; anything judged by taste or fun has no such evaluator, and evolved code overfits whatever the evaluator measures.", "Each run consumes thousands of LLM samples and program executions in earlier systems; ShinkaEvolve's roughly 150-sample result shows efficiency varies widely by problem.", "Evolved heuristics are tuned to the benchmark instances and must be re-validated on new maps, content or patches.", "No game studio use was found in this run's sources; all cited results are in mathematics, scheduling and systems code." ], "maturity": "research", "mechanism": "A human supplies a program skeleton, the part to evolve (one priority function in FunSearch; arbitrary blocks across a file in AlphaEvolve), and an evaluator that runs a candidate and returns scores. A programs database holds evaluated programs; samplers build prompts from high-scoring and diverse parents and ask an LLM for a modified program; evaluators execute and score it; results flow back into the database. Diversity is kept with island populations (FunSearch) or MAP-Elites grids (OpenEvolve). Samplers and evaluators run asynchronously in parallel. The output is readable source code, so a discovered heuristic can be inspected, tested and shipped like hand-written code.", "name": "LLM-guided evolutionary program search", "provenance": { "created": "2026-10-24-techniques-sim-math", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "evaluation", "compute-cost" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-romeraparedes-2023-funsearch", "src-deepmind-2025-alphaevolve-blog", "src-novikov-2025-alphaevolve", "src-georgiev-2025-math-discovery-at-scale", "src-lange-2025-shinkaevolve", "src-sharma-2025-openevolve" ], "status": "sourced", "summary": "An LLM proposes edits to code, an automated evaluator scores each candidate, and an evolutionary database keeps the best and most diverse programs, finding heuristics and constructions that beat human-designed ones.", "trl": 3, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 6, "families_affected": { "agents-behavior": 1, "conflict-defense": 1, "progression-goals": 3, "risk-failure": 1 }, "incoming": { "applied_in": [ "content-elastic-events", "generated-recipe-graphs", "language-modding", "mechanic-discovery-search", "parallel-autobalance", "qd-scenario-libraries" ] } }, "affects": [ { "effect": "tunes", "id": "adaptive-threat-scaling", "note": "Agent pass rates plus a player-population model estimate how a difficulty ramp drives churn, so threat and difficulty curves can be tuned before release instead of from live telemetry." }, { "effect": "tunes", "id": "scenario-objectives", "note": "Each scenario or level gets a predicted pass rate from human-like agents, letting designers order and adjust goals and constraints against a target difficulty curve." }, { "effect": "tunes", "id": "career-ladder", "note": "Search agents measure how many actions each career path needs to reach each level, exposing careers that are much faster or slower than the rest, as in The Sims Mobile study." }, { "effect": "tunes", "id": "agent-pathfinding", "note": "Curiosity and RL agents explore traversal exhaustively, finding stuck spots, missing collision, and shortcuts that scripted navmesh bots do not reach." }, { "effect": "tunes", "id": "milestone-unlocks", "note": "Agents replaying the early game measure when resource and progression thresholds are reached, so unlock pacing is set from distributions instead of a few manual runs." }, { "effect": "tunes", "id": "stat-combat-resolution", "note": "Agent playthroughs estimate win rates and dominant unit mixes in statistical combat before release." } ], "aliases": [ "RL playtesting", "human-like playtesting", "difficulty prediction agents", "game testing agents" ], "builds_on": [ "generalist-game-agents" ], "category": "automated-game-design", "determinism_impact": "preserves", "horizon": "now", "id": "automated-playtesting", "key_results": [ { "confidence": "high", "locator": "Abstract", "quote": "Compared to Monte Carlo Tree Search (MCTS) we show that this approach increases correlation with average level difficulty, giving more accurate predictions as well as requiring only a fraction of the computation time", "sources": [ "src-gudmundsson-2018-human-like-playtesting" ], "text": "King's CNN trained on player moves in Candy Crush Saga and Candy Crush Soda Saga predicted human moves and level difficulty, correlating better with average difficulty than MCTS at a fraction of the compute (IEEE CIG 2018)." }, { "confidence": "high", "locator": "Abstract", "quote": "deep RL models does not seamlessly transfer from the benchmark environments to target ones without heavily tuning their hyperparameters, leading to linear scaling of the engineering efforts and computational cost with the number of target domains", "sources": [ "src-zhao-2019-winning-isnt-everything" ], "text": "EA used A* and evolution-strategy agents to balance The Sims Mobile careers during development and reported that deep RL needed heavy per-game tuning, so effort grew roughly linearly with the number of target games." }, { "confidence": "high", "locator": "Section IV.A, Game exploits and logical bugs", "quote": "the RL agent quickly finds a way of exploiting the missing collision mesh", "sources": [ "src-bergdahl-2020-augmenting-automated-testing-drl", "src-gordillo-2021-curiosity-playtesting" ], "text": "EA SEED's RL test agents found a wall missing its collision mesh that scripted navmesh agents missed and located all five stuck areas in a test map, and curiosity rewards raised coverage of a complex 3D map." }, { "confidence": "high", "sources": [ "src-gillberg-2023-rl-testing-aaa" ], "text": "EA SEED added an experimental RL system to scripted-bot automated testing for AAA titles including Battlefield 2042 and Dead Space (2023) and reported that integration, not algorithms, consumed the most time." }, { "confidence": "high", "locator": "Abstract", "quote": "The AI players predict level difficulty, which is used to drive a player population model with simulated skill, persistence, and boredom", "sources": [ "src-roohi-2020-difficulty-churn" ], "text": "RL agents combined with a player-population model of skill, persistence, and boredom predicted per-level pass and churn rates for Angry Birds Dream Blast." }, { "confidence": "medium", "locator": "Looking to the Future: The Journey to General Embodied Intelligence", "quote": "We find the agents still face challenges with very long-horizon, complex tasks that require extensive, multi-step reasoning and goal verification. SIMA 2 also has a relatively short memory of its interactions", "sources": [ "src-deepmind-2025-sima-2-blog" ], "text": "SIMA 2 (2025) follows language instructions in commercial 3D games through screen input and virtual keyboard and mouse and transfers skills across games, but still struggles with long-horizon tasks, memory, and precise low-level control." } ], "limitations": [ "Agents are not players: skill and style differ, and imitation agents need large logs of real player moves, which a new game lacks before launch.", "Integration dominates cost: EA SEED found engineering RL into production builds was the largest time sink, partly because development builds differ from the shipped game.", "Deep RL transfers poorly between games without heavy hyperparameter tuning, so cost grows roughly linearly with the number of titles.", "Open-ended simulation games have no single goal to optimize, so coverage or curiosity rewards stand in for design intent and can miss balance problems that only appear over long play." ], "maturity": "shipped-niche", "mechanism": "Agents play the game through an API or the rendered screen, in three families. Human-like agents are trained by supervised imitation of logged player moves, as with King's CNN on Candy Crush data, and their success rate on a new level predicts its difficulty for people. Goal-driven search or RL agents (A*, evolution strategies, deep RL) optimize task completion, and their action counts and success rates measure progression balance, as in EA's career-balancing study for The Sims Mobile. Curiosity agents are rewarded for novel states, maximizing map coverage to find stuck spots, missing collision, and unintended shortcuts. Agent results can feed a model of the player population: Roohi et al. combined RL agent pass rates with simulated skill, persistence, and boredom to predict per-level churn. In production, agents run against new builds inside the existing test pipeline and their coverage and failures are visualized for designers.", "name": "Agent-based automated playtesting and balancing", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-24-technique-links", "2026-10-25-source-claims-pilot" ] }, "risks": [ "evaluation", "compute-cost", "labor" ], "runtime": [ "offline-authoring", "analytics" ], "sources": [ "src-gudmundsson-2018-human-like-playtesting", "src-zhao-2019-winning-isnt-everything", "src-bergdahl-2020-augmenting-automated-testing-drl", "src-gordillo-2021-curiosity-playtesting", "src-gillberg-2023-rl-testing-aaa", "src-roohi-2020-difficulty-churn", "src-deepmind-2025-sima-2-blog" ], "status": "sourced", "summary": "Studios train imitation, reinforcement learning, and search agents to play builds, estimating level difficulty, pass and churn rates, progression pacing, map coverage, and exploits before players reach them.", "trl": 7, "type": "technique", "year": 2018 }, { "_derived": { "affects_count": 5, "families_affected": { "governance-policy": 2, "narrative-generation": 1, "progression-goals": 1, "time-pacing": 1 }, "incoming": { "applied_in": [ "generated-inspection-shifts", "generated-recipe-graphs", "language-modding", "mechanic-discovery-search", "specification-programming-games" ] } }, "affects": [ { "effect": "enables-new", "id": "policy-toggles", "note": "Candidate policies can be written as rule code in the game's own description language and screened by agent playouts for dominance and balance, so a policy list can be grown by search rather than authored one law at a time." }, { "effect": "enables-new", "id": "player-legislation", "note": "Laws players propose could be translated by a code model into enforceable rule code and run through the same compile and playout tests GAVEL uses, rejecting laws that fail to compile or break play." }, { "effect": "tunes", "id": "turn-structure", "note": "Self-play over each generated variant measures first-player advantage, decisiveness, and the share of states with real choices, so turn order, phases, and move allowances can be adjusted by search." }, { "effect": "cuts-authoring-cost", "id": "scenario-objectives", "note": "End conditions are generated together with the rules and kept only if MCTS play reaches them without trivial draws or one-sided wins, replacing hand-tuned victory thresholds." }, { "effect": "deepens", "id": "generative-content", "note": "Generated content extends from text, items, and levels to complete executable rule sets that are compiled and filtered by playability and balance tests before anyone plays them." } ], "aliases": [ "Ludi", "Ludii", "GAVEL", "VGDL generation", "LLM rule generation" ], "builds_on": [ "quality-diversity-search", "monte-carlo-tree-search" ], "category": "automated-game-design", "determinism_impact": "preserves", "horizon": "3-5y", "id": "automated-rule-generation", "key_results": [ { "confidence": "medium", "locator": "Abstract", "quote": "We examine the concept of game quality, its automated measurement through self-play simulations, and its use in the evolutionary search for new high-quality games", "sources": [ "src-browne-2010-evolutionary-game-design" ], "text": "Browne and Maire (IEEE TCIAIG 2010) evolved board game rule sets and ranked them with measures computed from self-play." }, { "confidence": "high", "locator": "Abstract", "quote": "Ludii outperforms one of the most efficient Game Description Language (GDL) reasoners, based on a propositional network, in all games available in the Tiltyard GGP repository", "sources": [ "src-piette-2019-ludii" ], "text": "Ludii describes games with high-level ludemes and outperformed a propositional-network GDL reasoner on all games in the Tiltyard repository." }, { "confidence": "high", "locator": "Abstract", "quote": "including in regions of the potential rules space not covered by existing games in the Ludii dataset", "sources": [ "src-todd-2024-gavel" ], "text": "GAVEL (2024) fine-tuned CodeLlama-13b (about 40 hours on one RTX8000) on 574 Ludii games as a MAP-Elites mutation operator and generated playable games in rule-space regions not covered by existing Ludii games, scored by MCTS self-play metrics; samples are playable on the Ludii portal." }, { "confidence": "medium", "locator": "Abstract", "quote": "Based on video game description language, this paper proposes an LLM-based framework to generate game rules and levels simultaneously", "sources": [ "src-hu-2024-llm-game-generation-vgdl" ], "text": "An LLM framework built on VGDL generated game rules and levels together, extending LLM generation from levels for fixed games to whole games." } ], "limitations": [ "Generated rule sets are small abstract board games; GAVEL trained on 574 games after removing those over 1,024 tokens, far below the size of a simulation game's rule base.", "Self-play metrics such as balance, decisiveness, and agency describe agent play, not whether people enjoy the game, so human evaluation is still required.", "Every candidate needs compilation plus random and MCTS playouts, so the search is offline and compute-bound.", "Language models often emit rule code that fails to compile or is unplayable; GAVEL penalizes such candidates rather than preventing them.", "Simulation games rarely have a formal rule language like Ludii or VGDL, so applying this needs the game's systems exposed as data or a scripting layer the generator can edit." ], "maturity": "prototype", "mechanism": "Rules are written in a game description language. Ludii describes a game as a tree of ludemes (board, pieces, moves, end conditions) that compiles to a playable game; VGDL describes sprites, interactions, and termination for 2D video games. A generator proposes new rule trees: Browne and Maire's system evolved them by crossover and mutation, while GAVEL fine-tunes CodeLlama-13b with a fill-in-the-middle objective on 574 Ludii games and uses it as the mutation operator in MAP-Elites: a span of an existing game's rule code is masked and the model writes a replacement, and the archive is indexed by a 2D PCA projection of 510 Ludii concept features (1,600 cells). Candidates pass hierarchical tests (compiles, is playable, random playouts are not trivially one-sided and offer real choices) and are then scored by MCTS self-play on balance, decisiveness, completion, agency, coverage, and strategic depth. LLM frameworks built on VGDL generate rules and levels together from prompts. The output is a set of verified candidate rule sets for designers to play and select.", "name": "Automated generation of game rules and mechanics", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "evaluation", "controllability", "labor" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-browne-2010-evolutionary-game-design", "src-piette-2019-ludii", "src-todd-2024-gavel", "src-hu-2024-llm-game-generation-vgdl" ], "status": "sourced", "summary": "Systems that write complete game rule sets as code in a description language, using evolution or language models to propose rules and self-play agents to score balance, decisiveness, and strategic depth before a human sees them.", "trl": 5, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 5, "families_affected": { "conflict-defense": 1, "economy-finance": 1, "social-relations": 3 }, "incoming": { "applied_in": [ "agent-ambassadors-async", "natural-language-legislation", "negotiating-faction-ai" ] } }, "affects": [ { "effect": "enables-new", "id": "faction-diplomacy", "note": "Diplomacy becomes free-text negotiation with AI factions that propose specific joint moves, commit to them, and change plans when persuaded, instead of accept or decline treaty menus driven by opinion scores." }, { "effect": "deepens", "id": "natural-language-interaction", "note": "Typed messages are translated into predicted intents that feed a planner, so what the player says changes what the AI does on the map, and the AI's replies are constrained to match its own plan." }, { "effect": "deepens", "id": "reputation", "note": "An AI that models what others expect it to do can also track whether the player's stated intentions matched their moves, turning standing into a learned estimate of trustworthiness rather than a scalar moved by scripted events." }, { "effect": "enables-new", "id": "negotiation", "note": "Planner-grounded dialogue lets AI counterparties propose, counter, and keep terms in free text instead of accepting or rejecting fixed offers." }, { "effect": "deepens", "id": "war-score-peace", "note": "Peace terms could be negotiated in language against the war score, with the planner checking what the opponent will actually accept." } ], "aliases": [ "Cicero", "CICERO", "piKL", "Richelieu" ], "builds_on": [ "imperfect-information-solvers" ], "category": "game-theoretic-solvers", "determinism_impact": "breaks", "horizon": "1-2y", "id": "cicero-strategic-dialogue", "key_results": [ { "confidence": "high", "locator": "Abstract", "quote": "Across 40 games of an anonymous online Diplomacy league, Cicero achieved more than double the average score of the human players and ranked in the top 10% of participants who played more than one game", "sources": [ "src-fair-2022-cicero", "src-meta-2022-cicero-project" ], "text": "Cicero averaged more than double the human score over 40 games in an anonymous online Diplomacy league and ranked in the top 10% of participants who played more than one game (Science, November 2022)." }, { "confidence": "high", "sources": [ "src-meta-2022-cicero-project", "src-fair-2022-cicero" ], "text": "Cicero pairs a controllable dialogue model with a planning engine that models other players' likely moves from the board and the conversation and predicts what others think Cicero will do." }, { "confidence": "high", "locator": "README, License for Model Weights", "quote": "We are releasing model weights under a separate license: CC-BY-NC (version 4.0)", "sources": [ "src-facebookresearch-2022-diplomacy-cicero" ], "text": "The released code is MIT-licensed but model weights are CC-BY-NC 4.0, training scripts target a Slurm cluster (example configuration 256 GPUs), and the repository was archived read-only on 2025-04-17." }, { "confidence": "medium", "locator": "Abstract", "quote": "strategic planning with memory and reflection; 2) goal-oriented negotiation with social reasoning; and 3) augmenting memory through self-play games for self-evolution without human in the loop", "sources": [ "src-guan-2024-richelieu" ], "text": "Richelieu (2024) built an LLM Diplomacy agent combining planning with memory and reflection, goal-directed negotiation, and self-play memory augmentation without human game data." } ], "limitations": [ "Built for one game: the dialogue model and the move policy were trained on human Diplomacy games and messages, so another game needs comparable data or a substitute.", "Released weights are licensed for non-commercial use only (CC-BY-NC 4.0), and the repository is archived, so a studio cannot ship the released model.", "Training assumed a GPU cluster (example 256 GPUs) and inference loads large models per agent, which is costly for a game with several AI factions per player.", "Evaluation used an anonymous league; how players negotiate and whether they trust an opponent they know is an AI was not measured." ], "maturity": "research", "mechanism": "Each phase, a planning module predicts every player's likely actions from the board state and the dialogue so far. It starts from a policy model trained on human games and refines it with piKL, a regret-minimization search regularized by a KL penalty toward the human-like policy, so predictions stay plausible while improving expected value. The resulting plan fixes intents: the actions Cicero intends to take and the actions it proposes to each recipient. A dialogue language model, trained on human game messages and conditioned on the board, the message history, and these intents, generates candidate messages, which are filtered for consistency with the plan and the game state. Incoming messages update the predicted intents of others, which changes the plan and therefore what Cicero says next. The planner also models what others expect Cicero to do. Because messages are generated from intents rather than freely, what Cicero says tracks what it plans to do.", "name": "Language models grounded in strategic planning for negotiation games", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-24-technique-links", "2026-10-24-review-fixes", "2026-10-25-source-claims-pilot" ] }, "risks": [ "determinism", "consistency", "controllability", "content-safety", "compute-cost", "player-trust" ], "runtime": [ "server-realtime" ], "sources": [ "src-fair-2022-cicero", "src-meta-2022-cicero-project", "src-facebookresearch-2022-diplomacy-cicero", "src-guan-2024-richelieu" ], "status": "sourced", "summary": "Meta's Cicero couples a dialogue language model with a planning engine that infers other players' intentions from the board and conversation, so its free-text negotiation proposes and follows concrete joint plans, reaching human-level Diplomacy.", "trl": 5, "type": "technique", "year": 2022 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 1, "governance-policy": 1, "narrative-generation": 2, "social-relations": 1 }, "incoming": { "applied_in": [ "agent-ambassadors-async", "bottleneck-explainer", "causal-why-query", "content-elastic-events", "drama-aware-storyteller", "event-sourced-sim-runtime", "generated-inspection-shifts", "generative-decoration-catalogs", "intent-level-orders", "language-modding", "local-first-ai-sims", "natural-language-legislation", "negotiating-faction-ai", "placement-that-tells-stories", "salience-gated-llm-agents", "specification-programming-games" ] } }, "affects": [ { "effect": "enables-new", "id": "generative-content", "note": "Generated items, recipes, or characters can be written straight into typed game state (enum types, bounded stats, existing ids) instead of shown as text or parsed with retries, because invalid records cannot be emitted." }, { "effect": "deepens", "id": "natural-language-interaction", "note": "Free player text can be mapped onto the closed set of verbs and target ids legal in the current state, so conversation reliably triggers real game actions rather than only replies." }, { "effect": "replaces-substrate", "id": "utility-action-selection", "note": "An LLM can be dropped into the action-selection slot by constraining it to choose among the same enumerated actions the utility scorer would rank, keeping the engine's action set and preconditions intact." }, { "effect": "cuts-authoring-cost", "id": "dilemma-cards", "note": "Card prompts, two to four options, and per-meter effect values can be generated in the game's card schema with effect sizes bounded by the grammar, leaving designers to curate rather than write." }, { "effect": "enables-new", "id": "player-legislation", "note": "Laws proposed in plain language can be compiled into the world's structured condition-and-effect rule language by grammar-constrained generation, so players need not learn the rule editor." } ], "aliases": [ "structured outputs", "JSON mode", "grammar-guided generation", "GBNF", "Outlines", "XGrammar" ], "builds_on": [], "category": "llm-agents", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "now", "id": "constrained-decoding", "key_results": [ { "confidence": "high", "locator": "Introduction, before How to use Structured Outputs", "quote": "our new model gpt-4o-2024-08-06 with Structured Outputs scores a perfect 100%. In comparison, gpt-4-0613 scores less than 40%", "sources": [ "src-openai-2024-structured-outputs" ], "text": "OpenAI's Structured Outputs (August 2024) reached 100% on a complex JSON-schema-following eval with gpt-4o-2024-08-06, versus under 40% for gpt-4-0613 without constraints." }, { "confidence": "medium", "locator": "Abstract", "quote": "XGrammar can achieve up to 100x speedup over existing solutions", "sources": [ "src-dong-2024-xgrammar" ], "text": "XGrammar reports up to 100x faster grammar-constrained generation than prior engines and near-zero end-to-end overhead when integrated with an inference server." }, { "confidence": "high", "locator": "Abstract", "quote": "guiding text generation with regular expressions and context-free grammars by allowing the construction of an index over a language model's vocabulary", "sources": [ "src-willard-2023-outlines" ], "text": "Outlines showed that regex and CFG guidance can be implemented as finite-state-machine transitions with a precomputed vocabulary index and little per-token overhead." }, { "confidence": "medium", "locator": "Abstract", "quote": "we find that stricter format constraints generally lead to greater performance degradation in reasoning tasks", "sources": [ "src-tam-2024-let-me-speak-freely" ], "text": "Strict format restrictions such as JSON or XML significantly reduced LLM reasoning accuracy compared with free-form answers, with stricter formats causing larger drops." }, { "confidence": "high", "sources": [ "src-deepmind-2026-concordia-repo" ], "text": "Concordia's language-model interface includes a sample_choice call that scores a list of candidate responses and returns an index, raising an error after repeated invalid attempts, which is how its agents pick from closed option sets." }, { "confidence": "low", "sources": [ "src-ggml-2026-llama-cpp-gbnf", "src-geng-2025-jsonschemabench" ], "text": "No source found documents constrained decoding in a specific shipped game; adoption in games is inferred from engine support (llama.cpp GBNF, hosted APIs) rather than observed." } ], "limitations": [ "Validity is not correctness: a well-formed action can still be the wrong action or reference stale state, so engines must re-check preconditions.", "Format constraints can lower reasoning quality; a common workaround is free-form reasoning followed by a constrained final answer, which costs tokens.", "Schema coverage varies: hosted APIs support a JSON Schema subset, and llama.cpp silently skips unsupported features such as nested $ref and uniqueItems and slows down with many optional items.", "First use of a new schema adds compile latency (OpenAI reports under 10 seconds, up to a minute), so per-turn dynamic schemas need caching.", "Output can still fail validation if generation hits a token limit." ], "maturity": "research", "mechanism": "The output format (a JSON Schema, regular expression, or context-free grammar) is compiled into an automaton. At each decoding step the engine computes which vocabulary tokens keep the partial output a valid prefix and sets the logits of all others to minus infinity before sampling. Outlines precomputes an index from finite-state-machine states to allowed tokens; XGrammar handles full CFGs by prechecking context-independent tokens and checking context-dependent ones with a persistent stack, overlapping grammar work with GPU execution; hosted APIs convert schemas to CFGs and cache them; llama.cpp accepts GBNF grammars for local models. For games, the schema is built from current state: an enum of legal verbs, target ids that exist, and numeric fields with bounds, so every output maps to an action the engine can execute. A simpler variant scores a fixed list of candidate actions and returns the chosen index.", "name": "Grammar-constrained decoding for valid game actions", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "consistency", "latency", "evaluation" ], "runtime": [ "server-realtime", "client-realtime", "design-assist" ], "sources": [ "src-willard-2023-outlines", "src-dong-2024-xgrammar", "src-openai-2024-structured-outputs", "src-tam-2024-let-me-speak-freely", "src-geng-2025-jsonschemabench", "src-ggml-2026-llama-cpp-gbnf", "src-deepmind-2026-concordia-repo" ], "status": "sourced", "summary": "Decoding that masks every token which cannot extend a valid string of a schema, regular expression, or grammar, so a model's output always parses as a legal game action or data record.", "trl": 4, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 3, "families_affected": { "narrative-generation": 1, "population-lineage": 2 }, "incoming": { "applied_in": [ "open-ended-evolution-sandbox" ] } }, "affects": [ { "effect": "enables-new", "id": "natural-selection", "note": "In Flow-Lenia, rules travel with matter, so selection can act on creatures that emerge from the physics itself rather than on declared fitness stats." }, { "effect": "deepens", "id": "trait-inheritance", "note": "Rule parameters mix where creatures meet in Flow-Lenia, giving inheritance and hybridization as a spatial process instead of an explicit genome crossover function." }, { "effect": "enables-new", "id": "generative-content", "note": "ASAL turns a text prompt into a search over simulation rules that produce the described phenomenon, a route to generating new creature rule sets from player or designer descriptions at design time." } ], "aliases": [ "Lenia", "Flow-Lenia", "ASAL", "Automated Search for Artificial Life" ], "category": "artificial-life-research", "deployed_in": [], "determinism_impact": "preserves", "horizon": "3-5y", "id": "continuous-alife-and-asal", "key_results": [ { "confidence": "high", "locator": "Abstract", "quote": "More than 400 species in 18 families have been identified, many discovered via interactive evolutionary computation", "sources": [ "src-chan-2019-lenia" ], "text": "Lenia catalogued over 400 species in 18 families of geometric, segmented, resilient lifeforms, many found by interactive evolutionary search." }, { "confidence": "high", "locator": "Discussion", "quote": "Due to the mass conservative nature of Flow Lenia, most of patterns do not grow indefinitely into spatially global patterns (i.e patterns that diffuse on the entire grid also called Turing-like patterns), therefore SLPs are much more common and easier to find", "sources": [ "src-plantec-2023-flow-lenia" ], "text": "Flow-Lenia's mass conservation made localized creatures easier to find, and parameter localization let multiple species with different rules coexist and interact in one world." }, { "confidence": "medium", "locator": "Abstract", "quote": "the discovery of previously unseen Lenia and Boids lifeforms, as well as cellular automata that are open-ended", "sources": [ "src-kumar-2024-asal" ], "text": "ASAL found previously unseen Lenia and Boids lifeforms and open-ended cellular automata in the style of Game of Life, and quantified qualitative phenomena in ways that aligned with human judgment." } ], "limitations": [ "Lenia creatures exist in a narrow band of parameter space; most settings die out or explode, which is what made search tools necessary.", "Creatures are fields without discrete identity, needs or stats, so connecting them to resource economies or player goals needs extra interpretation layers.", "ASAL scores runs with a foundation model on rendered frames, which is offline-scale compute and inherits that model's notion of what looks interesting.", "Open-ended evolution in these substrates is a stated aim, not an established result." ], "maturity": "research", "mechanism": "Lenia stores a continuous field A(x) in [0,1]. Each step it convolves A with a smooth ring-shaped kernel K, passes the result through a bell-shaped growth function G, and updates A by a small time step dt times G(K*A), clipped to [0,1]; certain kernel and growth parameters yield moving, self-maintaining patterns ('species'). Flow-Lenia replaces additive growth with an affinity field that moves matter along its gradient, conserving total mass, and stores the rule parameters in the matter itself, so creatures carrying different rules can coexist, collide and mix. ASAL renders a simulation, embeds the frames with a vision-language foundation model, and scores parameter sets by similarity to a text prompt, by novelty of the embedding trajectory over time, or by diversity across a population; an evolutionary optimizer then searches the substrate's parameters (Boids, Particle Life, Game of Life, Lenia, NCA).", "name": "Continuous cellular automata and foundation-model search for artificial life", "provenance": { "created": "2026-10-24-techniques-sim-math", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "controllability", "legibility", "evaluation" ], "runtime": [ "design-assist", "offline-authoring", "client-realtime" ], "sources": [ "src-chan-2019-lenia", "src-plantec-2023-flow-lenia", "src-kumar-2024-asal" ], "status": "sourced", "summary": "Continuous cellular automata such as Lenia and mass-conserving Flow-Lenia produce lifelike self-organizing creatures, and ASAL uses vision-language foundation models to search their rule spaces for phenomena described in text or for open-ended novelty.", "trl": 3, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 3, "families_affected": { "agents-behavior": 1, "control-automation": 1, "physics-materials": 1 } }, "affects": [ { "effect": "deepens", "id": "vehicle-dynamics", "note": "AI drivers operate the same tire, suspension, and weight-transfer model as the player at champion level, so handling fidelity matters in races against the AI instead of being sidestepped by scripted racing lines." }, { "effect": "deepens", "id": "embodied-skill-control", "note": "Players practice against an opponent whose skill comes from the same controls and physics, at selectable levels (Beginner to Expert, then custom tire and fuel settings), raising the ceiling of a skill-centered mechanic." }, { "effect": "replaces-substrate", "id": "agent-pathfinding", "note": "For movement abilities such as grappling hooks, jetpacks, and double jumps, a learned policy replaces hand-placed navigation-mesh links, so NPCs can traverse spaces designers would otherwise annotate by hand." } ], "aliases": [ "Gran Turismo Sophy", "GT Sophy", "learned NPC controllers", "Unity ML-Agents" ], "builds_on": [ "massively-parallel-simulation" ], "category": "reinforcement-learning", "deployed_in": [ "gran-turismo-7" ], "determinism_impact": "mitigable", "horizon": "now", "id": "deep-rl-npc-policies", "key_results": [ { "confidence": "high", "locator": "Abstract", "quote": "We demonstrate the capabilities of our agent, Gran Turismo Sophy, by winning a head-to-head competition against four of the world’s best Gran Turismo drivers", "sources": [ "src-wurman-2022-gt-sophy" ], "text": "GT Sophy, trained with model-free deep RL across mixed racing scenarios and a reward designed around sportsmanship rules, beat four of the world's best Gran Turismo drivers head to head (Nature, February 2022)." }, { "confidence": "high", "sources": [ "src-polyphony-2023-gt-sophy-race-together" ], "text": "GT Sophy entered Gran Turismo 7 as the time-limited Race Together event in update 1.29 (February 2023), with four races against GT Sophy cars of different performance at Beginner, Intermediate, and Expert levels and a 1vs1 mode." }, { "confidence": "high", "sources": [ "src-polyphony-2023-gt-sophy-2" ], "text": "GT Sophy 2.0 became a permanent Quick Race rival in GT7's Spec II update (November 2023) on nine tracks and 340 cars, 95% of the car models, up from four cars and four tracks, on PS5 only." }, { "confidence": "high", "sources": [ "src-sony-2025-gt-sophy-2-1" ], "text": "GT Sophy 2.1 (update 1.57, March 2025) moved into Custom Race with 19 car and track options, lap count, parts and tuning changes, and adjustable tire and fuel consumption." }, { "confidence": "medium", "locator": "Abstract", "quote": "notably an order of magnitude larger than maps typically used in the Deep RL literature. One of these maps is directly modeled after a Ubisoft AAA game", "sources": [ "src-alonso-2020-drl-navigation-aaa" ], "text": "Ubisoft La Forge's deep RL navigation handled grappling hooks, jetpacks, teleporters, and double jumps on maps about ten times larger than typical RL research maps, one modeled on a Ubisoft AAA game, with at least 90% success." }, { "confidence": "high", "locator": "Section 5, The Unity ML-Agents Toolkit", "quote": "is an open source project which enables researchers and developers to create simulated environments using the Unity Editor and interact with them via a Python API", "sources": [ "src-juliani-2018-unity-ml-agents" ], "text": "Unity's open-source ML-Agents Toolkit lets developers train agents inside Unity scenes." } ], "limitations": [ "Coverage arrives per content item: GT Sophy grew from four cars and four tracks to nine tracks and 340 cars across updates, because each track and car needs training and validation.", "A policy trained to win needs reward shaping and later refinement for sportsmanship and enjoyment; Sony AI describes refining GT Sophy's behavior for etiquette and players of all skill levels after the research version.", "Availability was narrow at first: GT Sophy 2.0 ran only in Quick Race on PS5, not in other events on the same tracks.", "Learned policies are hard to direct for scripted moments, and physics or content patches can require retraining and re-validation.", "Deep RL transfers poorly between games without heavy tuning, as EA reported, so each title builds its own training setup." ], "maturity": "shipped-niche", "mechanism": "The game exposes observations (car state, track geometry, nearby opponents; or local terrain and goal direction for navigation) and actions (steering, throttle, brake; or movement and ability inputs) to a learning agent. Many instances run in parallel, and model-free deep RL learns a policy network that maps observations to actions to maximize a shaped reward. For GT Sophy the reward values course progress and was designed to encode racing's under-specified sportsmanship rules, and training mixed racing scenarios so one policy handles both speed and tactics. For NPCs with jetpacks, grappling hooks, or double jumps, where navigation meshes break down, the reward is reaching a goal. The trained network is small enough to run every frame. Difficulty is set by giving the agent cars of different performance or alternate checkpoints, and later GT Sophy versions expose settings such as tire and fuel consumption. Toolkits such as Unity ML-Agents connect engine scenes to training code.", "name": "Deep reinforcement learning policies for NPC opponents and navigation", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "compute-cost", "controllability", "evaluation" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-wurman-2022-gt-sophy", "src-polyphony-2023-gt-sophy-race-together", "src-polyphony-2023-gt-sophy-2", "src-sony-2025-gt-sophy-2-1", "src-alonso-2020-drl-navigation-aaa", "src-juliani-2018-unity-ml-agents", "src-zhao-2019-winning-isnt-everything" ], "status": "sourced", "summary": "Neural control policies trained with deep reinforcement learning inside the game's own simulation drive opponents and NPCs, as with Sony AI's Gran Turismo Sophy, a selectable rival in Gran Turismo 7 since 2023.", "trl": 7, "type": "technique", "year": 2022 }, { "_derived": { "affects_count": 6, "families_affected": { "environment-ecology": 1, "physics-materials": 3, "space-construction": 2 }, "incoming": { "applied_in": [ "bottleneck-explainer", "branch-preview-forecasts", "calibrated-serious-sims", "parallel-autobalance" ] } }, "affects": [ { "effect": "explains", "id": "design-performance-rating", "note": "The gradient of a design's rating with respect to each part parameter tells the player which strut, joint or track segment most changes the score, enabling hints like 'move this support left'." }, { "effect": "tunes", "id": "soft-body-deformation", "note": "Material stiffness and damping of node-beam or finite-element bodies can be fitted to reference motion or a target feel by gradient descent, as DiffTaichi does for elastic objects." }, { "effect": "tunes", "id": "vehicle-dynamics", "note": "Tire, suspension and drivetrain constants can be fitted to recorded telemetry by differentiating lap or trajectory error, replacing manual parameter sweeps." }, { "effect": "tunes", "id": "contagion", "note": "GradABM shows transmission parameters of a million-agent outbreak model calibrated to real case data by gradients, a route to epidemic games whose curves match real diseases or designer targets." }, { "effect": "tunes", "id": "fluid-gas-simulation", "note": "Differentiable fluid kernels (Warp, DiffTaichi) allow inverse design, solving for emitter or vent settings that produce a target flow, usable for level authoring or tuning puzzles." }, { "effect": "enables-new", "id": "modular-assembly", "note": "A controller for a player-built soft or articulated body can be optimized in tens of gradient iterations, so assembled machines can learn to walk or swim during play." } ], "aliases": [ "differentiable physics", "DiffTaichi", "NVIDIA Warp", "GradABM", "differentiable agent-based models" ], "category": "differentiable-simulation", "deployed_in": [], "determinism_impact": "preserves", "horizon": "3-5y", "id": "differentiable-simulation", "key_results": [ { "confidence": "high", "locator": "Abstract", "quote": "a differentiable elastic object simulator written in our language is 4.2x shorter than the hand-engineered CUDA version yet runs as fast, and is 188x faster than the TensorFlow implementation", "sources": [ "src-hu-2020-difftaichi" ], "text": "DiffTaichi implemented 10 differentiable simulators; its elastic-object simulator was 4.2x shorter than hand-written CUDA at similar speed and 188x faster than a TensorFlow version, and neural controllers optimized in tens of iterations." }, { "confidence": "medium", "locator": "Section 1, Summary of Contributions", "quote": "For an individual accelerator, Brax reaches millions of simulation steps per second", "sources": [ "src-freeman-2021-brax" ], "text": "Brax exposed gradients through a JAX rigid-body engine running at millions of steps per second on one accelerator." }, { "confidence": "medium", "locator": "Article body", "quote": "Customers such as Autodesk are using NVIDIA Warp to write large-scale fluid simulations up to 8x faster and 5x larger than traditional frameworks", "sources": [ "src-nvidia-2024-warp-blog" ], "text": "NVIDIA Warp lets developers write differentiable GPU kernels in Python; NVIDIA reports Autodesk fluid simulations 8x faster and 5x larger than with traditional frameworks." }, { "confidence": "high", "locator": "Abstract", "quote": "GradABM can quickly simulate million-size populations in few seconds on commodity hardware", "sources": [ "src-chopra-2023-gradabm" ], "text": "GradABM made a million-agent epidemic agent-based model differentiable, ran it in a few seconds on commodity hardware, and calibrated it to COVID-19 and influenza data by gradient descent." }, { "confidence": "medium", "locator": "README, MuJoCo API Compatibility", "quote": "Differentiability via Warp is not yet available", "sources": [ "src-deepmind-2026-mujoco-warp" ], "text": "As of 2026, gradient support is uneven across GPU engines: MuJoCo Warp lists differentiability via Warp as not yet available." } ], "limitations": [ "Collisions, contact switching, and integer or threshold game rules give zero, undefined, or misleading gradients and need smoothing or relaxations that change the dynamics being tuned.", "Reverse-mode differentiation must store or recompute intermediate states, so memory grows with timesteps times state size; long horizons also produce exploding or vanishing gradients.", "Support is incomplete in production-grade GPU engines; MuJoCo Warp lists differentiability via Warp as not yet available.", "Gradient descent finds local optima of a stated loss; matching a target curve is not the same as making a system fun or balanced." ], "maturity": "research", "mechanism": "The simulator step is written in a language with automatic differentiation: source-transformed kernels (DiffTaichi, NVIDIA Warp) or traced array programs (JAX in Brax). A loss is defined on the final or accumulated state, for example distance of a soft robot from a goal or error between simulated and observed infection counts. Reverse-mode differentiation propagates the gradient back through every timestep to the parameters, which are updated by gradient descent. Contacts and discrete events are smoothed or given surrogate gradients. For agent-based models, agent state is held in tensors and stochastic discrete events are replaced by differentiable relaxations so the whole population update has gradients (GradABM). The payoff is that tuning takes tens of gradient iterations instead of thousands of blind rollouts.", "name": "Differentiable simulation", "provenance": { "created": "2026-10-24-techniques-sim-math", "reviewed": [ "2026-10-25-source-claims-pilot" ] }, "risks": [ "controllability", "compute-cost" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-hu-2020-difftaichi", "src-freeman-2021-brax", "src-nvidia-2024-warp-blog", "src-chopra-2023-gradabm", "src-deepmind-2026-mujoco-warp" ], "status": "sourced", "summary": "Simulators written so automatic differentiation returns gradients of an outcome with respect to initial state, controls, or rule parameters, replacing black-box search with gradient descent for tuning and inverse design.", "trl": 3, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 5, "families_affected": { "control-automation": 1, "narrative-generation": 1, "physics-materials": 2, "world-generation": 1 }, "incoming": { "applied_in": [ "dream-replay-layer", "symbolic-sim-neural-render" ], "built_upon_by": [ "open-interactive-video-models", "real-time-video-restyling" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "direct-avatar", "note": "Avatar movement, camera response, and weapon use are predicted frame by frame from keyboard and mouse input instead of computed by engine code; responsiveness depends on denoising steps and frame rate (20 fps GameNGen and Oasis, about 10 fps DIAMOND CS:GO)." }, { "effect": "replaces-substrate", "id": "rigid-body-physics", "note": "Collision, gravity, and jumping are learned statistically from pixels rather than integrated; rules the data rarely violates can still be broken, as with DIAMOND's repeated mid-air jumps in CS:GO." }, { "effect": "enables-new", "id": "procedural-terrain", "note": "Terrain is invented on demand from the learned distribution of the training game with no seed or noise function (Oasis), but it is not persistent: regions that leave the short context window are regenerated differently on return." }, { "effect": "enables-new", "id": "generative-content", "note": "Every frame, enemy, and pickup is generated content conditioned on play, so a game can in principle contain situations its source game never had; in practice output stays inside the training game's distribution." }, { "effect": "replaces-substrate", "id": "voxel-destruction", "note": "In Oasis and MineWorld, breaking and placing Minecraft blocks is predicted as pixels with no stored voxel grid, so edits are not durable; Oasis placing a dirt block can bring up an entirely different environment." } ], "aliases": [ "neural game engine", "GameNGen", "DIAMOND", "Oasis", "MineWorld" ], "category": "generative-world-model", "determinism_impact": "breaks", "horizon": "3-5y", "id": "diffusion-game-engines", "key_results": [ { "confidence": "high", "sources": [ "src-valevski-2024-gamengen" ], "text": "GameNGen simulates DOOM at 20 frames per second on a single TPU, with next-frame PSNR of 29.43; human raters picked the real game over the simulation only 58 percent (1.6 s clips) and 60 percent (3.2 s clips) of the time." }, { "confidence": "high", "sources": [ "src-valevski-2024-gamengen" ], "text": "GameNGen conditions on 64 past frames (about 3 seconds) and was trained on about 70M frames recorded from a PPO agent trained for 50M environment steps; the authors state the model cannot be used to create new games." }, { "confidence": "high", "sources": [ "src-alonso-2024-diamond" ], "text": "DIAMOND reached a mean human-normalized score of 1.46 on Atari 100k for agents trained entirely inside a diffusion world model, and a CS:GO world model trained on 87 hours of human play runs at about 10 fps on an RTX 3090 but lets the player jump repeatedly in mid-air, which the real game forbids." }, { "confidence": "high", "sources": [ "src-decart-2024-oasis", "src-techcrunch-2024-decart-oasis", "src-infoq-2024-decart-etched-oasis" ], "text": "Oasis (Decart and Etched, October 2024) generated a Minecraft-like open world at 20 fps in a public browser demo with 500M-parameter open weights; reviewers found the layout rearranged after turning around, and Decart did not say it had Microsoft's permission to train on Minecraft footage." }, { "confidence": "high", "sources": [ "src-guo-2025-mineworld" ], "text": "MineWorld (Microsoft Research, 2025) released an open visual-action autoregressive transformer for Minecraft that generates 4 to 7 frames per second depending on model scale." } ], "limitations": [ "Memory is a few seconds: GameNGen's 64-frame context covers about 3 seconds, and Oasis shows about 3 seconds of memory in one example, so off-screen state such as map layout or enemy positions is lost.", "Requires an existing game to record: GameNGen needed about 70M agent frames from DOOM, DIAMOND 87 hours of CS:GO play; the approach clones a game rather than authoring a new one.", "Each player needs a dedicated accelerator (one TPU-v5 for GameNGen, H100s for Oasis); Maginative reports cost remains a significant barrier to deploying such models at scale, and no per-player-hour figure is published.", "State such as health, ammo, inventory, and score is only as reliable as the pixels that show it; Oasis inventory control is imprecise and distant geometry is fuzzy.", "Training on footage of commercial games raises unresolved licensing questions." ], "maturity": "prototype", "mechanism": "A game is recorded as pairs of frames and player actions (from an RL agent in GameNGen, human play in DIAMOND's CS:GO model, or gameplay video in Oasis). A generator is trained to predict frame t+1 from the last N frames and the action at t: a latent diffusion model in GameNGen, DIAMOND, and Oasis, or a transformer over interleaved image and action tokens in MineWorld. Context frames are noised during training (noise augmentation, Diffusion Forcing, dynamic noising) so the model can recover from its own errors during long rollouts, and sampling is cut to a few denoising steps for real-time play. There is no explicit game state: position, health, enemies, and map layout exist only as pixels or latents inside the context window, so anything that leaves the window is forgotten unless it is visible on screen.", "name": "Diffusion and autoregressive neural game engines", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "compute-cost", "controllability", "ip-provenance", "latency" ], "runtime": [ "server-realtime", "client-realtime" ], "sources": [ "src-valevski-2024-gamengen", "src-alonso-2024-diamond", "src-decart-2024-oasis", "src-techcrunch-2024-decart-oasis", "src-maginative-2024-decart-oasis", "src-infoq-2024-decart-etched-oasis", "src-guo-2025-mineworld" ], "status": "sourced", "summary": "Action-conditioned generative models trained on recordings of one existing game predict each next frame from recent frames and player input, so the model itself plays the role of the game engine.", "trl": 6, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 2, "families_affected": { "physics-materials": 1, "space-construction": 1 }, "incoming": { "applied_in": [ "generative-decoration-catalogs", "symbolic-sim-neural-render" ], "built_upon_by": [ "generated-3d-worlds" ] } }, "affects": [ { "effect": "enables-new", "id": "soft-body-deformation", "note": "PhysGaussian deforms captured objects with MPM directly on the rendered Gaussians, so a scanned real object can bend, dent, or flow without building a node-beam or finite-element mesh first." }, { "effect": "cuts-authoring-cost", "id": "decoration-beauty", "note": "Photoreal props and scenery can be captured from photos in under an hour of training instead of modeled and textured by hand, enlarging decor catalogs that beauty scores consume; lighting is baked into the capture." } ], "aliases": [ "3DGS", "Gaussian splats", "radiance field splatting" ], "category": "neural-rendering-3d", "determinism_impact": "preserves", "horizon": "1-2y", "id": "gaussian-splatting", "key_results": [ { "confidence": "high", "sources": [ "src-kerbl-2023-3d-gaussian-splatting" ], "text": "Kerbl et al. (2023) reached real-time (30 fps or more) novel-view synthesis at 1080p with state-of-the-art visual quality and competitive training times." }, { "confidence": "medium", "sources": [ "src-wikipedia-gaussian-splatting" ], "text": "A full 3DGS scene trains in about 35 to 45 minutes versus about 48 hours for Mip-NeRF 360, but uses more memory than NeRF methods, with peak training memory over 20 GB, and shows popping and elongated-splat artifacts." }, { "confidence": "medium", "sources": [ "src-xie-2023-physgaussian" ], "text": "PhysGaussian simulates elastic, metal, non-Newtonian, and granular behavior directly on Gaussian kernels with an MPM solver, avoiding meshing so the simulated and rendered representation are the same." }, { "confidence": "high", "sources": [ "src-khronos-2026-gltf-gaussian-splatting" ], "text": "Khronos released a KHR_gaussian_splatting glTF 2.0 extension as a release candidate on 2026-02-03 with contributors including Autodesk, Cesium, Esri, Niantic Spatial, and NVIDIA, expecting ratification in Q2 2026." }, { "confidence": "medium", "sources": [ "src-worldlabs-2025-marble", "src-worldlabs-2026-blog" ], "text": "Generated-world products export splats as their highest-fidelity format and render them in the browser (World Labs Marble with the Spark renderer)." } ], "limitations": [ "Colors store captured radiance, so lighting is baked in; dynamic time of day, player-placed lights, and relighting are not supported by the base method.", "Memory heavy: millions of Gaussians per scene and training peaks above 20 GB.", "No collision geometry or semantics: gameplay needs separate meshes or proxies for physics and pathfinding.", "Editing and stylizing splats is harder than editing meshes; tooling and a ratified interchange standard were still arriving in 2026." ], "maturity": "prototype", "mechanism": "Starting from sparse structure-from-motion points, each Gaussian stores position, anisotropic covariance, opacity, and view-dependent color as spherical harmonics. A differentiable tile-based rasterizer projects, depth-sorts, and alpha-blends the Gaussians; gradients from photo reconstruction error update parameters while adaptive density control splits, clones, and prunes Gaussians. Rendering needs no neural network at runtime, so it runs at real-time rates. Extensions attach physics directly to the Gaussians (PhysGaussian runs a Material Point Method on the kernels), stream level-of-detail on the web, and standardize the format in glTF.", "name": "3D Gaussian Splatting", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "compute-cost", "controllability", "ip-provenance" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-kerbl-2023-3d-gaussian-splatting", "src-wikipedia-gaussian-splatting", "src-xie-2023-physgaussian", "src-khronos-2026-gltf-gaussian-splatting", "src-worldlabs-2025-marble", "src-worldlabs-2026-blog" ], "status": "sourced", "summary": "Scenes are represented as millions of anisotropic 3D Gaussians optimized from photos and rasterized in real time, giving photoreal captured or generated environments at 1080p and 30+ fps.", "trl": 6, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 5, "families_affected": { "control-automation": 3, "progression-goals": 1, "social-relations": 1 }, "incoming": { "applied_in": [ "explained-delegation", "sims-as-agent-benchmarks", "teachable-sim-companion", "trainable-creatures-revival" ], "built_upon_by": [ "automated-playtesting", "open-ended-environment-generation" ] } }, "affects": [ { "effect": "enables-new", "id": "unit-commands", "note": "A companion can take open-ended spoken or typed orders, sketches, or emoji and carry them out with ordinary game controls, so the command vocabulary is no longer limited to the verbs a designer bound to buttons." }, { "effect": "enables-new", "id": "delegation-to-ai", "note": "The player can hand an avatar or task to an agent that pursues a stated goal in any game it can see, without game-specific AI code, and that explains what it intends to do." }, { "effect": "deepens", "id": "natural-language-interaction", "note": "SIMA 2 converses about its surroundings, answers questions about the scene, and explains its actions, grounding dialogue in what it sees on screen." }, { "effect": "tunes", "id": "tech-tree", "note": "Self-improving agents that advance through a game's tech tree can serve as automated playtesters of progression in builds they were not trained on." }, { "effect": "enables-new", "id": "programming-by-demonstration", "note": "Agents that follow language instructions and watch play could learn routines from a single narrated demonstration instead of recorded literal steps." } ], "aliases": [ "SIMA", "SIMA 2", "scalable instructable multiworld agent" ], "builds_on": [], "category": "llm-agents", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "3-5y", "id": "generalist-game-agents", "key_results": [ { "confidence": "high", "sources": [ "src-deepmind-2024-sima-blog", "src-deepmind-2024-sima-paper" ], "text": "SIMA 1 learned about 600 basic skills across Valheim, Teardown, No Man's Sky, Wobbly Life, Satisfactory, Goat Simulator 3, Hydroneer, Space Engineers, and Eco; the multi-game agent outperformed per-game specialists, and an agent trained without one game performed nearly as well on it as a specialist." }, { "confidence": "high", "sources": [ "src-deepmind-2024-sima-blog" ], "text": "SIMA 1 tasks completed in about 10 seconds and the agent could not do multi-step strategic planning." }, { "confidence": "high", "sources": [ "src-deepmind-2025-sima-2-paper" ], "text": "SIMA 2 roughly doubled SIMA 1's average task success on training games, approaching human players, and beat SIMA 1 by over 10% on held-out ASKA and MineDojo, completing tasks in 26 of 50 MineDojo categories." }, { "confidence": "high", "sources": [ "src-deepmind-2025-sima-2-paper" ], "text": "Through self-generated tasks and rewards in ASKA, SIMA 2 went from succeeding on under a quarter of a fixed task set to exceeding human reference scores and progressed further through ASKA's tech tree than earlier versions." }, { "confidence": "high", "sources": [ "src-deepmind-2025-sima-2-blog" ], "text": "SIMA 2 was released only as a limited research preview to a small group of academics and game developers." } ], "limitations": [ "Weak combat and precise low-level control; failures in held-out games are often due to poor exploration.", "Short memory bounded by the context window, and trouble with long multi-step tasks that need goal verification.", "Built on Gemini Flash-Lite for latency; fine-tuning cost general reasoning (about 25.5% drop on AIME after SFT).", "Runs as a cloud model reading the screen, so per-session inference cost and round-trip latency apply; no public cost figures.", "Not available to players; the reward model was calibrated on a small set of trajectories." ], "maturity": "research", "mechanism": "The agent sees only rendered frames and a language instruction and emits keyboard and mouse actions in real time, with no access to game source, state APIs, or per-game reward code. SIMA (2024) was trained on human gameplay with language annotations across nine commercial games and four research environments covering about 600 short skills (navigate, interact with objects, use menus). SIMA 2 (2025) puts a fine-tuned Gemini Flash-Lite model at the core, so the agent reasons about goals, states its plan, converses, and follows sketches or emoji. It then improves without human data: a Gemini task setter proposes tasks likely achievable from the current state, the agent attempts them, a Gemini reward model scores each trajectory from 0 to 100 (50 or more counts as success), and the agent trains on the scored experience.", "name": "Generalist instructable game agents (SIMA)", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "compute-cost", "latency", "evaluation", "controllability" ], "runtime": [ "server-realtime", "analytics" ], "sources": [ "src-deepmind-2024-sima-blog", "src-deepmind-2024-sima-paper", "src-deepmind-2025-sima-2-blog", "src-deepmind-2025-sima-2-paper" ], "status": "sourced", "summary": "Agents that read screen pixels and natural-language instructions and output keyboard and mouse actions across many commercial 3D games, with later versions reasoning, conversing, and improving through self-generated tasks.", "trl": 4, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 4, "families_affected": { "physics-materials": 1, "progression-goals": 1, "space-construction": 1, "world-generation": 1 } }, "affects": [ { "effect": "enables-new", "id": "procedural-terrain", "note": "Maps come from a prompt, photo, or video rather than a seed and noise pipeline, and unlike frame-based world models the result is baked geometry that persists, can be saved, and can be expanded region by region." }, { "effect": "cuts-authoring-cost", "id": "scenario-objectives", "note": "Bespoke scenario maps can be produced from a reference image or description and exported with collider meshes, reducing hand modeling per scenario; objectives and rules still need authoring." }, { "effect": "enables-new", "id": "procedural-build-assist", "note": "Chisel turns a coarse blockout into a detailed styled scene, the same rough-input to detailed-output contract as Townscaper-style build assist but with a learned generator; today it is an editor mode, not a runtime player tool." }, { "effect": "cuts-authoring-cost", "id": "rigid-body-physics", "note": "Exported collider meshes give generated scenes collision geometry automatically; Marble describes them as coarse, so precise physics still needs cleanup." } ], "aliases": [ "Marble", "World Labs", "HunyuanWorld", "RTFM", "Atlas" ], "builds_on": [ "gaussian-splatting" ], "category": "neural-rendering-3d", "determinism_impact": "preserves", "horizon": "1-2y", "id": "generated-3d-worlds", "key_results": [ { "confidence": "high", "sources": [ "src-worldlabs-2025-marble" ], "text": "Marble became generally available on 2025-11-12, taking text, images, video, or coarse 3D layouts and exporting Gaussian splats, triangle meshes including collider meshes for coarse physics, and rendered video; Chisel separates blocked-out structure from prompted style." }, { "confidence": "high", "sources": [ "src-tencent-2025-hunyuanworld-1" ], "text": "HunyuanWorld 1.0 (Tencent, July 2025) generates explorable worlds from text or images via panoramic proxies converted to a semantically layered 3D mesh exportable to standard graphics pipelines, listing game development as an application." }, { "confidence": "high", "sources": [ "src-worldlabs-2025-rtfm" ], "text": "RTFM (October 2025) renders generated or real scenes from a single image at interactive frame rates on one H100, using posed frames as spatial memory; it models static scenes only." }, { "confidence": "medium", "sources": [ "src-worldlabs-2026-blog", "src-worldlabs-2026-atlas" ], "text": "World Labs launched a public World API in January 2026, released Atlas to early-access partners in September 2026, and announced it is joining AMD on 2026-09-28." } ], "limitations": [ "Scenes are static: no dynamic objects, agents, or rules are generated, and RTFM lists dynamics and interaction as future work.", "Single-image or text prompts force the model to invent unseen regions; Marble notes artifacts and less crisp far regions that expansion may fix.", "Splat output bakes lighting into colors, so day-night cycles and relighting need conversion to meshes or other methods.", "Collider meshes are coarse and need manual work for gameplay-critical collision.", "Corporate change (World Labs joining AMD in 2026) leaves the hosted product's future terms uncertain." ], "maturity": "prototype", "mechanism": "An image or video diffusion model generates views of a scene (a 360-degree panorama in HunyuanWorld, multi-view images in Marble) from the prompt; the views are lifted to an explicit 3D representation (3D Gaussian splats, or a semantically layered mesh) that can be walked through and re-rendered without regeneration. Marble also emits triangle meshes and coarse collider meshes for physics, can expand or merge worlds, and in Chisel mode takes a blockout of boxes and planes as structure and a text prompt as style. A parallel line renders frames directly instead of building geometry: RTFM keeps generated frames with camera poses as spatial memory and retrieves nearby frames as context. Atlas (2026) is a multimodal model that outputs video, point clouds, and splats.", "name": "Prompt-to-persistent 3D world generation", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "consistency", "controllability", "ip-provenance", "labor", "player-trust" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-worldlabs-2025-marble", "src-tencent-2025-hunyuanworld-1", "src-worldlabs-2025-rtfm", "src-worldlabs-2026-blog", "src-worldlabs-2026-atlas" ], "status": "sourced", "summary": "Models such as World Labs Marble and HunyuanWorld turn text, images, video, or coarse 3D blockouts into persistent explorable scenes exported as Gaussian splats and meshes with collision proxies.", "trl": 6, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 9, "families_affected": { "agents-behavior": 4, "information-legibility": 1, "population-lineage": 1, "social-relations": 3 }, "incoming": { "applied_in": [ "living-workforce-factories", "llm-society-terrarium", "salience-gated-llm-agents", "scientists-not-tech-trees" ], "built_upon_by": [ "human-behavior-simulation-models", "llm-agent-societies" ] } }, "affects": [ { "effect": "enables-new", "id": "memories-thoughts", "note": "Memories become free-text records with LLM-rated importance, retrieved by recency, importance, and relevance, plus reflections that summarize many records into beliefs, instead of a fixed catalog of timed mood modifiers." }, { "effect": "replaces-substrate", "id": "daily-schedules", "note": "Each agent writes its own day plan from its description and memories, decomposed into hour and 5-15 minute steps and revised on events, in place of authored routine tables; the player still sees rhythms of work, meals, and leisure." }, { "effect": "replaces-substrate", "id": "utility-action-selection", "note": "The next action is chosen by prompting with the current plan step, perceived surroundings, and retrieved memories rather than by scoring options against need curves, so choices can reference past events and other agents." }, { "effect": "deepens", "id": "gossip-propagation", "note": "Information moves through generated conversations whose content comes from memory retrieval, so what spreads, to whom, and how it is reworded emerge from dialogue; in Smallville party awareness rose from 4% to 52% of agents with 1.3% hallucinated answers." }, { "effect": "replaces-substrate", "id": "relationship-values", "note": "Relationships are held as remembered interactions and reflections about another agent rather than a numeric affinity, so a relationship's state is a set of retrievable facts; Smallville's acquaintance network density rose from 0.167 to 0.74 in two days." }, { "effect": "enables-new", "id": "social-exchanges", "note": "Agents open-endedly converse, invite, and coordinate (for example planning and attending a party) without a predefined catalog of social moves and preconditions." }, { "effect": "explains", "id": "agent-inspection", "note": "Because an agent's state is natural-language memories, plans, and reflections, the player or designer can read them directly or interview the agent about why it acted, which is how the original paper evaluated believability." }, { "effect": "deepens", "id": "offscreen-life-progression", "note": "Memory streams summarized at low frequency give off-screen characters histories that persist when the player returns." }, { "effect": "enables-new", "id": "aspiration-wants", "note": "Wants generated from an agent's own memories and reflections rather than rolled from a fixed table tied to an aspiration." } ], "aliases": [ "generative agent architecture", "Smallville agents", "plan-act-reflect agents" ], "builds_on": [], "category": "llm-agents", "deployed_in": [ "generative-agents-smallville", "inzoi" ], "determinism_impact": "breaks", "horizon": "now", "id": "generative-agents", "key_results": [ { "confidence": "high", "sources": [ "src-park-2023-generative-agents" ], "text": "In Smallville, 25 agents on gpt-3.5-turbo were rated for believability by 100 human raters (TrueSkill): full architecture 29.89, no reflection 26.88, no reflection or planning 25.64, human crowdworkers role-playing 22.95, no memory, planning, or reflection 21.21." }, { "confidence": "high", "sources": [ "src-park-2023-generative-agents" ], "text": "Over two game days, knowledge of a seeded party spread from 1 to 13 of 25 agents and of a mayoral candidacy from 1 to 8, network density rose from 0.167 to 0.74, five of twelve invited agents attended the party, and 1.3% of 453 awareness answers were hallucinated." }, { "confidence": "high", "sources": [ "src-park-2023-generative-agents", "src-gallotta-2024-llms-games-survey" ], "text": "The simulation cost thousands of dollars in token credits and took multiple days of wall-clock time for 25 agents." }, { "confidence": "high", "sources": [ "src-yu-2024-affordable-generative-agents" ], "text": "Affordable Generative Agents used 31.1% of the baseline's tokens for a 3-agent town over two game days and 42.7% for the 25-agent town, by replaying cached plan-action policies (embedding match at cosine 0.97) and compressing about 2,000 prompt tokens of social context to about 100." }, { "confidence": "medium", "sources": [ "src-nvidia-2026-ace-developer-page", "src-steam-2025-inzoi" ], "text": "NVIDIA describes inZOI's Smart Zois as agents that plan, act, and reflect on their decisions using small language models, and inZOI's Steam disclosure says character actions and thoughts can be influenced through an on-device SLM." } ], "limitations": [ "Cost and wall-clock time: 25 agents over two game days cost thousands of dollars and took days to run on gpt-3.5-turbo (2023 prices).", "Retrieval failures and embellishment: agents sometimes miss relevant memories or add fabricated details to recalled events.", "Instruction-tuned models make agents overly formal and cooperative, and agents misread location norms such as single-occupancy bathrooms or closing hours.", "Each agent step is several sequential LLM calls (importance rating, retrieval-conditioned action, dialogue, reflection), so per-agent latency is seconds, which limits real-time populations unless calls are cached or batched.", "Behavior is not reproducible from a seed with hosted models; replays and saves must store generated outputs." ], "maturity": "shipped-niche", "mechanism": "Each agent appends observations as natural-language records to a memory stream. To act or speak, it retrieves records by a score that sums min-max normalized recency (exponential decay, factor 0.995 per game hour), importance (an LLM rating from 1 to 10 made when the record is stored), and relevance (embedding cosine similarity to the current situation), all with weight 1. When the summed importance of recent records exceeds 150, roughly two or three times a game day, the agent asks the LLM for high-level questions and inferences and stores the answers as reflections that can be retrieved like observations. Planning is top-down: a daily sketch is decomposed into hour chunks and then 5-15 minute actions, and the plan is revised when a perceived event warrants a reaction. Dialogue is generated from retrieved memories about the other agent. Cost-reduction variants replay stored plan-action sequences when a new plan matches an old one by embedding similarity and compress social context into short summaries, cutting LLM calls.", "name": "Generative agents (memory stream, reflection, planning)", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links", "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "compute-cost", "consistency", "latency", "controllability", "evaluation" ], "runtime": [ "server-realtime", "client-realtime" ], "sources": [ "src-park-2023-generative-agents", "src-yu-2024-affordable-generative-agents", "src-gallotta-2024-llms-games-survey", "src-nvidia-2026-ace-developer-page", "src-steam-2025-inzoi" ], "status": "sourced", "summary": "LLM-driven characters that store observations in a natural-language memory stream, retrieve them by recency, importance, and relevance, synthesize reflections, and plan their day top-down.", "trl": 7, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 7, "families_affected": { "control-automation": 2, "environment-ecology": 1, "narrative-generation": 2, "physics-materials": 1, "world-generation": 1 }, "incoming": { "applied_in": [ "dream-replay-layer", "symbolic-sim-neural-render" ], "built_upon_by": [ "open-ended-environment-generation" ] } }, "affects": [ { "effect": "enables-new", "id": "weather", "note": "Genie 3's promptable world events change weather by text during a session with no weather simulation or authored weather assets; the change is visual only and has no stored state for crops, agents, or travel to read." }, { "effect": "cuts-authoring-cost", "id": "storyteller-director", "note": "A director could inject an event (a new character, object, or condition) as a text prompt that Genie 3 renders into the running world, removing per-event art and animation; event selection and pacing still need a separate system with real state." }, { "effect": "enables-new", "id": "procedural-terrain", "note": "Worlds come from a text or image prompt instead of a seed and noise pipeline, in any visual style; persistence is minutes long (about one minute of visual memory in Genie 3, 60-second sessions in Project Genie)." }, { "effect": "replaces-substrate", "id": "direct-avatar", "note": "First- or third-person character movement is rendered by the model from WASD, camera, and jump input; the action space is small and collisions are not guaranteed (characters passing through walls)." }, { "effect": "replaces-substrate", "id": "rigid-body-physics", "note": "Object responses such as doors opening, balloons bursting, water, and smoke are produced by learned video dynamics (Genie 2) rather than a physics engine, with no conservation or collision guarantees." }, { "effect": "enables-new", "id": "generative-content", "note": "Whole explorable scenes, characters, and events are generated at runtime from player text, sketches, or photos, which extends generative content from text and items to playable space." }, { "effect": "enables-new", "id": "god-powers", "note": "Promptable world events work as a free-form god power: typed text adds characters, objects, or weather to the running world; Genie 3 attaches no cost or rule to them and their effects exist only in rendered frames." } ], "aliases": [ "Genie", "Genie 2", "Genie 3", "Project Genie" ], "category": "generative-world-model", "determinism_impact": "breaks", "horizon": "3-5y", "id": "genie-world-models", "key_results": [ { "confidence": "high", "sources": [ "src-bruce-2024-genie" ], "text": "Genie has 11B parameters trained on 30k hours of 2D platformer clips with 8 latent actions; it generates about 1 frame per second and has 16 frames of memory, which the paper lists as limitations." }, { "confidence": "high", "sources": [ "src-deepmind-2024-genie-2" ], "text": "Genie 2 turns one image into a playable 3D world with water, smoke, gravity, lighting, door opening, and NPCs; worlds stay coherent for up to a minute, but most examples last 10 to 20 seconds." }, { "confidence": "high", "sources": [ "src-deepmind-2025-genie-3" ], "text": "Genie 3 renders 720p at 24 fps in real time, stays consistent for a few minutes with visual memory reaching back about one minute, and supports text prompts that change weather or add objects and characters mid-session; DeepMind lists limited action space and multi-agent interaction as open problems." }, { "confidence": "high", "sources": [ "src-techcrunch-2026-project-genie", "src-wikipedia-project-genie" ], "text": "Project Genie opened Genie 3 to US Google AI Ultra subscribers on 2026-01-29 with a 60-second limit per world set by compute cost; a reviewer saw characters walk through walls and sometimes unresponsive controls, and copyrighted settings were refused." }, { "confidence": "medium", "sources": [ "src-wikipedia-project-genie" ], "text": "Shares of video game companies, including Unity, fell after the Project Genie launch." }, { "confidence": "medium", "sources": [ "src-wikipedia-genie-ai-model" ], "text": "Waymo built a world model on Genie 3 (February 2026) to simulate driving edge cases, controllable by driving action, scene layout, and language." } ], "limitations": [ "Session length: Genie 2 coherence is usually 10 to 20 seconds; Genie 3 manages a few minutes; Project Genie caps exploration at 60 seconds because of compute.", "No inspectable state: there are no variables for inventory, health, or object identity, so mechanics that need stocks or rules must live in a separate system.", "Limited action space and no reliable multi-agent interaction (DeepMind's stated Genie 3 limitations).", "Access is a hosted research prototype for paying subscribers; no public API, price per hour, or on-device version is published.", "Guardrails refuse copyrighted settings and nudity, which constrains user-generated worlds." ], "maturity": "prototype", "mechanism": "Genie (2024) learned from 30k hours of unlabeled 2D platformer video: a video tokenizer, a latent action model that infers 8 discrete actions between frames without labels, and a 10.1B-parameter dynamics model that predicts the next frame tokens given an action. Genie 2 (December 2024) is an autoregressive latent diffusion model with a transformer dynamics model that turns a single image into a 3D world controlled by keyboard and mouse. Genie 3 (August 2025) generates frames auto-regressively in real time from a text prompt, keeps visual memory of earlier frames for about a minute so revisited places look the same, and accepts text 'promptable world events' during play that add objects, characters, or weather changes. No explicit 3D scene or game state is stored; consistency comes from the frame history.", "name": "Genie foundation world models", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "compute-cost", "controllability", "ip-provenance", "content-safety" ], "runtime": [ "server-realtime" ], "sources": [ "src-bruce-2024-genie", "src-deepmind-2024-genie-2", "src-deepmind-2025-genie-3", "src-techcrunch-2026-project-genie", "src-wikipedia-project-genie", "src-wikipedia-genie-ai-model" ], "status": "sourced", "summary": "Google DeepMind's Genie models generate playable worlds from a text or image prompt and render them in real time from player input, with Genie 3 reaching 720p at 24 fps and about a minute of visual memory.", "trl": 6, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 6, "families_affected": { "environment-ecology": 1, "physics-materials": 4, "space-construction": 1 }, "incoming": { "applied_in": [ "branch-preview-forecasts", "learned-material-surrogates", "learned-weather-layer" ], "built_upon_by": [ "ml-weather-models" ] } }, "affects": [ { "effect": "accelerates", "id": "fluid-gas-simulation", "note": "Learned fluid surrogates (GNS particles, FNO fields) reproduce solver-quality flow at large speedups, raising resolution or extent of water and gas within a frame budget." }, { "effect": "accelerates", "id": "soft-body-deformation", "note": "MeshGraphNets predicts cloth and structural deformation 10-100x faster than its training solver, allowing finite-element-like deformation where games now use coarse node-beam models." }, { "effect": "deepens", "id": "material-interactions", "note": "GNS simulates several materials (fluids, rigid solids, deformables) interacting in one learned model, so mixed-material reactions need no hand-written pairwise rules." }, { "effect": "deepens", "id": "flight-aerodynamics", "note": "MeshGraphNets learned aerodynamic flow around airfoils, pointing to flight models driven by surrogate CFD fields instead of fixed lift and drag coefficient tables." }, { "effect": "deepens", "id": "structural-integrity", "note": "Learned structural-mechanics surrogates output stress and deformation fields, which could replace binary support checks with graded bending, sagging and stress-driven failure." }, { "effect": "accelerates", "id": "hydrology", "note": "Neural operators solve parametric flow PDEs in one forward pass at any grid resolution, a candidate replacement for per-cell iterative water-flow updates over large terrains." } ], "aliases": [ "GNS", "MeshGraphNets", "Fourier Neural Operator", "FNO", "learned physics simulators", "neural PDE surrogates" ], "category": "neural-physics", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "3-5y", "id": "graph-network-simulators", "key_results": [ { "confidence": "high", "sources": [ "src-sanchezgonzalez-2020-gns" ], "text": "GNS learned fluids, rigid solids and deformable materials as interacting particles; models trained on single steps with thousands of particles generalized to thousands of rollout steps and at least 10x more particles." }, { "confidence": "high", "sources": [ "src-pfaff-2021-meshgraphnets" ], "text": "MeshGraphNets predicted cloth, structural mechanics and aerodynamics on adaptive meshes and ran 1-2 orders of magnitude faster than the simulator it was trained on." }, { "confidence": "high", "sources": [ "src-li-2020-fno" ], "text": "The Fourier Neural Operator modeled Navier-Stokes turbulence with zero-shot super-resolution and was up to three orders of magnitude faster than traditional PDE solvers." } ], "limitations": [ "Errors accumulate over long rollouts; GNS needed training-noise injection to stay stable, and nothing guarantees conservation of mass or energy.", "Each material or domain needs a training set generated by a numerical solver, and behavior outside the training distribution (new materials, extreme speeds, novel geometry) is unreliable.", "Reported speedups are against the high-accuracy solver used for training, not against the cheap approximations games already ship, so the net gain for a game can be much smaller.", "Inference runs on the GPU and competes with rendering for the frame budget.", "Results are cited from 2020-2021 primary papers; this run found no shipped game using GNS or neural operators at runtime." ], "maturity": "research", "mechanism": "Graph network simulators treat particles (GNS) or mesh vertices (MeshGraphNets) as nodes, with edges to neighbors within a radius or along mesh connectivity. An encoder embeds positions, velocities and material type; several rounds of learned message passing compute pairwise interactions; a decoder outputs per-node accelerations that a simple integrator applies. Models are trained on one-step targets from a numerical solver, with noise added to inputs so long rollouts stay stable. MeshGraphNets also learns where to refine the mesh. Neural operators such as the Fourier Neural Operator learn a map from one function to another (for example initial vorticity to vorticity at a later time) by applying learned kernels in Fourier space, so one model covers a family of PDE instances and can be evaluated at grid resolutions not seen in training.", "name": "Graph network simulators and neural operators", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "consistency", "controllability", "compute-cost" ], "runtime": [ "client-realtime", "offline-authoring" ], "sources": [ "src-sanchezgonzalez-2020-gns", "src-pfaff-2021-meshgraphnets", "src-li-2020-fno" ], "status": "sourced", "summary": "Learned simulators that predict the next state of particles, meshes, or continuous fields with message-passing networks or Fourier-space operators, trained on numerical solver output and run 10-1000x faster than that solver.", "trl": 3, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 9, "families_affected": { "agents-behavior": 2, "economy-finance": 3, "governance-policy": 2, "logistics-networks": 1, "narrative-generation": 1 }, "incoming": { "applied_in": [ "ai-liquidity-in-player-markets", "human-calibrated-customers", "participatory-planning-sandboxes" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "customer-value-choice", "note": "Customer purchase decisions could be sampled from a model trained on human choice data, giving realistic heterogeneity and biases, instead of utility = value x preference - price with hand-set weights." }, { "effect": "deepens", "id": "crowd-preferences", "note": "Visitor populations could be drawn from person-grounded profiles whose preferences and tolerances are measured against real people rather than sampled from designer distributions." }, { "effect": "deepens", "id": "elections", "note": "Simulated electorates could be built from survey-grounded agents that reproduce subgroup opinion distributions, so vote shares respond to issues the way measured populations do." }, { "effect": "deepens", "id": "interest-groups", "note": "A group's preferences over policies could be derived from simulated members (silicon samples) rather than a single authored ideology vector." }, { "effect": "deepens", "id": "personality-traits", "note": "A character's disposition can be a rich text profile whose predicted Big Five scores and economic-game behavior match a real respondent, rather than a few discrete trait tags." }, { "effect": "tunes", "id": "dilemma-cards", "note": "Before ship, designers could estimate option pick rates for each card by querying a choice model on the card text; untested on game content." }, { "effect": "deepens", "id": "negotiation", "note": "Counterparties whose acceptance follows models of human bargaining (fairness, anchoring) rather than a single reservation price." }, { "effect": "deepens", "id": "advertising-awareness", "note": "Segment-level response to advertising predicted from models of human choice instead of a flat awareness multiplier." }, { "effect": "tunes", "id": "travel-mode-choice", "note": "Mode-choice weights for time, cost, and comfort can be fitted to predicted human choices instead of hand-set." } ], "aliases": [ "Centaur", "interview-grounded agents", "silicon samples", "Psych-101" ], "builds_on": [ "generative-agents" ], "category": "player-modeling", "deployed_in": [], "determinism_impact": "preserves", "horizon": "3-5y", "id": "human-behavior-simulation-models", "key_results": [ { "confidence": "high", "sources": [ "src-stanford-hai-2025-simulating-human-behavior", "src-park-2024-1000-people" ], "text": "Agents built from two-hour interviews with 1,052 U.S. adults replicated General Social Survey answers 85% as accurately as participants replicated their own answers two weeks later, and performed comparably on Big Five traits, five economic games, and five experiments, with smaller accuracy gaps across racial and ideological groups than demographics-only agents." }, { "confidence": "medium", "sources": [ "src-park-2024-1000-people" ], "text": "The later arXiv version reports normalized GSS accuracy of 83% for interview-only, 82% survey-only, 86% combined, and 74% demographics-only agents." }, { "confidence": "high", "sources": [ "src-binz-2025-centaur" ], "text": "Centaur predicted held-out participants better than 14 cognitive and statistical models and generalized to new cover stories, structural task changes, and new domains; the authors propose it for in silico prototyping of experiments." }, { "confidence": "medium", "sources": [ "src-liu-2025-centaur-instruction-understanding", "src-orr-2025-not-even-wrong" ], "text": "With task instructions removed or replaced by misleading ones, Centaur still generally beat domain-specific cognitive models, which its critics read as reliance on statistical cues rather than understanding the task." }, { "confidence": "high", "sources": [ "src-argyle-2023-out-of-one-many", "src-aher-2022-turing-experiments" ], "text": "Earlier work showed GPT-3 conditioned on survey backstories reproduces subgroup response distributions (algorithmic fidelity), while GPT-4-era models replicated classic experiments but showed a hyper-accuracy distortion on wisdom-of-crowds questions." } ], "limitations": [ "Psych-101 is skewed toward WEIRD participants and toward learning and decision-making tasks; individual differences such as age were not in the prompts.", "Critics show Centaur performs well even without valid instructions, so its predictions on new game situations may track surface statistics.", "Interview-grounded agents need hours of data per person, which raises consent and privacy questions for player data.", "Centaur's 70B base needed about five days on one A100 to fine-tune; the 8B Minitaur variant generalized less robustly.", "No study yet validates these models on choices inside games, where options are visual, timed, and repeated." ], "maturity": "research", "mechanism": "Two approaches. Grounded agents put a person's two-hour interview transcript (or survey answers) in the prompt of a generative agent and ask it to answer new questions as that person; accuracy is normalized by how consistently the real person answers the same items two weeks later. Fine-tuned choice models transcribe experiments into natural language and train a model to predict each trial's choice: Centaur adds rank-8 QLoRA adapters (about 0.15% of parameters) to a frozen 4-bit Llama 3.1 70B and trains one epoch on Psych-101 (160 experiments, 60,092 participants, 10,681,650 choices), scoring held-out participants by negative log-likelihood against cognitive models. Earlier silicon-sample work conditions a model on demographic backstories to reproduce subgroup response distributions. In games the same pipeline would predict how players or simulated citizens choose among options described in text.", "name": "LLM models of individual human choice", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "evaluation", "consistency", "ip-provenance" ], "runtime": [ "analytics", "design-assist" ], "sources": [ "src-park-2024-1000-people", "src-stanford-hai-2025-simulating-human-behavior", "src-binz-2025-centaur", "src-liu-2025-centaur-instruction-understanding", "src-orr-2025-not-even-wrong", "src-argyle-2023-out-of-one-many", "src-aher-2022-turing-experiments" ], "status": "sourced", "summary": "Language models conditioned on interviews or fine-tuned on millions of experimental choices that predict how specific people or populations answer surveys, play economic games, and choose in decision tasks.", "trl": 3, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 6, "families_affected": { "conflict-defense": 1, "economy-finance": 3, "information-legibility": 2 }, "incoming": { "applied_in": [ "negotiating-faction-ai" ], "built_upon_by": [ "cicero-strategic-dialogue" ] } }, "affects": [ { "effect": "deepens", "id": "fog-of-war", "note": "AI players reason over unseen state as a belief distribution and play strategies that are hard to exploit, instead of cheating with full map knowledge or ignoring hidden units, so concealment and feints work against the AI as they would against a person." }, { "effect": "deepens", "id": "scouting-knowledge", "note": "Equilibrium strategies put a value on information, so AI opponents scout, hide their own attributes, and avoid revealing them through their actions, giving the player's scouting system a real counterpart." }, { "effect": "replaces-substrate", "id": "auctions", "note": "AI bidders play mixed strategies over private valuations instead of bidding a fixed fraction of value, which removes simple exploits such as always outbidding the AI by one unit." }, { "effect": "deepens", "id": "military-operations", "note": "Deception becomes a stable part of AI play, as DeepNash bluffed with hidden unit identities in Stratego, so operational wargames could field opponents that feint and conceal force composition at calibrated rates." }, { "effect": "deepens", "id": "rival-firms", "note": "Competitors with private costs or plans can bluff in price wars and market entry, and designers can measure how exploitable a fixed AI pricing script is by computing a best response against it." }, { "effect": "deepens", "id": "negotiation", "note": "Bargaining with hidden reservation values is an imperfect-information game; equilibrium strategies bluff and concede in ways scripted thresholds cannot." } ], "aliases": [ "CFR", "MCCFR", "Libratus", "DeepStack", "Pluribus", "ReBeL", "DeepNash", "Student of Games" ], "category": "game-theoretic-solvers", "determinism_impact": "mitigable", "horizon": "3-5y", "id": "imperfect-information-solvers", "key_results": [ { "confidence": "high", "sources": [ "src-brown-2018-libratus" ], "text": "Libratus beat top professionals at heads-up no-limit Texas hold'em (Science, January 2018)." }, { "confidence": "high", "sources": [ "src-moravcik-2017-deepstack" ], "text": "DeepStack beat professional players with statistical significance over 44,000 hands of heads-up no-limit hold'em using recursive re-solving with a learned value function." }, { "confidence": "high", "sources": [ "src-brown-2019-pluribus", "src-facebook-2019-pluribus-blog" ], "text": "Pluribus beat elite professionals at six-player no-limit hold'em, winning about 5 big blinds per 100 hands over 10,000 hands; its blueprint took 8 days of Monte Carlo CFR on a 64-core server (under $150 of cloud compute) and it played on two CPUs at about 20 seconds per hand." }, { "confidence": "high", "sources": [ "src-brown-2020-rebel" ], "text": "ReBeL combined self-play RL with search, provably converging to a Nash equilibrium in two-player zero-sum games, and reached superhuman heads-up no-limit hold'em with less domain knowledge." }, { "confidence": "high", "sources": [ "src-perolat-2022-deepnash" ], "text": "DeepNash reached a top-3 all-time ranking among expert humans on the Gravon Stratego platform using model-free Regularised Nash Dynamics and no search (Science, December 2022)." }, { "confidence": "high", "sources": [ "src-schmid-2023-student-of-games" ], "text": "Student of Games (Science Advances, November 2023) used one sound algorithm across chess, Go, heads-up no-limit poker, and Scotland Yard, beating the strongest openly available poker agent and the prior best Scotland Yard agent." } ], "limitations": [ "Convergence guarantees hold for two-player zero-sum games; Pluribus's six-player play works empirically without them, and cooperative or general-sum simulation games fall outside the theory.", "Large games need designed abstractions of actions and states; a simulation with continuous economies and many units has no obvious abstraction to solve.", "Runtime search costs seconds per decision (Pluribus averaged about 20 seconds per hand on two CPUs), far above per-frame budgets for many agents.", "Equilibrium play avoids being exploited rather than exploiting weak players, which can feel passive, and randomized bluffs read as arbitrary unless the game explains them." ], "maturity": "research", "mechanism": "A decision point is an information set: all histories the player cannot tell apart. CFR traverses the game tree repeatedly; at each information set it computes each action's counterfactual regret (how much better it would have done, weighted by the probability that others reach that set) and sets the next strategy by regret matching, proportional to positive cumulative regret. The average strategy converges to a Nash equilibrium in two-player zero-sum games, and Monte Carlo CFR samples the tree to scale. Large games are abstracted (bucketed hands and bet sizes) into an offline blueprint that is refined during play by subgame search: Libratus used nested subgame solving, DeepStack re-solved to a depth limit with a learned value function, and Pluribus searched a few moves ahead while letting each opponent pick among four continuation strategies at the leaves. ReBeL and Student of Games learn values over public belief states and search with them, reducing to AlphaZero-style search in perfect-information games. DeepNash instead learns a policy model-free with Regularised Nash Dynamics and no search. Outputs are mixed strategies, so the agent bluffs at calibrated rates.", "name": "Regret minimization and search for imperfect-information games", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "compute-cost", "legibility", "controllability" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-brown-2018-libratus", "src-moravcik-2017-deepstack", "src-brown-2019-pluribus", "src-facebook-2019-pluribus-blog", "src-brown-2020-rebel", "src-perolat-2022-deepnash", "src-schmid-2023-student-of-games" ], "status": "sourced", "summary": "Self-play algorithms that minimize counterfactual regret over information sets, refined by real-time subgame search, compute hard-to-exploit strategies that bluff and hide information and beat top humans at poker, Stratego, and Scotland Yard.", "trl": 5, "type": "technique", "year": 2019 }, { "_derived": { "affects_count": 4, "families_affected": { "narrative-generation": 1, "progression-goals": 1, "risk-failure": 1, "world-generation": 1 }, "incoming": { "applied_in": [ "learned-build-assist", "qd-scenario-libraries" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "procedural-terrain", "note": "Hand-written noise and placement rules give way to a generator trained on example maps or prompted with text, so a seed becomes a prompt plus a sample, and validity must be checked by a solver instead of guaranteed by construction." }, { "effect": "cuts-authoring-cost", "id": "scenario-objectives", "note": "Scenario maps can be produced from a written description of the situation, Word2World-style story to world, with an agent playthrough confirming the stated goal is reachable." }, { "effect": "deepens", "id": "adaptive-threat-scaling", "note": "Difficulty-conditioned generators let the next map or wave be built at a target difficulty instead of multiplying enemy counts, though EA SEED's match-3 model hit the requested difficulty only a third of the time." }, { "effect": "deepens", "id": "generative-content", "note": "Text-to-level models small enough to run on the client (MarioGPT has 96M parameters) let players request maps in words at runtime, extending generated content from text and items to playable space." } ], "aliases": [ "PCGRL", "MarioGPT", "ARLPCG", "Word2World", "text-to-level", "LLM level generation" ], "category": "procedural-generation", "determinism_impact": "mitigable", "horizon": "1-2y", "id": "learned-level-generation", "key_results": [ { "confidence": "high", "sources": [ "src-khalifa-2020-pcgrl" ], "text": "PCGRL (2020) framed level design as a Markov decision process with three editing representations tested on three games, giving a fast generator that needs few or no example levels." }, { "confidence": "high", "sources": [ "src-sudhakaran-2023-mariogpt" ], "text": "MarioGPT (2023) fine-tuned a 96M-parameter DistilGPT2 model on 37 Super Mario Bros levels; 88.4% of 250 generated levels were completed by an A* agent versus about 31% for an LSTM baseline, and output followed text prompts." }, { "confidence": "high", "sources": [ "src-todd-2023-llm-level-generation" ], "text": "LLM Sokoban level generation improved sharply with training-set size, and scarce level data was the main obstacle (FDG 2023)." }, { "confidence": "medium", "sources": [ "src-nasir-2024-word2world" ], "text": "Word2World (2024) used off-the-shelf LLMs without fine-tuning to write a story and place tiles into a coherent, playable 2D world." }, { "confidence": "high", "sources": [ "src-villanueva-2024-match3-conditional-generation" ], "text": "EA SEED's difficulty-conditioned match-3 generator produced 51.39% valid levels versus 46.15% in its training data, but matched the requested difficulty only 33.33% of the time (mean gap 10.19 moves)." }, { "confidence": "medium", "sources": [ "src-gisslen-2021-arlpcg" ], "text": "ARLPCG trained a generator adversarially against a solver agent in a 3D platformer and a racing game, producing new environments shown to be solvable." } ], "limitations": [ "Example data is scarce: MarioGPT trained on 37 levels and LLM Sokoban results scaled with dataset size, so a new game's handful of authored maps may not suffice.", "Validity is not guaranteed: about 12% of MarioGPT levels were not completed by an A* agent, so every output needs a solver check.", "Conditional control is weak: difficulty conditioning hit its target a third of the time and lowered size and tile-distribution accuracy.", "Results are mostly 2D tile grids for small games; large simulation maps with resources, rivers, and settlement suitability carry global constraints these models do not represent.", "Training on existing commercial levels such as Super Mario Bros raises IP and platform disclosure questions for shipped use." ], "maturity": "research", "mechanism": "PCGRL treats level design as a sequential decision problem: an agent edits a tile grid step by step (narrow, turtle, or wide representations of where it may edit) and is rewarded by the change in a quality score such as playability, path length, or region count, so the trained generator builds a level in fast forward passes without example levels. ARLPCG adds a solver agent and rewards the generator for levels that are challenging but solvable, with auxiliary control inputs. Language-model generators serialize levels as text and train or prompt a model to continue them: MarioGPT fine-tunes DistilGPT2 with cross-attention to a prompt naming pipes, enemies, blocks, and elevation; Sokoban generators fine-tune on level corpora; Word2World prompts a general LLM to write a story, extract characters and tiles, and place them. Outputs are checked by search agents or solver bots and rejected or repaired if unplayable. Conditional generators such as EA SEED's match-3 cVAE take size, symmetry, and target difficulty as inputs.", "name": "Learned level and map generation (PCGRL, MarioGPT, LLM generators)", "provenance": { "created": "2026-10-24-techniques-rl-design" }, "risks": [ "controllability", "evaluation", "ip-provenance", "player-trust" ], "runtime": [ "offline-authoring", "design-assist", "client-realtime" ], "sources": [ "src-khalifa-2020-pcgrl", "src-sudhakaran-2023-mariogpt", "src-todd-2023-llm-level-generation", "src-nasir-2024-word2world", "src-villanueva-2024-match3-conditional-generation", "src-gisslen-2021-arlpcg" ], "status": "sourced", "summary": "Generators trained by reinforcement learning or on example levels, including small and large language models that emit tile maps from text prompts, produce playable levels and maps without hand-written generation rules.", "trl": 4, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 2, "families_affected": { "control-automation": 1, "narrative-generation": 1 } }, "affects": [ { "effect": "deepens", "id": "direct-avatar", "note": "Motion-matching-quality responsive locomotion fits in a few MB of weights, making it affordable on memory-constrained platforms and for more character types." }, { "effect": "deepens", "id": "generative-content", "note": "Players author new character motions from their own video (inZOI), so the animation set available to life-sim characters is no longer fixed by the studio's mocap library." } ], "aliases": [ "learned motion matching", "HyperMotion", "video-to-motion" ], "category": "animation-motion", "deployed_in": [ "inzoi" ], "determinism_impact": "preserves", "horizon": "now", "id": "learned-motion-synthesis", "key_results": [ { "confidence": "high", "sources": [ "src-ubisoft-2020-learned-motion-matching" ], "text": "Learned motion matching shrank a 590 MB motion matching locomotion system to about 17 MB of weights (8.5 MB with 16-bit quantization), about 70 times smaller, with nearly identical output and constant CPU cost." }, { "confidence": "medium", "sources": [ "src-wikipedia-fifa-22" ], "text": "FIFA 22's HyperMotion used a machine learning algorithm over more than 8.7 million frames of 22-player match capture to generate player animation in real time, on PS5, Xbox Series X/S, and Stadia only." }, { "confidence": "high", "sources": [ "src-steam-2025-inzoi" ], "text": "inZOI lets players create character motions by uploading video files, disclosed as an AI-assisted feature." }, { "confidence": "high", "sources": [ "src-epic-2026-motion-matching-docs" ], "text": "Unreal Engine ships database motion matching through its Pose Search plugin, the non-learned baseline that learned motion matching compresses." } ], "limitations": [ "Kinematic output ignores physics, so characters still clip or float when the environment differs from the training motion.", "Learned motion matching saves memory but not per-character CPU time, which still limits crowd counts.", "Video-to-motion quality depends on the uploaded footage; occlusion and camera motion produce artifacts that the game must filter or let players discard." ], "maturity": "shipped-niche", "mechanism": "Motion matching picks the next pose by searching a database of mocap frames for the best match to a query of current pose and desired trajectory. Learned motion matching replaces the database with three small networks: a Decompressor that maps a feature vector to a full pose, a Stepper that advances features one frame, and a Projector that maps a query to the nearest valid features, giving nearly identical output at constant CPU cost. HyperMotion trains a model on frames of full-team match capture to generate transitions in real time. Video-to-motion tools estimate 3D poses from ordinary video and retarget them onto a game rig. No physics is simulated.", "name": "Learned kinematic motion synthesis", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "controllability", "evaluation", "labor" ], "runtime": [ "client-realtime", "offline-authoring", "design-assist" ], "sources": [ "src-ubisoft-2020-learned-motion-matching", "src-wikipedia-fifa-22", "src-steam-2025-inzoi", "src-epic-2026-motion-matching-docs" ], "status": "sourced", "summary": "Neural networks generate or compress kinematic character animation, as in Ubisoft's learned motion matching, EA's HyperMotion, and inZOI's motion extraction from uploaded video.", "trl": 7, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 8, "families_affected": { "economy-finance": 2, "governance-policy": 3, "multiplayer-persistence": 1, "population-lineage": 1, "social-relations": 1 }, "incoming": { "applied_in": [ "agent-ambassadors-async", "ai-liquidity-in-player-markets", "llm-society-terrarium", "natural-language-legislation", "participatory-planning-sandboxes" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "elections", "note": "Voters become LLM agents that hear persuasion from other agents and vote on amendments, as in Project Sid's constitutional votes, instead of computing a vote from satisfaction and bloc membership." }, { "effect": "deepens", "id": "gossip-propagation", "note": "Memes, beliefs, and a seeded religion spread through populations of 500 to one million agents, with content mutating in conversation and spread shaped by social structure or a recommender feed." }, { "effect": "enables-new", "id": "labor-market", "note": "Roles emerge from agents' social context and goals (farmers, miners, guards in 30-agent villages) rather than from job slots filled by education level; AgentSociety adds agent-level employment and consumption decisions." }, { "effect": "enables-new", "id": "player-legislation", "note": "NPC populations can propose, debate, and vote on rule changes in natural language and then comply or not, so legislation is no longer limited to human players in a shared world." }, { "effect": "replaces-substrate", "id": "policy-influence-network", "note": "A policy's effect on groups is measured by running it on a population of agents (AgentSociety's universal basic income and hurricane experiments) instead of reading hand-authored weighted edges from policies to voter groups." }, { "effect": "deepens", "id": "taxation", "note": "Tax compliance becomes an agent decision that responds to law changes and influencer messaging, as in Project Sid's 20% tax experiment, rather than an automatic deduction at the set rate." }, { "effect": "deepens", "id": "persistent-world", "note": "A shared world can be populated by hundreds of autonomous agents that keep roles, culture, and rules while human players come and go; PIANO was built for real-time interaction with humans and other agents." }, { "effect": "replaces-substrate", "id": "offscreen-life-progression", "note": "Unplayed households can be advanced by cheap language-agent societies that generate plausible life events instead of fixed event tables." } ], "aliases": [ "Project Sid", "PIANO", "AgentSociety", "Concordia", "OASIS", "generative agent-based modeling" ], "builds_on": [ "generative-agents" ], "category": "multi-agent-social-sim", "deployed_in": [ "project-sid" ], "determinism_impact": "breaks", "horizon": "3-5y", "id": "llm-agent-societies", "key_results": [ { "confidence": "high", "sources": [ "src-altera-2024-project-sid" ], "text": "In Minecraft, PIANO agents in 30-agent villages specialized into persistent roles (farmer, miner, guard, blacksmith) only when the social module was present; agents largely obeyed a 20% tax law and changed compliance after voted constitutional amendments, with pro- and anti-tax influencer agents shifting outcomes; a religion seeded by 20 priests kept spreading through a 500-agent population for two hours." }, { "confidence": "high", "sources": [ "src-altera-2024-project-sid" ], "text": "Project Sid runs above 1,000 agents exceeded Minecraft server limits and agents became sporadically unresponsive, so the main cultural analysis used 500 agents." }, { "confidence": "medium", "sources": [ "src-piao-2025-agentsociety" ], "text": "AgentSociety simulated over 10,000 agents with about 5 million interactions and reproduced qualitative findings on polarization, inflammatory message spread, universal basic income, and a hurricane shock." }, { "confidence": "medium", "sources": [ "src-yang-2024-oasis" ], "text": "OASIS ran up to one million agents on X- and Reddit-like platforms, replicated information spreading, group polarization, and herd effects, and found stronger group dynamics at larger agent counts." }, { "confidence": "high", "sources": [ "src-vezhnevets-2023-concordia", "src-vezhnevets-2025-multi-actor-game-engine" ], "text": "Concordia formalizes a Game Master entity that grounds natural-language actions in physical, social, or digital space; its 2025 design pattern builds the Game Master from the same components as other entities so designers can configure scenarios without engine code." } ], "limitations": [ "Scale ceilings come from the host game as well as the model: Project Sid agents became unresponsive above 1,000 agents on a Minecraft server.", "PIANO agents lack vision and spatial reasoning and have no innate drives such as survival or curiosity, so building and navigation are weak.", "Agents built on pretrained models reproduce known human institutions (taxes, votes, religion) rather than inventing new ones, so outcomes are bounded by training data.", "Per-agent LLM calls multiply with population; AgentSociety chose an API without request limits to avoid rate-limit artifacts, and none of these papers reports per-agent-hour cost in a form a game could budget against.", "Validation is qualitative (matching stylized findings), so designers cannot easily tune or test collective outcomes." ], "maturity": "research", "mechanism": "Scaling generative agents from tens to thousands needs architecture and infrastructure changes. Project Sid's PIANO runs about ten modules concurrently (memory, action awareness, goal generation, social awareness, talking, skill execution) that read and write a shared agent state; fast modules use small non-LLM models and slow ones call an LLM, and a cognitive controller acts as a bottleneck that broadcasts one high-level decision so speech and actions stay consistent. AgentSociety gives each agent a profile plus emotion, needs, and cognition layers and simulates mobility, social contact, employment, and consumption across linked urban, social, and economic spaces, distributing agents over Ray processes and routing messages over MQTT. Concordia has agents state intended actions in natural language and a Game Master agent decide outcomes. OASIS runs up to one million agents inside a recommender-driven social feed.", "name": "Many-agent LLM societies", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links", "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "compute-cost", "consistency", "controllability", "evaluation", "legibility" ], "runtime": [ "server-realtime" ], "sources": [ "src-altera-2024-project-sid", "src-piao-2025-agentsociety", "src-yang-2024-oasis", "src-vezhnevets-2023-concordia", "src-vezhnevets-2025-multi-actor-game-engine" ], "status": "sourced", "summary": "Simulations of hundreds to a million LLM-driven agents with concurrent cognitive modules, shared environments, and message buses that produce role specialization, rule changes, cultural spread, and polarization.", "trl": 4, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 5, "families_affected": { "information-legibility": 4, "narrative-generation": 1 }, "incoming": { "applied_in": [ "bottleneck-explainer", "causal-why-query", "event-sourced-sim-runtime", "explained-delegation", "illustrated-chronicles", "intent-level-orders", "sim-instructor-coach" ] } }, "affects": [ { "effect": "deepens", "id": "advisors", "note": "Advisors can answer arbitrary player questions about the current situation and recommend actions in context, instead of speaking a fixed catalog of threshold-triggered messages." }, { "effect": "explains", "id": "agent-inspection", "note": "The inspection panel can answer why an agent did something by summarizing its needs, itemized mood terms, and recent thought log into a short causal account." }, { "effect": "explains", "id": "causal-tooltips", "note": "A long nested modifier chain can be condensed into a ranked plain-language list of the largest causes and what would change them, while the numbers still come from the simulation." }, { "effect": "explains", "id": "stats-dashboards", "note": "Graph changes can be annotated with narrated causes drawn from the event log, such as which policy or disaster preceded a drop." }, { "effect": "deepens", "id": "history-chronicle", "note": "Recorded events can be retold as biographies, recaps, or highlight summaries for the player, a commentator and reteller role." } ], "aliases": [ "AI advisor chat", "natural-language explanations of agent behavior", "LLM commentator", "log summarization" ], "builds_on": [], "category": "llm-narrative-dialogue", "deployed_in": [], "determinism_impact": "preserves", "horizon": "1-2y", "id": "llm-explanation-of-state", "key_results": [ { "confidence": "medium", "sources": [ "src-zhang-2023-explaining-agent-behavior" ], "text": "Explanations generated from a learned behavior representation plus an LLM were rated about as helpful as a human domain expert's, with little hallucination, and supported clarification and counterfactual follow-ups." }, { "confidence": "medium", "sources": [ "src-nvidia-2026-ace-developer-page" ], "text": "NVIDIA lists a Total War: PHARAOH AI advisor that helps players learn the game's systems using a small on-device model connected to game data; no public release date was found." }, { "confidence": "high", "sources": [ "src-deepmind-2025-sima-2-blog" ], "text": "SIMA 2 explains its intended steps and answers questions about its behavior and environment in commercial 3D games." }, { "confidence": "high", "sources": [ "src-peng-2024-player-driven-emergence" ], "text": "GPT-4 converted 28 players' logs into narrative graphs, letting designers count 53 player-created narrative nodes that their design did not include." }, { "confidence": "high", "sources": [ "src-wang-2024-text-world-simulators" ], "text": "GPT-4 predicted whole text-game state transitions correctly at most 59.9% of the time and environment-driven changes at most 49.7%, so explanations of dynamics must be computed by the simulation and narrated, not inferred by the model." }, { "confidence": "medium", "sources": [ "src-gallotta-2024-llms-games-survey" ], "text": "A 2024 survey found research on LLM commentators and retellers of game events rather limited." } ], "limitations": [ "Wrong but fluent causes: models are unreliable at inferring dynamics, so explanations are only trustworthy when causes are passed in from the simulation.", "Serializing large state (thousands of modifiers, long logs) into context is slow and costly; selection of what to include determines what can be explained.", "Hard to QA: the answer space is open-ended, and a confident wrong explanation damages trust more than no explanation.", "Shipped evidence is thin: known game examples are tech demos or research previews." ], "maturity": "prototype", "mechanism": "The game serializes the relevant slice of state (indicator values and their contributing modifiers, an agent's thought log and action trace, recent events) into the prompt, or exposes it through retrieval or tool calls over game data and documentation, and the model answers a player question or writes a summary. Zhang et al. first learn a compact representation of an agent's behavior from observed states and actions, then have the LLM explain from that representation, which limits hallucination and supports follow-up and what-if questions. For designers, a model can segment play logs into strategies and merge them into a graph to compare with the intended path. The explanation layer reads state and does not write it, so the simulation stays deterministic.", "name": "LLM explanation of simulation state", "provenance": { "created": "2026-10-24-techniques-llm-agents" }, "risks": [ "consistency", "evaluation", "latency", "player-trust" ], "runtime": [ "client-realtime", "server-realtime", "analytics" ], "sources": [ "src-zhang-2023-explaining-agent-behavior", "src-nvidia-2026-ace-developer-page", "src-deepmind-2025-sima-2-blog", "src-peng-2024-player-driven-emergence", "src-wang-2024-text-world-simulators", "src-gallotta-2024-llms-games-survey" ], "status": "sourced", "summary": "Language models that read simulation state, modifier breakdowns, agent traces, or event logs and answer why and what-if questions or summarize what happened for players and designers.", "trl": 5, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 6, "families_affected": { "narrative-generation": 4, "social-relations": 1, "world-generation": 1 }, "incoming": { "applied_in": [ "content-elastic-events", "drama-aware-storyteller", "generated-inspection-shifts", "illustrated-chronicles", "placement-that-tells-stories" ] } }, "affects": [ { "effect": "enables-new", "id": "storyteller-director", "note": "The director can author an event in language for the exact current state, check it against canon and rules, and narrate it, instead of choosing among a finite catalog of pre-written incidents by a pacing budget." }, { "effect": "replaces-substrate", "id": "triggered-event-chains", "note": "Follow-up beats are generated from story state and player intent rather than scheduled by authored templates with trigger predicates, while the player still experiences multi-step chains." }, { "effect": "deepens", "id": "villager-requests", "note": "Characters can ask for things tied to the specific story so far, not only fetch templates with filled slots; ungrounded models invent quests that do not exist, so requests must be checked against state." }, { "effect": "deepens", "id": "generative-content", "note": "Story continuation from Do, Say, Story, and See inputs makes the whole narrative generated content; grounded engines turn it into typed state changes." }, { "effect": "deepens", "id": "natural-language-interaction", "note": "Any typed action is adjudicated for plausibility and consequence rather than matched against a parser's verb list, so unanticipated approaches produce new narrative paths (53 emergent nodes in one 28-player study)." }, { "effect": "cuts-authoring-cost", "id": "procedural-characters", "note": "NPCs with backstories, goals, and personality are generated from designer criteria and voiced in dialogue, removing per-character writing." } ], "aliases": [ "AI game master", "AI Dungeon", "Concordia Game Master", "story engine", "AI narrator" ], "builds_on": [], "category": "llm-narrative-dialogue", "claims": [ { "confidence": "high", "sources": [ "src-valve-2026-steamworks-content-survey", "src-valve-2024-ai-content-on-steam" ], "text": "Steam requires games with live-generated AI content to disclose it and describe guardrails against illegal output, and does not permit live-generated adult sexual content." }, { "confidence": "medium", "sources": [ "src-zhu-2023-calypso" ], "text": "Dungeon Masters using CALYPSO presented some generated text almost directly and treated rougher output as ideas while keeping creative control." }, { "confidence": "medium", "sources": [ "src-buongiorno-2024-pangea" ], "text": "PANGeA generated settings, key items, and Big Five personality-biased NPCs from designer criteria and used the model to keep free-text player input within the narrative, tested in a browser demo and a Unity game." }, { "confidence": "medium", "sources": [ "src-totallyhuman-2024-steam-genai-count", "src-totallyhuman-2025-steam-genai-count", "src-totallyhuman-2025-steam-ai-revenue" ], "text": "Steam generative AI disclosures rose from about 1,000 games in April 2024 to 7,818 (about 7% of the catalog, a little under 20% of 2025 releases) in July 2025 and 10,258 (about 8%) in November 2025; about 60% of disclosures concern visual assets, so live-generated runtime systems such as LLM narration are a small share." } ], "deployed_in": [ "ai-dungeon" ], "determinism_impact": "breaks", "horizon": "now", "id": "llm-game-master", "key_results": [ { "confidence": "high", "sources": [ "src-wikipedia-ai-dungeon" ], "text": "AI Dungeon 2 (December 2019) ran the full 1.5B-parameter GPT-2 and moved to GPT-3 in mid-2020, reaching over 100,000 players in its first week and 1.5 million by June 2020." }, { "confidence": "high", "sources": [ "src-wikipedia-ai-dungeon" ], "text": "AI Dungeon's 2021 content filter, added to block sexual content involving minors, produced false positives on innocuous text and led human moderators to read flagged private stories, prompting review bombing." }, { "confidence": "high", "sources": [ "src-peng-2024-player-driven-emergence" ], "text": "In a 28-player study with five GPT-4 NPCs holding backstories, clues, and ordered goals, play produced 53 narrative nodes absent from the designers' graph; about 15 seconds of lag frustrated 40% of players and NPCs forgot earlier statements in 10% of reports." }, { "confidence": "medium", "sources": [ "src-hiddendoor-2026-blog-index", "src-hiddendoor-2026-homepage" ], "text": "Hidden Door announced early access in August 2025 and offers licensed worlds such as Dracula; its engine checks every story turn against canon and rules before display and reports a tuned 1.7B model beating a 120B general model on one core task." }, { "confidence": "high", "sources": [ "src-vezhnevets-2023-concordia", "src-vezhnevets-2025-multi-actor-game-engine", "src-deepmind-2026-concordia-repo" ], "text": "Concordia's Game Master converts agents' natural-language actions into outcomes and checks physical plausibility, and its 2025 design pattern classifies uses as Simulationist, Dramatist, or Evaluationist." }, { "confidence": "medium", "sources": [ "src-gallotta-2024-llms-games-survey" ], "text": "Surveyed LLM game masters tend to accommodate even bizarre player requests that human GMs would curb, and LLM NPCs can invent quests that do not exist in the game." } ], "limitations": [ "Long-session consistency: models forget or contradict earlier facts unless state is externalized into memory or a typed world model.", "Accommodation bias: models rarely say no, which removes challenge and lets players talk past any obstacle.", "Content moderation is hard: AI Dungeon's filter both missed and over-blocked, and Steam requires guardrail disclosure for live-generated content.", "Latency of several to 15 seconds per turn frustrated players in a GPT-4 study; multi-step grounded pipelines add calls per turn.", "Every turn is an inference cost for the life of the service, which pushes products to subscriptions or small tuned models." ], "maturity": "shipped-niche", "mechanism": "The player types an intent; the system assembles a prompt from recent story, pinned memories, and setting notes and asks the model to continue the story, as AI Dungeon has since 2019. Grounded variants split the job: actors state intended actions in language, a Game Master component checks plausibility against rules and state, decides what happens, writes state changes, and only then narrates (Concordia). Hidden Door runs intent through plan, canon check, and world-state update steps over a typed, auditable world model, using many small tuned models for about 100 narrative tasks; its example blocks Dracula from entering a house uninvited. Research systems generate setting, items, and NPCs from designer criteria and use the model to validate free-text input against the story (PANGeA), or assist a human Dungeon Master (CALYPSO).", "name": "LLM game master and narrative director", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "content-safety", "controllability", "compute-cost", "evaluation" ], "runtime": [ "server-realtime" ], "sources": [ "src-wikipedia-ai-dungeon", "src-vezhnevets-2023-concordia", "src-vezhnevets-2025-multi-actor-game-engine", "src-deepmind-2026-concordia-repo", "src-hiddendoor-2026-homepage", "src-hiddendoor-2026-blog-index", "src-peng-2024-player-driven-emergence", "src-buongiorno-2024-pangea", "src-zhu-2023-calypso", "src-gallotta-2024-llms-games-survey", "src-valve-2026-steamworks-content-survey", "src-valve-2024-ai-content-on-steam", "src-totallyhuman-2024-steam-genai-count", "src-totallyhuman-2025-steam-genai-count", "src-totallyhuman-2025-steam-ai-revenue" ], "status": "sourced", "summary": "A language model that adjudicates player and NPC intentions, decides outcomes, generates events and quests, and narrates them, either free-form or checked against a typed world state and canon.", "trl": 8, "type": "technique", "year": 2019 }, { "_derived": { "affects_count": 5, "families_affected": { "narrative-generation": 1, "physics-materials": 1, "progression-goals": 2, "resources-production": 1 } }, "affects": [ { "effect": "enables-new", "id": "hand-crafting", "note": "The recipe space becomes unbounded: any pair of items yields a generated result rather than a lookup in a finite designer table, and the first answer is cached as the permanent recipe." }, { "effect": "enables-new", "id": "material-interactions", "note": "Interactions between arbitrary materials or elements can be resolved by a model rather than an authored pair table, but whole-state accuracy near 60% means results must be bounded to simple outputs and cached." }, { "effect": "deepens", "id": "generative-content", "note": "Generated results are written back as shared, persistent rules rather than one-off text, so the world's content grows with play and stays consistent across players." }, { "effect": "enables-new", "id": "collection-goals", "note": "Collections become open-ended races to discover never-generated items, with a First Discovery credit for the first player to create each one." }, { "effect": "enables-new", "id": "combination-discovery", "note": "A language model decides the result of each new combination and caches it, turning a finite authored recipe graph into an open one (Infinite Craft)." } ], "aliases": [ "LLM as rules engine", "open-ended combination", "text world simulator", "memoized generation" ], "builds_on": [], "category": "generative-world-model", "deployed_in": [ "infinite-craft" ], "determinism_impact": "mitigable", "horizon": "now", "id": "llm-rule-adjudication", "key_results": [ { "confidence": "high", "sources": [ "src-wikipedia-infinite-craft" ], "text": "Infinite Craft (January 2024) starts from water, fire, wind, and earth and generates a new element and emoji with Llama 2 and Llama 3.1 for any untried pair, storing the result so each pair gives every player the same answer." }, { "confidence": "high", "sources": [ "src-wang-2024-text-world-simulators" ], "text": "On 76,369 transitions from 31 text games, GPT-4 predicted the full next state correctly at most 59.9% of the time, up to 77.1% for action-driven and at most 49.7% for environment-driven changes; humans scored about 80% on a subset where GPT-4 scored about 50%." }, { "confidence": "high", "sources": [ "src-vezhnevets-2023-concordia" ], "text": "Concordia's Game Master uses a language model to decide whether stated actions are physically plausible and what their effects are." } ], "limitations": [ "Accuracy: as a general state simulator GPT-4 is right on whole-state transitions at most 59.9% of the time, so errors compound over multi-step dynamics.", "Legibility: generated rules follow word association rather than a learnable system, so players cannot predict outcomes from principles.", "Cost and load scale with unique queries; Infinite Craft's traffic overloaded its servers.", "Outputs must be filtered for offensive results because every cached answer becomes permanent shared content." ], "maturity": "shipped-niche", "mechanism": "The game describes the inputs (two items, or an object state plus an action) as text and asks the model for the result, constrained to a short name, an icon, or a small state diff. Infinite Craft first looks the pair up in a database; only a never-tried pair goes to the model (Llama 2, later Llama 3.1), and the answer is stored, so the rule becomes fixed and shared by all players after the first query and the first finder is credited. The cache turns a stochastic model into a growing deterministic rule table and bounds cost to unique queries. Benchmarks of LLMs as text world simulators measure how often the predicted next state matches the true one.", "name": "LLM-adjudicated state transitions (LLM as game mechanic)", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links", "2026-10-24-review-fixes" ] }, "risks": [ "consistency", "legibility", "content-safety", "compute-cost" ], "runtime": [ "server-realtime" ], "sources": [ "src-wikipedia-infinite-craft", "src-wang-2024-text-world-simulators", "src-vezhnevets-2023-concordia" ], "status": "sourced", "summary": "A language model computes the outcome of an interaction (two elements combined, an action on an object) from a text description of the inputs, with results cached so every player sees the same rule.", "trl": 7, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 8, "families_affected": { "agents-behavior": 3, "logistics-networks": 1, "physics-materials": 1, "population-lineage": 1, "space-construction": 2 }, "incoming": { "applied_in": [ "branch-preview-forecasts", "calibrated-serious-sims", "event-sourced-sim-runtime", "learned-material-surrogates", "living-workforce-factories", "mean-field-crowds", "neural-npc-motion", "open-ended-evolution-sandbox", "parallel-autobalance" ], "built_upon_by": [ "deep-rl-npc-policies", "physical-world-foundation-models", "physics-based-character-control" ] } }, "affects": [ { "effect": "accelerates", "id": "rigid-body-physics", "note": "Batched GPU rigid-body solvers step thousands of worlds or bodies per kernel launch, so physics-heavy evaluation that was CPU-bound becomes cheap enough for large sweeps." }, { "effect": "enables-new", "id": "creature-learning", "note": "Locomotion policies train in seconds to minutes on one GPU (Brax, MuJoCo Playground), so a creature could learn to move or act through RL during a play session instead of from canned animation." }, { "effect": "enables-new", "id": "modular-assembly", "note": "A player-built creature or machine can be dropped into thousands of parallel worlds and have a controller trained for its specific body, so assembled designs move competently without the player programming them." }, { "effect": "deepens", "id": "natural-selection", "note": "Fitness can be measured by simulating every individual of a large population in its own physics world each generation, so selection acts on actual behavior rather than on summed stat values." }, { "effect": "deepens", "id": "design-performance-rating", "note": "A player design can be scored over hundreds of randomized loads, terrains or seeds in parallel, giving a robustness distribution instead of one deterministic test run." }, { "effect": "tunes", "id": "attribute-match-engine", "note": "Millions of simulated matches per hour let designers check how attribute weights and tactic modifiers shift outcome distributions before shipping a match-engine change." }, { "effect": "accelerates", "id": "traffic-congestion", "note": "GPUDrive shows multi-agent vehicle simulation above one million steps per second, which points to agent-level traffic at counts where city builders now fall back to aggregate flow." }, { "effect": "tunes", "id": "utility-action-selection", "note": "Cheap batched rollouts make it practical to search utility-curve weights, or train a policy to replace them, against metrics such as survival or need satisfaction over many simulated colonies." } ], "aliases": [ "batch simulation", "many-world simulation", "GPU-accelerated simulators", "Isaac Gym", "Brax", "MJX", "Madrona" ], "category": "differentiable-simulation", "deployed_in": [], "determinism_impact": "preserves", "horizon": "1-2y", "id": "massively-parallel-simulation", "key_results": [ { "confidence": "high", "sources": [ "src-makoviychuk-2021-isaac-gym" ], "text": "Isaac Gym ran physics and policy training on a single GPU and reported 2-3 orders of magnitude faster RL training than pipelines using a CPU simulator." }, { "confidence": "high", "sources": [ "src-freeman-2021-brax" ], "text": "Brax reported millions of simulation steps per second on one accelerator and hundreds of millions across accelerators, training Ant locomotion in about ten seconds versus close to half an hour for standard PPO." }, { "confidence": "high", "sources": [ "src-shacklett-2023-madrona" ], "text": "Madrona's GPU batch ECS ran OpenAI Hide and Seek at over 1.9 million steps per second, Overcooked-AI at 40 million and Hanabi at 20 million on an RTX 4090, 5-33x faster than strong 32-thread CPU baselines." }, { "confidence": "high", "sources": [ "src-matthews-2024-craftax" ], "text": "Craftax reimplemented the survival-crafting game Crafter in JAX, ran up to 250x faster than the Python original, and trained PPO for 1 billion interactions in under an hour on one GPU." }, { "confidence": "high", "sources": [ "src-kazemkhani-2024-gpudrive" ], "text": "GPUDrive, built on Madrona, simulated multi-agent driving from real road scenarios at over one million steps per second." }, { "confidence": "high", "sources": [ "src-zakka-2025-mujoco-playground" ], "text": "MuJoCo Playground (2025) trained locomotion and manipulation policies in minutes on a single GPU and transferred them zero-shot to real robots." }, { "confidence": "low", "sources": [ "src-kemper-2024-genesis-430000x", "src-genesis-2024-genesis-world" ], "text": "Genesis (December 2024) was publicized as reaching 43 million FPS (430,000x real time) on simple scenes; press noted skepticism and that contact-rich dexterous scenes remain hard." }, { "confidence": "high", "sources": [ "src-thacker-2025-physx-gpu-open-source", "src-larabel-2025-newton-linux-foundation" ], "text": "GPU physics became more open in 2025: NVIDIA released PhysX 5.6 GPU source under BSD-3 in April, and the Warp-based Newton engine (NVIDIA, Google DeepMind, Disney Research) moved to the Linux Foundation in September." } ], "limitations": [ "Speedups come from running many independent worlds in parallel; a single large coupled world, such as one city, gains much less and per-world latency does not improve.", "Game logic must be rewritten in a GPU-friendly form (JAX in Craftax, C++ compiled to CUDA in GPUDrive and Madrona), which is a port rather than a recompile.", "Madrona requires all component archetypes to be declared up front and does not support adding or removing components at runtime; it is labeled research code with breaking API changes.", "Headline frame rates are for simple scenes; contact-rich and dexterous scenes remain hard and much slower (reported caveat on Genesis).", "Much of the stack is NVIDIA-specific: MuJoCo Warp needs an NVIDIA GPU for speed, and porting PhysX's 500+ CUDA kernels to other vendors is described as a lot of work.", "Policies trained in a batched simulator exploit its physics quirks, so behavior must be re-validated in the shipping engine if that engine differs." ], "maturity": "research", "mechanism": "World state for N independent instances is stored as structure-of-arrays tensors in accelerator memory. Each tick launches one set of kernels that integrates physics, evaluates game logic, and computes observations and rewards for every instance at once, so policy inference and training read the state without CPU-GPU copies. Implementations compile the step function: JAX/XLA (Brax, MJX, Craftax), CUDA kernels written in Python (NVIDIA Warp, MuJoCo Warp), or a GPU entity-component-system with a task graph (Madrona, where game logic is C++ compiled to CUDA). Throughput comes from parallelism across worlds, not from making one world faster, so the main uses are reinforcement learning (billions of steps), parameter sweeps, and evaluating many randomized variants of one design.", "name": "Massively parallel GPU simulation", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "controllability", "evaluation", "determinism" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-makoviychuk-2021-isaac-gym", "src-freeman-2021-brax", "src-shacklett-2023-madrona", "src-madrona-2026-repository", "src-matthews-2024-craftax", "src-kazemkhani-2024-gpudrive", "src-zakka-2025-mujoco-playground", "src-deepmind-2026-mujoco-warp", "src-genesis-2024-genesis-world", "src-kemper-2024-genesis-430000x", "src-thacker-2025-physx-gpu-open-source", "src-larabel-2025-newton-linux-foundation" ], "status": "sourced", "summary": "Simulators that keep thousands of world instances as tensors on a GPU or TPU and step them all in one kernel, raising throughput to millions of steps per second for agent training and Monte Carlo sweeps.", "trl": 4, "type": "technique", "year": 2021 }, { "_derived": { "affects_count": 6, "families_affected": { "agents-behavior": 2, "economy-finance": 1, "logistics-networks": 3 }, "incoming": { "applied_in": [ "ai-liquidity-in-player-markets", "mean-field-crowds" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "traffic-congestion", "note": "Trip routing changes from each agent reacting to current travel times, which oscillates, to an equilibrium in which drivers anticipate congestion, computed once for the population at a cost independent of car count." }, { "effect": "deepens", "id": "crowd-preferences", "note": "Park visitors can choose attractions in equilibrium with expected queues and crowding rather than greedily, producing realistic spreading of crowds across a park." }, { "effect": "deepens", "id": "service-queues", "note": "Queue choice and arrival timing become equilibrium behavior, so changing capacity at one service shifts load across the others as people anticipate the change." }, { "effect": "deepens", "id": "supply-demand-pricing", "note": "Prices can come from many small rational producers and consumers in equilibrium (as in MFG models of energy markets), instead of a fixed elasticity formula." }, { "effect": "accelerates", "id": "agent-pathfinding", "note": "One equilibrium policy over positions serves any number of agents, like a flow field that already includes congestion effects, replacing per-agent searches." }, { "effect": "deepens", "id": "travel-mode-choice", "note": "Mode split solved as a population equilibrium in which each traveler responds to congestion created by everyone else." } ], "aliases": [ "MFG", "mean field MARL", "online mirror descent for MFGs" ], "category": "game-theoretic-solvers", "deployed_in": [], "determinism_impact": "preserves", "horizon": "3-5y", "id": "mean-field-games", "key_results": [ { "confidence": "high", "sources": [ "src-lasry-2007-mean-field-games" ], "text": "Lasry and Lions (2007) formulated mean field games as coupled forward-backward nonlinear PDEs for many rational players reacting to aggregates and showed the problems are generally well-posed." }, { "confidence": "high", "sources": [ "src-perolat-2021-mfg-omd" ], "text": "Online mirror descent converged to MFG Nash equilibria under monotonicity, outperformed fictitious play, and solved MFGs with hundreds of billions of states." }, { "confidence": "high", "sources": [ "src-yang-2018-mean-field-marl" ], "text": "Mean-field MARL let each agent learn against its neighbors' mean action, scaling multi-agent learning to many-agent battle games and solving the Ising model with model-free RL." }, { "confidence": "medium", "sources": [ "src-carmona-2020-mfg-finance-economics" ], "text": "MFGs are applied to energy markets, bitcoin mining, macro-finance models and contract theory." }, { "confidence": "medium", "sources": [ "src-lauriere-2022-learning-mfg-survey" ], "text": "Learning methods for static, stationary and evolutive MFG equilibria and social optima are now surveyed as an established toolkit." } ], "limitations": [ "Assumes anonymous, symmetric, individually negligible agents; named characters with unique state, or a player large enough to move the aggregate, need major-player extensions.", "Equilibrium describes settled behavior; the learning, overshoot and panic dynamics games often want are absent unless modeled separately.", "The forward-backward coupling needs the whole horizon at once, so player interventions mid-run require re-solving or a fast approximate solver.", "Convergence guarantees such as those for online mirror descent need monotonicity assumptions that game rules can violate.", "No game implementation was found in this run's sources." ], "maturity": "research", "mechanism": "Each agent is small, anonymous and interchangeable, so the game depends on others only through the population distribution m(t,x). An equilibrium is a fixed point: the representative agent's value function u solves a backward Hamilton-Jacobi-Bellman equation given m, and m evolves forward by a Fokker-Planck equation under the resulting policy. Numerical and learning methods iterate this pair: fictitious play averages past best responses; online mirror descent accumulates Q-values and updates the policy by a softmax of them; RL variants learn from samples when the model is unknown. Mean-field multi-agent RL applies the same idea locally, with each agent learning against the mean action of its neighbors. Cost scales with the size of the individual state space, not with the number of agents.", "name": "Mean field games and mean-field learning", "provenance": { "created": "2026-10-24-techniques-sim-math", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "legibility", "controllability" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-lasry-2007-mean-field-games", "src-perolat-2021-mfg-omd", "src-yang-2018-mean-field-marl", "src-carmona-2020-mfg-finance-economics", "src-lauriere-2022-learning-mfg-survey" ], "status": "sourced", "summary": "Mean field game theory replaces an N-player game with one representative agent responding to the population distribution, and recent learning methods compute these equilibria for state spaces of hundreds of billions of states.", "trl": 3, "type": "technique", "year": 2021 }, { "_derived": { "affects_count": 5, "families_affected": { "environment-ecology": 3, "information-legibility": 1, "risk-failure": 1 }, "incoming": { "applied_in": [ "learned-weather-layer" ] } }, "affects": [ { "effect": "deepens", "id": "weather", "note": "Weather becomes a coherent global field with fronts, cyclones and multi-day evolution across hundreds of variables, instead of a per-region Markov chain of weather states." }, { "effect": "enables-new", "id": "forecasting", "note": "Ensembles (50 GenCast members) give calibrated probabilities, so an in-game forecast can carry real uncertainty that players plan against and that occasionally misses honestly." }, { "effect": "deepens", "id": "random-disasters", "note": "Storms emerge from atmospheric state with forecastable tracks (Aurora beat operational centres on 5-day cyclone tracks), turning disasters from dice rolls into threats with lead time." }, { "effect": "deepens", "id": "pollution-spread", "note": "Aurora's fine-tuned air-quality forecasts run about 100,000x faster than CAMS, a path to pollution plumes transported by simulated wind at playable cost." }, { "effect": "deepens", "id": "planetary-climate", "note": "Earth-system foundation models fine-tuned to several domains suggest climate sims coupling atmosphere, ocean waves and air chemistry, though they are trained for days-ahead weather, not decades." } ], "aliases": [ "GraphCast", "GenCast", "Aurora", "AIFS", "FGN", "WeatherNext", "ML weather prediction" ], "builds_on": [ "graph-network-simulators" ], "category": "neural-physics", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "3-5y", "id": "ml-weather-models", "key_results": [ { "confidence": "high", "sources": [ "src-lam-2023-graphcast", "src-deepmind-2023-graphcast-blog" ], "text": "GraphCast produced 0.25 degree, 10-day global forecasts in under one minute on one TPU v4 machine and beat ECMWF HRES on over 90% of 1,380 verification targets." }, { "confidence": "high", "sources": [ "src-price-2024-gencast" ], "text": "GenCast's 50-member diffusion ensembles were more skillful than ECMWF's ENS on 97.2% of 1,320 targets; one 15-day forecast took about 8 minutes on a Cloud TPU v5." }, { "confidence": "high", "sources": [ "src-bodnar-2025-aurora" ], "text": "Aurora (1.3B parameters) matched or beat operational systems on 74% of air-quality targets, 86% of ocean-wave variables and 92% of 0.1 degree weather targets, beat seven centres on 5-day cyclone tracks, and ran about 100,000x faster than CAMS for air quality." }, { "confidence": "high", "sources": [ "src-ecmwf-2025-aifs-operational" ], "text": "ECMWF made its AIFS machine-learned model operational on 25 February 2025, reporting about 1,000x less energy per forecast than its physics model, at 28 km grid spacing versus 9 km." }, { "confidence": "medium", "sources": [ "src-alet-2025-fgn" ], "text": "FGN (2025) reported state-of-the-art ensemble skill and cyclone tracks while being trained only on per-location targets." } ], "limitations": [ "Trained on Earth's reanalysis record, so they reproduce Earth's atmosphere only; fictional planets, altered geography or deliberate climate change lie outside the training data.", "Grids are coarse for game maps: 0.25 degree (about 28 km at the equator), and operational AIFS runs at 28 km.", "Inference needs an accelerator and gigabytes of weights (Aurora has 1.3B parameters), heavy for client-side use.", "Designers cannot author dramatic timing; the model follows the statistics of real weather, so pacing must be imposed by selecting or conditioning forecasts.", "Autoregressive rollouts are validated to 10-15 days; behavior over seasons or years is not established by these sources." ], "maturity": "research", "mechanism": "Models take the current and recent atmospheric state on a global 0.25 degree latitude-longitude grid (temperature, wind, humidity, pressure at many levels plus surface variables) and predict the state one fixed step later; rolling the prediction forward gives a 10-15 day forecast. GraphCast maps grid cells onto a multi-resolution icosahedral mesh and runs learned message passing there. GenCast is a conditional diffusion model that samples stochastic trajectories in 12-hour steps, so repeated samples form an ensemble with calibrated spread. FGN (2025) builds ensembles from learned model perturbations trained on CRPS. Aurora is a 1.3B-parameter foundation model pretrained on over a million hours of geophysical data and fine-tuned for air quality, ocean waves, cyclone tracks and 0.1 degree weather. Inference replaces hours on a supercomputer with seconds to minutes on one accelerator.", "name": "Machine-learned weather forecasting", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "compute-cost", "controllability" ], "runtime": [ "offline-authoring", "server-realtime" ], "sources": [ "src-lam-2023-graphcast", "src-deepmind-2023-graphcast-blog", "src-price-2024-gencast", "src-bodnar-2025-aurora", "src-ecmwf-2025-aifs-operational", "src-alet-2025-fgn" ], "status": "sourced", "summary": "Neural networks trained on decades of atmospheric reanalysis forecast global weather 10-15 days ahead in seconds to minutes on one accelerator, matching or beating physics-based operational models on most verification targets.", "trl": 3, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 1, "conflict-defense": 1, "economy-finance": 1, "governance-policy": 1, "progression-goals": 1 }, "incoming": { "applied_in": [ "explained-delegation", "trainable-creatures-revival" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "rival-firms", "note": "Competitor AI becomes a policy trained by self-play against the game's own rules, so it re-adapts when economic rules change in a patch instead of needing rewritten heuristics; training compute per rule set is the cost." }, { "effect": "tunes", "id": "tech-tree", "note": "Self-play agents trained on each candidate tree reveal dominant research orders and unused branches, the same comparison AlphaZero made across chess rule variants." }, { "effect": "tunes", "id": "policy-toggles", "note": "Each toggle combination is a rule variant; converged self-play under each variant measures how much a policy shifts outcomes and flags toggles that are strictly dominant or never worth enacting." }, { "effect": "deepens", "id": "creature-learning", "note": "Sample-efficient model-based learners (EfficientZero needed about two hours of experience) make it plausible for a creature to learn new behaviors from player reward within a session, by planning in a learned model of its world rather than reacting with a fixed network." }, { "effect": "deepens", "id": "military-operations", "note": "Opponents plan with a learned value function plus search over unit actions, giving strong play in operational simulations where hand-written doctrine is brittle; latent models allow planning where exact rollouts are too slow." } ], "aliases": [ "AlphaZero", "MuZero", "EfficientZero", "DreamerV3", "Dreamer 4" ], "builds_on": [ "monte-carlo-tree-search" ], "category": "reinforcement-learning", "determinism_impact": "mitigable", "horizon": "3-5y", "id": "model-based-planning-muzero", "key_results": [ { "confidence": "high", "sources": [ "src-silver-2017-alphazero" ], "text": "AlphaZero reached superhuman chess, shogi, and Go within 24 hours from random play given only the rules, beating a world-champion program in each game." }, { "confidence": "high", "sources": [ "src-schrittwieser-2020-muzero" ], "text": "MuZero matched AlphaZero in Go, chess, and shogi and surpassed R2D2 on Atari while learning its own model instead of being given the rules (Nature, December 2020)." }, { "confidence": "medium", "sources": [ "src-wikipedia-muzero" ], "text": "MuZero's board-game training used 16 TPUs for training and 1,000 for self-play at 800 simulations per move, and it performed poorly on 6 Atari games." }, { "confidence": "high", "sources": [ "src-ye-2021-efficientzero" ], "text": "EfficientZero reached 194.3% mean and 109.0% median human-normalized score on Atari 100k from about two hours of real-time game experience." }, { "confidence": "high", "sources": [ "src-tomasev-2020-alphazero-chess-variants" ], "text": "AlphaZero trained from scratch on nine chess rule variants was used to compare decisiveness, piece values, and dynamism across variants as an in silico game balance assessment." }, { "confidence": "high", "sources": [ "src-hafner-2025-dreamerv3" ], "text": "DreamerV3 (Nature, April 2025) beat specialized methods on over 150 tasks with one configuration and was the first algorithm to collect Minecraft diamonds from scratch without human data, first at about 30M environment steps (about 17 days of play)." }, { "confidence": "medium", "sources": [ "src-hafner-2025-dreamer-4" ], "text": "Dreamer 4 (September 2025) learned a Minecraft world model that runs interactively on one GPU and obtained diamonds from offline data only, choosing sequences of over 20,000 keyboard and mouse actions from pixels." }, { "confidence": "high", "sources": [ "src-deepmind-2022-muzero-youtube" ], "text": "MuZero-RC chose per-frame VP9 quantisation on part of YouTube live traffic for an average 4% bitrate reduction, its first production deployment, outside games." } ], "limitations": [ "Training is expensive: MuZero's board-game runs used 16 TPUs for training and 1,000 for self-play, and each significant rule change needs retraining or fine-tuning.", "Self-play converges to strong play, not human play, so balance conclusions can reflect exploits humans would never find or use.", "Learned dynamics drift over long rollouts; MuZero did poorly on 6 Atari games and DreamerV3 needed about 30M steps (17 days of play) for its first Minecraft diamond.", "Most simulation games are many-player, simultaneous-move, partly hidden, and long-horizon, which the two-player perfect-information AlphaZero setup does not cover directly.", "Policies are opaque, so designers cannot read why an AI rival chose a strategy, and giving it a personality requires reward shaping rather than editing rules." ], "maturity": "research", "mechanism": "AlphaZero pairs MCTS with a network that outputs a move prior p and a value v for a state; games generated by the search train p toward search visit counts and v toward outcomes, so search and network improve each other given only the rules. MuZero drops the need for rules at planning time: a representation network encodes observations into a latent state, a dynamics network predicts the next latent state and reward for an action, and a prediction network outputs policy and value, so MCTS runs entirely in the learned latent space. EfficientZero adds self-supervised consistency and value-prefix losses for sample efficiency. Dreamer learns a recurrent world model and trains an actor-critic on imagined trajectories; DreamerV3 adds symlog and return normalization so one hyperparameter set works across domains, and Dreamer 4 trains its world model mostly on unlabeled video and learns the policy offline inside it. For design work the same loop can be retrained on each rule variant and the converged play compared.", "name": "Self-play planning with learned models (AlphaZero, MuZero, Dreamer)", "provenance": { "created": "2026-10-24-techniques-rl-design" }, "risks": [ "compute-cost", "controllability", "legibility", "evaluation" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-silver-2017-alphazero", "src-schrittwieser-2020-muzero", "src-wikipedia-muzero", "src-ye-2021-efficientzero", "src-tomasev-2020-alphazero-chess-variants", "src-hafner-2025-dreamerv3", "src-hafner-2025-dreamer-4", "src-deepmind-2022-muzero-youtube" ], "status": "sourced", "summary": "Agents that combine tree search or imagined rollouts with neural policy and value networks trained by self-play, where MuZero and Dreamer also learn the environment's dynamics, reach superhuman play from rules or pixels alone.", "trl": 4, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 1, "conflict-defense": 1, "control-automation": 1, "economy-finance": 1, "social-relations": 1 }, "incoming": { "built_upon_by": [ "automated-rule-generation", "model-based-planning-muzero" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "utility-action-selection", "note": "Agents choose actions by sampling futures from the game's forward model and backing up outcomes, instead of scoring one-step options with hand-tuned utility curves, so choices account for consequences several actions ahead at a per-decision compute cost." }, { "effect": "deepens", "id": "military-operations", "note": "Tactical AI plans multi-unit action sequences within a time budget, as in Fable Legends' real-time tactical planning, rather than following scripted doctrine; the iteration budget becomes a difficulty lever." }, { "effect": "deepens", "id": "faction-diplomacy", "note": "AI factions can weigh war, peace, and alliance options by simulating future turns of the campaign, the kind of high-level campaign decision Total War: Rome II gave to MCTS, so choices respond to the actual strategic position instead of fixed opinion thresholds." }, { "effect": "deepens", "id": "rival-firms", "note": "AI competitors can pick prices, expansions, and acquisitions by rolling the economic model forward, so rival play adapts to the player's position and to rule changes without bespoke heuristics per market rule." }, { "effect": "deepens", "id": "delegation-to-ai", "note": "Governors or managers given a subsystem can search over action sequences toward the player-set goal with the same forward model, staying competent in rule configurations designers did not script for." } ], "aliases": [ "MCTS", "UCT", "information-set MCTS" ], "category": "game-theoretic-solvers", "determinism_impact": "mitigable", "horizon": "now", "id": "monte-carlo-tree-search", "key_results": [ { "confidence": "high", "sources": [ "src-kocsis-2006-uct", "src-wikipedia-monte-carlo-tree-search" ], "text": "Kocsis and Szepesvari introduced UCT, applying the UCB1 bandit rule to node selection in Monte Carlo planning (ECML 2006); Coulom coined the name Monte Carlo tree search the same year." }, { "confidence": "medium", "sources": [ "src-browne-2012-mcts-survey" ], "text": "A 2012 survey in IEEE TCIAIG catalogued MCTS variants, enhancements, and applications in board, card, and video games after its first six years." }, { "confidence": "medium", "sources": [ "src-wikipedia-monte-carlo-tree-search" ], "text": "Total War: Rome II used MCTS in its high-level campaign AI, and Fable Legends used real-time MCTS for tactical planning (2015 talk)." } ], "limitations": [ "Needs a fast, copyable forward model; a full simulation state with thousands of agents is expensive to clone and step, so practical use usually requires an abstracted model or macro-actions.", "Random playouts estimate value poorly in long-horizon systems with delayed consequences such as economies and colony survival unless a heuristic rollout policy or evaluation function is supplied.", "Time-budgeted search returns different moves on different hardware, which breaks replays and lockstep multiplayer unless budgets are counted in iterations with seeded randomness.", "Determinized search over hidden information averages over sampled worlds and can make choices that are only good if it could see the hidden state (strategy fusion)." ], "maturity": "shipped-niche", "mechanism": "Each iteration has four steps. Selection descends from the root by a tree policy; UCT picks the child that maximizes mean reward plus c*sqrt(ln N / n), trading exploitation against exploration. Expansion adds an untried child. Simulation plays forward from that node with a cheap default policy (random or heuristic) to a terminal state or depth limit. Backpropagation adds the outcome to the visit count and value of every node on the path. The search is anytime: it stops after an iteration count or time budget and returns the most-visited action. It needs only a forward model that can copy a state, list legal actions, and apply them, so it works for rule sets without a known evaluation function. Variants add chance nodes, determinization or information-set trees for hidden information, simultaneous moves, and macro-actions that search over plans instead of single moves.", "name": "Monte Carlo tree search for game AI", "provenance": { "created": "2026-10-24-techniques-rl-design" }, "risks": [ "latency", "controllability", "legibility" ], "runtime": [ "client-realtime" ], "sources": [ "src-kocsis-2006-uct", "src-browne-2012-mcts-survey", "src-wikipedia-monte-carlo-tree-search" ], "status": "sourced", "summary": "A best-first search that grows a game tree from repeated simulated playouts and picks branches with a bandit rule such as UCT, letting AI players plan from a forward model without a hand-tuned evaluation function.", "trl": 7, "type": "technique", "year": 2006 }, { "_derived": { "affects_count": 3, "families_affected": { "control-automation": 1, "narrative-generation": 1, "physics-materials": 1 } }, "affects": [ { "effect": "replaces-substrate", "id": "direct-avatar", "note": "In WHAMM the avatar's movement and camera are rendered by the model from controller input, at 10+ fps with a 0.9 second context." }, { "effect": "replaces-substrate", "id": "rigid-body-physics", "note": "Movement and collisions follow learned game dynamics; the Nature paper's consistency criterion is that generated play respects controller input and the game's physics, which is measured statistically rather than enforced." }, { "effect": "enables-new", "id": "generative-content", "note": "Designers insert an object or character into a frame and the model generates plausible gameplay around it (persistency), giving a preview of how a content change might play before it is built." } ], "aliases": [ "Muse", "WHAM", "WHAMM", "World and Human Action Model" ], "category": "generative-world-model", "determinism_impact": "breaks", "horizon": "3-5y", "id": "muse-wham", "key_results": [ { "confidence": "high", "sources": [ "src-kanervisto-2025-wham" ], "text": "The Nature paper (February 2025) frames consistency, diversity, and persistency of user modifications as the three capabilities generative models need to support game ideation, and shows WHAM provides them." }, { "confidence": "high", "sources": [ "src-microsoft-2025-muse-blog" ], "text": "WHAM was trained on more than 1 billion images and controller actions, over 7 years of continuous play, from Bleeding Edge matches recorded under the game's EULA; model weights, sample data, and the WHAM Demonstrator were released on Azure AI Foundry." }, { "confidence": "high", "sources": [ "src-microsoft-2025-whamm" ], "text": "WHAMM ran a playable section of one Quake II level at 10+ fps and 640x360 after training on one week of gameplay data; enemies were blurry, damage could be wrong, context was 0.9 seconds, and health counters were unreliable." } ], "limitations": [ "WHAM-1.6B generates about one image per second at 300x180, too slow and too small for play; WHAMM trades data scope (one level section) for speed.", "Needs years of recorded human play with consent terms for a single game; few studios hold such datasets.", "Context of under a second (WHAMM) means objects out of view are forgotten.", "Counts, damage, and other exact quantities are unreliable in generated frames." ], "maturity": "prototype", "mechanism": "Frames are encoded into discrete image tokens and controller inputs into action tokens; a decoder-only transformer models the interleaved sequence, so it predicts both how the world evolves and what a human player is likely to do next. Prompting with a few frames yields many plausible continuations (diversity); a designer can edit a frame, such as inserting a character, and the model carries the edit forward (persistency). WHAM-1.6B produces about one frame per second at 300x180. WHAMM replaces token-by-token decoding with MaskGIT-style parallel prediction (a 500M backbone plus a 250M refiner) to reach more than 10 fps at 640x360 on Quake II.", "name": "Muse World and Human Action Model (WHAM)", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "compute-cost", "controllability", "labor", "player-trust" ], "runtime": [ "design-assist", "server-realtime" ], "sources": [ "src-kanervisto-2025-wham", "src-microsoft-2025-muse-blog", "src-microsoft-2025-whamm" ], "status": "sourced", "summary": "Microsoft Research's WHAM is a transformer trained on seven years of Bleeding Edge play that generates both future frames and the player's controller actions, aimed at gameplay ideation, with a real-time Quake II follow-up (WHAMM).", "trl": 6, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 4, "families_affected": { "physics-materials": 2, "space-construction": 1, "world-generation": 1 }, "incoming": { "applied_in": [ "learned-material-surrogates", "open-ended-evolution-sandbox" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "falling-sand-materials", "note": "Local per-cell material rules can be learned from examples of desired behavior instead of hand-coded; DiffLogic CA recovered Game of Life exactly as a compact logic circuit cheap enough for per-pixel use." }, { "effect": "enables-new", "id": "voxel-destruction", "note": "Regenerating NCA regrow damaged shapes and 3D NCA regrow parts of machines, enabling destructible structures and organisms that heal from local rules rather than respawning." }, { "effect": "enables-new", "id": "procedural-build-assist", "note": "A seed block placed by the player can grow into a castle or tree over time under a learned 3D rule, a growth-based alternative to instant procedural detailing." }, { "effect": "replaces-substrate", "id": "procedural-terrain", "note": "Tile maps and levels can be grown by NCA generators selected by quality-diversity search for solvability and diversity, replacing hand-tuned noise and placement rules." } ], "aliases": [ "NCA", "Growing NCA", "DiffLogic CA", "differentiable logic cellular automata", "self-organising textures" ], "category": "artificial-life-research", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "3-5y", "id": "neural-cellular-automata", "key_results": [ { "confidence": "high", "sources": [ "src-mordvintsev-2020-growing-nca" ], "text": "Growing NCA grew an image from one seed cell with a shared update network of about 8,000 parameters; models trained on damaged samples regrew erased regions, including damage types not seen in training." }, { "confidence": "high", "sources": [ "src-mordvintsev-2021-nca-textures" ], "text": "NCA texture models with 5,868 parameters synthesized textures interactively in real time in a WebGL demo." }, { "confidence": "high", "sources": [ "src-sudhakaran-2021-3d-nca-minecraft" ], "text": "3D NCA grew Minecraft castles, apartment blocks and trees, some over 3,000 blocks, and regrew parts of simple functional machines." }, { "confidence": "high", "sources": [ "src-earle-2021-nca-level-generation" ], "text": "Quality-diversity search (CMA-ME) produced diverse NCA level generators for maze, Sokoban and Zelda domains, often meeting solvability criteria, and explored level space better than a CPPN baseline." }, { "confidence": "high", "sources": [ "src-miotti-2025-difflogic-ca" ], "text": "DiffLogic CA (2025) learned Conway's Game of Life from all 512 3x3 configurations as a 336-gate binary circuit and grew patterns such as a lizard with 577 gates, with asynchronous and self-healing variants." } ], "limitations": [ "Each trained rule targets one pattern, texture or structure family; changing the target requires retraining.", "Growth unfolds over many steps, so getting a stable result has a latency that instant procedural methods do not have.", "Learned rules are opaque to designers and players, apart from small logic-gate circuits, which makes failures hard to diagnose.", "Demonstrations are small 2D images, 3D structures of a few thousand blocks, and simple tile domains; robustness under player interference at game scale is not shown." ], "maturity": "research", "mechanism": "Each grid cell holds a state vector (16 channels in Growing NCA: RGB, an alpha channel that marks living cells, and hidden signaling channels). Every step a cell perceives its 3x3 neighborhood through fixed Sobel gradient filters, a shared network of about 8,000 parameters maps the perception vector to a state increment, and a random mask applies the update to roughly half the cells, so no global clock is needed. Cells with no living neighbor are zeroed. The rule is trained by backpropagating through many steps toward a target image or style, with damaged samples in the training pool to teach regeneration. Variants use 3D convolutions for voxel structures, quality-diversity search instead of gradients for game levels, or differentiable logic gates that compile to a discrete binary circuit (DiffLogic CA).", "name": "Neural cellular automata", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "controllability", "legibility", "evaluation" ], "runtime": [ "client-realtime", "offline-authoring" ], "sources": [ "src-mordvintsev-2020-growing-nca", "src-mordvintsev-2021-nca-textures", "src-sudhakaran-2021-3d-nca-minecraft", "src-earle-2021-nca-level-generation", "src-miotti-2025-difflogic-ca" ], "status": "sourced", "summary": "Cellular automata whose local update rule is a small learned network or logic circuit, trained so that cells grow, maintain, and regenerate a target pattern, texture, structure, or level from a seed.", "trl": 4, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 3, "families_affected": { "control-automation": 1, "environment-ecology": 1, "space-construction": 1 }, "incoming": { "applied_in": [ "symbolic-sim-neural-render" ] } }, "affects": [ { "effect": "accelerates", "id": "embodied-skill-control", "note": "Upscaling and frame generation raise displayed frame rates in GPU-heavy flight and driving sims for the same render cost; generated frames do not sample new input, so frame generation adds latency that hurts precise hand-flying or driving." }, { "effect": "deepens", "id": "day-night-cycle", "note": "Neural denoising and the neural radiance cache make dynamic multi-bounce lighting affordable, so player-built interiors and towns can be lit correctly across a full diurnal cycle where baked lighting cannot follow player construction." }, { "effect": "deepens", "id": "decoration-beauty", "note": "Up to 7x VRAM savings from neural texture compression and cheaper layered materials let more distinct high-detail decor assets be resident at once, raising the variety a beauty-driven builder can show on screen." } ], "aliases": [ "DLSS", "DLSS 4", "DLSS 5", "neural shaders", "RTX Neural Materials", "neural texture compression", "neural radiance cache" ], "category": "neural-rendering-3d", "deployed_in": [ "microsoft-flight-simulator-2020", "cities-skylines-2" ], "determinism_impact": "preserves", "horizon": "now", "id": "neural-rendering-shaders", "key_results": [ { "confidence": "high", "sources": [ "src-nvidia-2025-dlss-4" ], "text": "DLSS 4 (January 2025) introduced Multi Frame Generation of up to three extra frames per rendered frame on RTX 50 GPUs and transformer models for super resolution and ray reconstruction, with 75 games and apps at launch and up to 8x frame rate over brute-force rendering." }, { "confidence": "high", "sources": [ "src-nvidia-2025-rtx-neural-rendering" ], "text": "RTX neural shaders (January 2025) included neural texture compression using up to 7x less VRAM than block-compressed textures at equal quality, neural materials up to 5x faster, and a neural radiance cache for multi-bounce indirect lighting, with DirectX cooperative vector support announced." }, { "confidence": "high", "sources": [ "src-nvidia-2026-dlss-4-5" ], "text": "By January 2026 over 250 titles supported Multi Frame Generation and over 400 games and apps supported DLSS Super Resolution; DLSS 4.5 added a 6X mode generating up to five frames per rendered frame." }, { "confidence": "high", "sources": [ "src-nvidia-2026-dlss-5" ], "text": "DLSS 5 launched on 2026-09-03 in NBA 2K27 as a deterministic one-frame-in, one-frame-out neural pass that adds lighting and material detail, with structure and tone intensity controls and semantic masks for developers." }, { "confidence": "medium", "sources": [ "src-wikipedia-deep-learning-super-sampling" ], "text": "DLSS 5's reveal drew largely negative reaction from players and developers over altered faces and art direction, and a reported 50 to 60 percent frame-rate cost; frame generation in general adds input latency and can introduce artifacts." }, { "confidence": "medium", "sources": [ "src-pcgamingwiki-2026-msfs-2020", "src-pcgamingwiki-2026-cities-skylines-2" ], "text": "Simulation titles use DLSS: Microsoft Flight Simulator (2020) supports DLSS upscaling and DLSS Frame Generation, and Cities: Skylines II supports DLSS upscaling." } ], "limitations": [ "Changes presentation, not simulation: CPU-bound simulation games (agent counts, pathfinding, economy ticks) gain nothing from GPU-side neural rendering.", "Frame generation adds input latency and artifacts; DLSS 5's final pass reportedly costs 50 to 60 percent of frame rate and can alter art direction.", "Hardware-gated: Multi Frame Generation and DLSS 5 require RTX 50-series GPUs at launch, and neural shader features depend on vendor SDKs and cooperative vector support." ], "maturity": "shipped-mainstream", "mechanism": "DLSS Super Resolution reconstructs a high-resolution frame from a lower-resolution render plus motion vectors with a transformer network (from DLSS 4); Multi Frame Generation predicts up to three (DLSS 4) or five (DLSS 4.5) frames between rendered frames; Ray Reconstruction replaces hand-tuned denoisers. Neural shaders embed small MLPs inside shader code and run them on tensor units: neural texture compression decodes textures from learned latents, neural materials evaluate layered BRDFs, and the neural radiance cache learns multi-bounce indirect light online. DirectX cooperative vectors expose this to cross-vendor shaders. DLSS 5 adds a final network pass that adds lighting and material detail from color and motion vectors, with developer intensity controls and masks.", "name": "Neural rendering in the graphics pipeline", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "player-trust", "latency", "compute-cost", "controllability" ], "runtime": [ "client-realtime" ], "sources": [ "src-nvidia-2025-dlss-4", "src-nvidia-2025-rtx-neural-rendering", "src-nvidia-2026-dlss-4-5", "src-nvidia-2026-dlss-5", "src-wikipedia-deep-learning-super-sampling", "src-pcgamingwiki-2026-msfs-2020", "src-pcgamingwiki-2026-cities-skylines-2" ], "status": "sourced", "summary": "Small neural networks inside the renderer upscale frames, generate extra frames, denoise ray tracing, decode compressed textures and layered materials, and since DLSS 5 re-light the final frame, all without touching simulation state.", "trl": 9, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 2, "control-automation": 1, "information-legibility": 1, "social-relations": 1 }, "incoming": { "applied_in": [ "intent-level-orders", "living-workforce-factories", "llm-society-terrarium", "local-first-ai-sims", "salience-gated-llm-agents", "scientists-not-tech-trees", "sim-instructor-coach", "teachable-sim-companion", "trainable-creatures-revival" ] } }, "affects": [ { "effect": "accelerates", "id": "natural-language-interaction", "note": "Free-text and voice input to characters runs on the player's GPU, removing per-token server cost and network round trips, so open conversation can be offered to every player of a premium game." }, { "effect": "replaces-substrate", "id": "utility-action-selection", "note": "Smart Zois pick actions with a small model that reads needs, surroundings, and inner thoughts rather than only scored need curves, while the game's action set stays the same." }, { "effect": "deepens", "id": "memories-thoughts", "note": "Characters generate text thoughts about events and reflect on past decisions, giving each agent a running inner monologue rather than a list of authored mood modifiers." }, { "effect": "replaces-substrate", "id": "advisors", "note": "Total War: PHARAOH's ACE advisor answers questions about the game's systems from a small on-device model connected to game data, instead of firing authored lines when indicators cross thresholds." }, { "effect": "enables-new", "id": "unit-commands", "note": "A co-playable teammate such as PUBG Ally takes spoken orders about movement, looting, and strategy mid-match and acts on them, a command channel that needs no menus or ping wheel." } ], "aliases": [ "NVIDIA ACE", "Smart Zoi", "PUBG Ally", "co-playable characters", "SLM NPCs", "quantized local LLMs" ], "builds_on": [], "category": "inference-efficiency", "deployed_in": [ "inzoi" ], "determinism_impact": "mitigable", "horizon": "now", "id": "on-device-language-models", "key_results": [ { "confidence": "high", "sources": [ "src-sreenivas-2024-minitron", "src-nvidia-2025-ace-companions" ], "text": "Minitron pruning and distillation compressed Mistral NeMo 12B to an 8B model and Llama 3.1 8B to 4B, and NVIDIA's ACE character stack lists Minitron 8B, 4B, and 2B instruct models, the 2B fitting in about 1.5 GB of VRAM." }, { "confidence": "high", "sources": [ "src-frantar-2022-gptq", "src-lin-2023-awq" ], "text": "GPTQ quantizes weights to 3-4 bits with minimal accuracy loss and generates 3.25-4.5x faster than FP16; AWQ's 4-bit runtime is over 3x faster than FP16 on desktop and mobile GPUs." }, { "confidence": "high", "sources": [ "src-steam-2025-inzoi", "src-nvidia-2025-gamescom-ace" ], "text": "inZOI (early access March 2025) ships an on-device SLM through which typed prompts influence Zois' actions and thoughts; an August 2025 update made Smart Zois observe surroundings and react to events better and align actions with inner thoughts." }, { "confidence": "high", "sources": [ "src-krafton-2026-pubg-ally-beta" ], "text": "KRAFTON ran a public PUBG Ally beta (Ally Duo, 17 June to 1 July 2026) in which an AI teammate takes voice instructions via an on-device small language model built on NVIDIA ACE with speech-to-text and text-to-speech." }, { "confidence": "high", "sources": [ "src-nvidia-2026-nvigi-core" ], "text": "NVIDIA's in-game inference SDK requires an RTX 30-series class GPU or newer with at least 8 GB of VRAM (12 GB recommended), and some plugins need RTX 40-series." }, { "confidence": "low", "sources": [ "src-hiddendoor-2026-homepage" ], "text": "Hidden Door reports that a 1.7B model tuned on real play beat a 120B general model on one of its core narrative tasks." } ], "limitations": [ "Hardware floor: the inference SDK needs an RTX 30-series class GPU with 8 GB of VRAM minimum, which excludes many players and consoles, and the model competes with rendering for VRAM and GPU time.", "Small models are less capable at multi-step reasoning, so vendors offload strategic planning to cloud models, reintroducing cost and latency.", "Quality and safety filters must also run locally; local weights can be extracted or modified by players.", "Player reception risk: inZOI drew negative user reviews in May 2025 after Krafton announced its generative AI use.", "Exact model sizes and latency in shipped titles are mostly undisclosed." ], "maturity": "shipped-niche", "mechanism": "A 12B or 8B model is pruned in depth and width and distilled into an 8B, 4B, or 2B student (Minitron), then weights are quantized to 3-4 bits with second-order (GPTQ) or activation-aware (AWQ) methods, shrinking memory so a 2B model fits in about 1.5 GB of VRAM. An in-game inference SDK runs the models in-process and schedules them alongside graphics on the same GPU. A character loop chains on-device speech recognition, the small model (perceive, decide, act, reflect), text-to-speech, and facial animation, with an in-process embedding model for memory retrieval; slow strategic planning can be offloaded to a larger cloud model while fast decisions stay local. Fine-tuning a small model on a narrow game task can beat much larger general models on that task.", "name": "On-device small language models for characters", "provenance": { "created": "2026-10-24-techniques-llm-agents" }, "risks": [ "consistency", "latency", "controllability", "content-safety", "player-trust" ], "runtime": [ "client-realtime" ], "sources": [ "src-sreenivas-2024-minitron", "src-frantar-2022-gptq", "src-lin-2023-awq", "src-nvidia-2025-ace-companions", "src-nvidia-2025-gamescom-ace", "src-nvidia-2026-ace-developer-page", "src-nvidia-2026-nvigi-core", "src-steam-2025-inzoi", "src-wikipedia-inzoi", "src-krafton-2026-pubg-ally-beta", "src-hiddendoor-2026-homepage" ], "status": "sourced", "summary": "Pruned, distilled, and 4-bit quantized language, speech, and embedding models that run on the player's GPU beside rendering to drive character decisions, dialogue, and voice without per-token server cost.", "trl": 7, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 1, "narrative-generation": 1, "progression-goals": 1, "risk-failure": 1, "world-generation": 1 }, "incoming": { "applied_in": [ "sims-as-agent-benchmarks" ] } }, "affects": [ { "effect": "enables-new", "id": "scenario-objectives", "note": "Objectives stop being a fixed authored list: a model writes each task and its success check as code, admitted only if it is neither already solved nor out of reach, so a sandbox can keep issuing new challenges." }, { "effect": "deepens", "id": "procedural-terrain", "note": "Terrain parameters are mutated against measured agent or player performance, as POET did with walker terrain roughness, gaps, and stumps, so maps grow harder along the axes a player has already mastered." }, { "effect": "deepens", "id": "adaptive-threat-scaling", "note": "Challenge scales with measured learning progress per task family rather than with wealth or time survived, using a band that admits only tasks the current agent or player can partly solve." }, { "effect": "deepens", "id": "creature-learning", "note": "A creature or companion can improve through self-proposed practice tasks with model-estimated rewards, the loop SIMA 2 used, rather than only through explicit player reward and punishment." }, { "effect": "deepens", "id": "villager-requests", "note": "Requests can be generated beyond fixed templates, with completion conditions written as generated code over simulation state and an archive of past requests used to avoid repeats." } ], "aliases": [ "POET", "Enhanced POET", "OMNI-EPIC", "SIMA 2 self-improvement", "automatic curriculum generation" ], "builds_on": [ "genie-world-models", "generalist-game-agents" ], "category": "reinforcement-learning", "determinism_impact": "mitigable", "horizon": "3-5y", "id": "open-ended-environment-generation", "key_results": [ { "confidence": "high", "sources": [ "src-wang-2019-poet" ], "text": "POET (2019) co-generated environment challenges and agents, producing behaviors that solved challenges direct optimization and a direct-path curriculum could not, and transfer between environments was essential." }, { "confidence": "high", "sources": [ "src-wang-2020-enhanced-poet" ], "text": "Enhanced POET (2020) added a domain-general measure of meaningful novelty, an efficient transfer heuristic, a more flexible environment encoding, and a measure of ongoing open-ended innovation." }, { "confidence": "medium", "sources": [ "src-faldor-2024-omni-epic" ], "text": "OMNI-EPIC (2024) used foundation models to write environment and reward code for tasks judged both learnable and interesting, kept proposing new tasks, and adjusted difficulty to the agent's learning progress." }, { "confidence": "medium", "sources": [ "src-deepmind-2025-sima-2-blog" ], "text": "SIMA 2 (November 2025) learned new games through self-directed play from Gemini-proposed tasks and estimated rewards without new human data, improved on previously failed tasks across generations including in Genie 3 worlds, and remained a limited research preview." } ], "limitations": [ "Cost is dominated by training or evaluating an agent on every candidate task, so the loop runs offline and is too slow to respond to a live player directly.", "Interestingness and success are judged by foundation models or generated code that can be wrong, and generated reward code can be satisfied by unintended behavior.", "SIMA 2 still struggles with very long-horizon tasks, short memory, precise low-level control, and robust visual understanding, which limits the tasks it can practice.", "Worlds generated by video world models are approximate and short-horizon rather than rules-exact simulations, so success checks inside them are approximate." ], "maturity": "research", "mechanism": "POET keeps a population of environment-agent pairs. It mutates environment parameters (for a 2D walker: roughness, gap width, stump height), admits a child environment only if current agents score within a band that is neither too easy nor too hard, optimizes each agent in its paired environment, and periodically tests whether an agent from another pair does better there, transferring it if so. Enhanced POET adds a novelty measure based on how agents rank on an environment and a more flexible environment encoding. OMNI-EPIC replaces parameter mutation with a foundation model that writes environment and reward code for a physics simulator, conditioned on an archive of solved and failed tasks; a model of interestingness rejects near-duplicates, an RL agent trains on each task, and a success check decides whether the task enters the archive. SIMA 2 has Gemini propose tasks and estimate rewards in commercial games and in Genie 3 generated worlds, then trains later agents on its own experience. The shared loop is: generate a task, check it is learnable, train, archive, repeat.", "name": "Open-ended co-generation of environments and agents", "provenance": { "created": "2026-10-24-techniques-rl-design" }, "risks": [ "controllability", "evaluation", "compute-cost" ], "runtime": [ "offline-authoring", "server-realtime" ], "sources": [ "src-wang-2019-poet", "src-wang-2020-enhanced-poet", "src-faldor-2024-omni-epic", "src-deepmind-2025-sima-2-blog" ], "status": "sourced", "summary": "Systems that keep generating new environments or tasks at the edge of an agent's ability and train agents on them, from POET's co-evolved terrains to OMNI-EPIC's foundation-model-written task code and SIMA 2 practicing in Genie 3 worlds.", "trl": 4, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 6, "families_affected": { "control-automation": 1, "multiplayer-persistence": 1, "narrative-generation": 1, "physics-materials": 2, "world-generation": 1 }, "incoming": { "applied_in": [ "dream-replay-layer" ] } }, "affects": [ { "effect": "replaces-substrate", "id": "vehicle-dynamics", "note": "Matrix-Game 2.0 was fine-tuned on 574 hours of GTA driving; car handling is reproduced from video rather than tire and suspension models, with no guarantee of consistent grip or mass." }, { "effect": "replaces-substrate", "id": "direct-avatar", "note": "Walking, turning, and jumping are rendered from a unified camera-action space; GameCraft's action space lacks shooting, throwing, and explosions." }, { "effect": "enables-new", "id": "procedural-terrain", "note": "A world starts from any image, sketch, or photo (Mirage 2, GameCraft); pose-indexed memory in Matrix-Game 3.0 and WorldPlay lets revisited areas stay consistent for about a minute, still far short of a save file." }, { "effect": "enables-new", "id": "generative-content", "note": "Typed text during play adds or restyles objects and scenes (Yan editing, Mirage 2 commands, Odyssey-2 prompts), so content is requested at runtime rather than authored." }, { "effect": "enables-new", "id": "content-sharing", "note": "Mirage 2 lets players save and share generated worlds; what is shared is a prompt and seed image for a model rather than an asset bundle, so the receiver needs the same hosted model." }, { "effect": "replaces-substrate", "id": "rigid-body-physics", "note": "Contact and motion responses come from video priors; out-of-domain views degrade (Matrix-Game 2.0 reports over-saturation under sustained upward or forward camera motion)." } ], "aliases": [ "Matrix-Game", "Hunyuan-GameCraft", "WorldPlay", "Yan", "Mirage", "Odyssey-2" ], "builds_on": [ "diffusion-game-engines" ], "category": "generative-world-model", "determinism_impact": "breaks", "horizon": "3-5y", "id": "open-interactive-video-models", "key_results": [ { "confidence": "high", "sources": [ "src-he-2025-matrix-game-2" ], "text": "Matrix-Game 2.0 (Skywork, August 2025) streams 352x640 video at 25 fps on one H100 with about 1.8B parameters trained on about 1,200 hours of Unreal Engine, GTA driving, Minecraft, and Temple Run data, and releases weights and code; it has no explicit memory." }, { "confidence": "high", "sources": [ "src-li-2025-hunyuan-gamecraft", "src-tencent-2025-hunyuan-gamecraft-weights" ], "text": "Hunyuan-GameCraft trained on over 1M recordings from more than 100 AAA titles; distillation raised speed from 0.25 to 6.6 fps with under 5 seconds latency per action, and weights released in August 2025 need at least 24 GB of GPU memory (80 GB recommended)." }, { "confidence": "medium", "sources": [ "src-tencent-2025-yan" ], "text": "Yan (Tencent, August 2025) reports 1080p at 50 to 60 fps for its simulation module trained on the game Yuan Meng Star and an editing module that separates interactive mechanics simulation from visual rendering so text can restyle the world during play." }, { "confidence": "high", "sources": [ "src-wang-2026-matrix-game-3", "src-sun-2025-worldplay" ], "text": "Matrix-Game 3.0 (April 2026) reaches up to 40 fps at 720p with a 5B model on an 8+1 GPU setup and keeps minute-long consistency by retrieving memory frames by camera pose; WorldPlay (Tencent Hunyuan) streams 720p at 24 fps with a reconstituted context memory for long-term geometric consistency." }, { "confidence": "medium", "sources": [ "src-thedecoder-2025-mirage-2", "src-odyssey-2025-odyssey-2" ], "text": "Mirage 2 (Dynamics Lab, August 2025) lets players turn photos or sketches into browser-playable worlds and change them by typing commands, with saved and shared worlds, but struggles with precise controls and visual stability; Odyssey-2 (October 2025) generates each frame in under 50 ms from text prompts." } ], "limitations": [ "Resolution and speed trade against hardware: Matrix-Game 2.0 is 352x640 on one H100, Matrix-Game 3.0's 720p 40 fps needs 9 GPUs, and GameCraft runs at 6.6 fps after distillation.", "Consistency horizon is about a minute even with retrieval memory; earlier open models have no explicit memory at all.", "Training data includes recordings of commercial games (GTA V, 100+ AAA titles), so output provenance is contested.", "Stronger history conditioning improves consistency but reduces responsiveness to new actions (GameCraft ablation).", "Action spaces cover camera movement and basic locomotion; game verbs such as shooting, building, or trading are not modeled." ], "maturity": "prototype", "mechanism": "A pretrained text- or image-to-video diffusion transformer (SkyReels-V2 for Matrix-Game 2.0, HunyuanVideo for GameCraft) gets action modules: keyboard and mouse mapped to camera translation and rotation, injected through cross-attention, token addition, or Plucker camera embeddings. Training uses gameplay recordings from Unreal Engine scenes, GTA V, Minecraft, Temple Run, or 100+ AAA titles. The bidirectional model is then distilled into a causal few-step generator (Self-Forcing with distribution matching, or phased consistency models) that streams chunks with a rolling KV cache. 2026 systems add memory by retrieving past frames whose camera pose overlaps the current view (Matrix-Game 3.0, WorldPlay). Yan, Mirage 2, and Odyssey-2 accept typed text during play to edit the world.", "name": "Open real-time interactive video world models", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "compute-cost", "ip-provenance", "controllability", "latency" ], "runtime": [ "server-realtime" ], "sources": [ "src-he-2025-matrix-game-2", "src-li-2025-hunyuan-gamecraft", "src-tencent-2025-hunyuan-gamecraft-weights", "src-tencent-2025-yan", "src-wang-2026-matrix-game-3", "src-sun-2025-worldplay", "src-thedecoder-2025-mirage-2", "src-odyssey-2025-odyssey-2" ], "status": "sourced", "summary": "Open and startup interactive video models (Matrix-Game, Hunyuan-GameCraft, WorldPlay, Yan, Mirage, Odyssey) adapt pretrained video diffusion transformers to keyboard and mouse control and stream frames in real time, adding pose-indexed memory and text edits during play.", "trl": 6, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 3, "families_affected": { "agents-behavior": 1, "physics-materials": 2 } }, "affects": [ { "effect": "replaces-substrate", "id": "vehicle-dynamics", "note": "Cosmos is positioned for autonomous-vehicle training and closed-loop evaluation; a driving model post-trained from it predicts scene video under driving inputs with no tire or suspension model, which suits traffic scenarios more than the handling fidelity a racing sim needs." }, { "effect": "replaces-substrate", "id": "rigid-body-physics", "note": "Learned video dynamics could stand in for a physics engine in open-ended scenes, but near-chance IntPhys 2 scores show current models do not reliably reject impossible motion." }, { "effect": "replaces-substrate", "id": "utility-action-selection", "note": "V-JEPA 2-AC scores candidate action sequences by predicted distance to a goal image in latent space, a learned alternative to hand-written utility curves for agents whose goals can be pictured." } ], "aliases": [ "NVIDIA Cosmos", "V-JEPA 2", "world foundation models" ], "builds_on": [ "massively-parallel-simulation" ], "category": "generative-world-model", "determinism_impact": "breaks", "horizon": "5y-plus", "id": "physical-world-foundation-models", "key_results": [ { "confidence": "high", "sources": [ "src-nvidia-2025-cosmos-wfm" ], "text": "Cosmos (January 2025) pretrained diffusion and autoregressive world foundation models on about 100M clips curated from about 20M hours of video and released them as open weights with permissive licenses." }, { "confidence": "high", "sources": [ "src-nvidia-2026-cosmos-page" ], "text": "Cosmos 3 ships Super 64B, Nano 16B, and Edge 4B models under the OpenMDW1.1 license, positioned for autonomous vehicle and robot synthetic data, closed-loop evaluation, and robot policy training, with no mention of games." }, { "confidence": "high", "sources": [ "src-assran-2025-v-jepa-2" ], "text": "V-JEPA 2-AC, post-trained on under 62 hours of DROID robot video, performed zero-shot pick-and-place with image goals on Franka arms in two labs without task-specific data or reward." }, { "confidence": "high", "sources": [ "src-meta-2025-v-jepa-2-blog" ], "text": "On Meta's IntPhys 2 benchmark of physical plausibility, current video models including V-JEPA 2 score at or near chance while humans score 85 to 95 percent across the released benchmarks." } ], "limitations": [ "Designed for robots and vehicles; no game integration or game-specific evaluation exists in the cited sources.", "Physical plausibility is unreliable: video models are at or near chance on IntPhys 2.", "Large models (up to 64B in Cosmos 3) are far from per-player real-time budgets.", "V-JEPA 2 predicts at a single time scale, limiting long-horizon planning (Meta's stated limitation)." ], "maturity": "research", "mechanism": "Cosmos curates about 20M hours of raw video into about 100M clips, compresses them with causal video tokenizers, and pretrains diffusion (Text2World, Video2World) and autoregressive world models that developers post-train into action- or camera-conditioned simulators; Cosmos 3 (2026) is an omni family at 4B, 16B, and 64B parameters. V-JEPA 2 instead predicts masked video in a learned embedding space, not pixels, from over 1M hours of video; its action-conditioned variant V-JEPA 2-AC is post-trained on under 62 hours of robot video and plans by searching for action sequences whose predicted embedding approaches a goal image.", "name": "Physical-world foundation models", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "consistency", "compute-cost", "evaluation", "controllability" ], "runtime": [ "offline-authoring", "server-realtime" ], "sources": [ "src-nvidia-2025-cosmos-wfm", "src-nvidia-2026-cosmos-page", "src-assran-2025-v-jepa-2", "src-meta-2025-v-jepa-2-blog" ], "status": "sourced", "summary": "Video-pretrained models of physical dynamics, NVIDIA Cosmos for generating action-conditioned training video and Meta's V-JEPA 2 for latent prediction and planning, target robotics and driving rather than games.", "trl": 3, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 4, "families_affected": { "agents-behavior": 1, "biology-health": 1, "control-automation": 1, "physics-materials": 1 }, "incoming": { "applied_in": [ "neural-npc-motion" ] } }, "affects": [ { "effect": "cuts-authoring-cost", "id": "smart-objects", "note": "When an object advertises an interaction (sit, lie, lean), a goal-conditioned controller can perform it on new furniture geometry without a bespoke animation per object, as MaskedMimic shows for chairs, stools, and sofas." }, { "effect": "deepens", "id": "direct-avatar", "note": "The controlled character becomes a physical body that adapts footing to terrain and reacts to pushes, instead of playing canned clips that slide over geometry." }, { "effect": "deepens", "id": "rigid-body-physics", "note": "Characters join the rigid-body simulation as actuated bodies, so carried, pushed, and collided objects affect them and vice versa (InterMimic human-object interaction)." }, { "effect": "enables-new", "id": "injury-body-model", "note": "With a simulated body, a wound could change gait by limiting joint torque instead of swapping in an authored limp; the cited work does not demonstrate injury-conditioned motion, so this remains a projection." } ], "aliases": [ "MaskedMimic", "InterMimic", "learned physics controllers", "RL motion imitation" ], "builds_on": [ "massively-parallel-simulation" ], "category": "animation-motion", "determinism_impact": "mitigable", "horizon": "3-5y", "id": "physics-based-character-control", "key_results": [ { "confidence": "high", "sources": [ "src-tessler-2024-maskedmimic" ], "text": "MaskedMimic (SIGGRAPH Asia 2024) formulates physics-based control as motion inpainting: one controller generates motion from keyframes, text, VR head or hand tracking, paths, or objects, and interacts with chairs, stools, tables, and sofas while transferring to irregular terrain; code is released as ProtoMotions." }, { "confidence": "medium", "sources": [ "src-xu-2025-intermimic" ], "text": "InterMimic (2025) trains a single policy for physics-based human-object interaction from hours of imperfect motion capture by distilling subject-specific teacher policies and refining with RL; the authors report zero-shot generalization and compatibility with kinematic motion generators." } ], "limitations": [ "Every character needs a physics simulation step plus policy inference, which is costly at the hundreds-of-agents scale colony and management sims use; no crowd-scale runtime figures are published in the cited work.", "Training requires large mocap datasets and GPU-parallel RL; retargeting to non-human or stylized rigs requires retraining.", "Physical controllers can fall or jitter in edge cases, which is hard to QA across open-ended situations." ], "maturity": "research", "mechanism": "A ragdoll with joint actuators is simulated in a physics engine; a policy network outputs PD targets each step and is trained with RL in massively parallel GPU simulation to track motion capture clips. MaskedMimic first trains a full-body tracker, then distills it into a conditional generative student that receives a masked goal: any subset of future joint positions, a text description, a head and hand signal from VR, a path, or a target object. The same controller therefore handles locomotion, steering, and sitting on varied furniture and transfers flat-ground motions to rough terrain. InterMimic distills per-subject teachers into one student refined by RL for whole-body human-object interaction from imperfect mocap.", "name": "Learned physics-based character control", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "compute-cost", "controllability", "evaluation" ], "runtime": [ "offline-authoring", "client-realtime" ], "sources": [ "src-tessler-2024-maskedmimic", "src-xu-2025-intermimic" ], "status": "sourced", "summary": "Neural policies trained with reinforcement learning drive simulated humanoids by joint torques to imitate motion capture, and MaskedMimic-style controllers accept partial goals such as keyframes, text, or an object to sit on.", "trl": 4, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 6, "families_affected": { "economy-finance": 1, "progression-goals": 1, "resources-production": 1, "space-construction": 2, "world-generation": 1 }, "incoming": { "applied_in": [ "generated-recipe-graphs", "learned-build-assist", "mechanic-discovery-search", "open-ended-evolution-sandbox", "parallel-autobalance", "qd-scenario-libraries" ], "built_upon_by": [ "automated-rule-generation" ] } }, "affects": [ { "effect": "tunes", "id": "modular-assembly", "note": "Mapping the build space by behavior descriptors, as was done for Hearthstone decks, shows the best design in each style niche, exposing dominant part combinations and empty niches before players find them." }, { "effect": "tunes", "id": "design-performance-rating", "note": "Searching for the top-scoring design in every descriptor cell stress-tests the rating formula and surfaces degenerate designs that score well, such as rides or bridges that game the metric." }, { "effect": "deepens", "id": "procedural-terrain", "note": "Instead of one seed-to-map generator, a game can ship an archive of maps spanning designer axes, such as resource density by chokepoint count, each verified by simulation and served to players by trait." }, { "effect": "deepens", "id": "rival-firms", "note": "An archive of competitor strategies, each the strongest found for its niche (aggressive expander, low-price discounter), gives AI rivals distinct styles without hand-tuning each persona." }, { "effect": "cuts-authoring-cost", "id": "scenario-objectives", "note": "Scenario maps and starting conditions can be generated across a difficulty-by-style grid and filtered by agent playthroughs, replacing hand-built scenario sets with a curated archive." }, { "effect": "tunes", "id": "machine-boosting", "note": "Searches over module and beacon layouts expose dominant boosting configurations that collapse design choice." } ], "aliases": [ "MAP-Elites", "CMA-ME", "MESB", "DSAGE", "QDAIF", "illumination search" ], "category": "procedural-generation", "determinism_impact": "preserves", "horizon": "now", "id": "quality-diversity-search", "key_results": [ { "confidence": "high", "sources": [ "src-mouret-2015-map-elites" ], "text": "MAP-Elites (2015) returns a map of high-performing solutions across user-chosen dimensions of variation, revealing trade-offs and often finding a better best solution than standard search." }, { "confidence": "high", "sources": [ "src-gravina-2019-pcg-quality-diversity" ], "text": "Gravina et al. (IEEE CoG 2019) framed procedural content generation as a quality-diversity problem and reviewed early applications to content generation and game playing." }, { "confidence": "high", "sources": [ "src-fontaine-2019-hearthstone-mesb" ], "text": "MAP-Elites with sliding boundaries mapped Hearthstone deck spaces along play-style dimensions, found diverse strong decks, and was proposed as an aid to rebalancing." }, { "confidence": "high", "sources": [ "src-fontaine-2020-mario-gan-qd" ], "text": "CMA-ME searching a GAN's latent space produced playable Super Mario Bros levels spanning designer-chosen measures such as enemy and obstacle counts, and a user study found those measures changed perceived difficulty." }, { "confidence": "medium", "sources": [ "src-bhatt-2022-dsage" ], "text": "DSAGE (2022) used a deep surrogate to predict agent behavior in generated environments and beat earlier QD environment-generation methods at finding environments that elicit diverse agent behaviors." }, { "confidence": "medium", "sources": [ "src-bradley-2023-qdaif" ], "text": "QDAIF (2023) used language models to generate variations and to judge quality and diversity of creative text, covering more of the target space than non-QD baselines with reasonable agreement with human raters." } ], "limitations": [ "Every candidate needs a fitness and descriptor evaluation, usually a simulated playthrough, so large archives cost many simulations; surrogate methods such as DSAGE exist to cut that cost.", "Results depend on the chosen descriptors: poor choices give diversity along axes players do not perceive, and the grid grows exponentially with the number of descriptors.", "Fitness measured with AI agents reflects agent skill, not human experience, so cells marked playable or balanced may not match player perception.", "Game applications in the cited work are research prototypes on small games (Mario levels, Hearthstone decks); no documented shipped use in a simulation game was found in this run." ], "maturity": "research", "mechanism": "The designer picks k behavior descriptors (for a level: enemy count and number of jumps; for a deck: mean mana cost and minion count) and discretizes them into a grid. Each cell stores the highest-fitness solution seen with those descriptor values. Each iteration selects elites from the archive, mutates or recombines them, evaluates the child's fitness and descriptors (usually by simulation or an agent playthrough), and inserts it if its cell is empty or it beats the incumbent. The archive shows which trait combinations can be good and how good. Variants slide cell boundaries to follow where solutions cluster (MESB), search a generative model's latent space with CMA-ES emitters (CMA-ME on a Mario GAN), use a learned surrogate to predict agent behavior and skip simulations (DSAGE), or use language models both to propose variations and to judge quality and diversity (QDAIF). GAVEL applies MAP-Elites to game rule sets.", "name": "Quality-diversity search (MAP-Elites) for content and balance", "provenance": { "created": "2026-10-24-techniques-rl-design", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "evaluation", "compute-cost" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-mouret-2015-map-elites", "src-gravina-2019-pcg-quality-diversity", "src-fontaine-2019-hearthstone-mesb", "src-fontaine-2020-mario-gan-qd", "src-bhatt-2022-dsage", "src-bradley-2023-qdaif" ], "status": "sourced", "summary": "Evolutionary search that keeps the best solution found in each cell of a grid of designer-chosen behavior descriptors, returning an archive of diverse high-quality levels, decks, or strategies instead of one optimum.", "trl": 4, "type": "technique", "year": 2015 }, { "_derived": { "affects_count": 2, "families_affected": { "narrative-generation": 1, "space-construction": 1 } }, "affects": [ { "effect": "enables-new", "id": "procedural-build-assist", "note": "Blocky player builds are rendered as detailed structures in a prompted style (for example Roman), a visual-only form of build assist: collision and block state stay coarse while the view is detailed." }, { "effect": "enables-new", "id": "generative-content", "note": "Players or in-world logic change the whole world's look at runtime by text, without new textures or models; the game's systems cannot read the new look, so it carries no gameplay meaning." } ], "aliases": [ "MirageLSD", "Oasis 2.0", "live-stream diffusion", "video-to-video game skin" ], "builds_on": [ "diffusion-game-engines" ], "category": "neural-rendering-3d", "deployed_in": [ "minecraft" ], "determinism_impact": "mitigable", "horizon": "1-2y", "id": "real-time-video-restyling", "key_results": [ { "confidence": "high", "sources": [ "src-decart-2025-miragelsd" ], "text": "MirageLSD responds in under 40 ms per frame at 24 fps, conditioning each frame on recent output, the current input frame, and a prompt; Decart lists limited memory, weak control of specific objects or regions, and structure distortion under extreme style shifts." }, { "confidence": "high", "sources": [ "src-decart-2025-oasis-2" ], "text": "Oasis 2.0 is a Minecraft 1.21.8 mod that restyles the player's world in real time from prompts set by chat command or by in-world command blocks." } ], "limitations": [ "Short frame window and no long-term memory, so restyled objects can change appearance between views.", "Rendered detail can contradict game state (a block that looks like a door but is solid), hurting legibility.", "Cloud-served stream adds network latency and per-player GPU cost; no price per hour is published." ], "maturity": "prototype", "mechanism": "The host game renders normally and streams its frames to a video-to-video diffusion model. Each output frame is conditioned on the current input frame, a window of recently generated frames, and a text prompt; Diffusion Forcing and history augmentation keep the stream from drifting, and distillation, pruning, and custom GPU kernels bring response under 40 ms at 24 fps. Because the model only replaces pixels, collision, inventory, and world state stay in the host engine. Decart's Oasis 2.0 packages this as a Minecraft mod with chat commands and command-block hooks that switch the prompt.", "name": "Real-time generative restyling of game video", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "consistency", "legibility", "latency", "compute-cost", "controllability" ], "runtime": [ "server-realtime" ], "sources": [ "src-decart-2025-miragelsd", "src-decart-2025-oasis-2" ], "status": "sourced", "summary": "A live video-to-video diffusion model repaints each frame of a running game from a text prompt in under 40 ms, leaving the game's own simulation and state untouched.", "trl": 6, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 4, "families_affected": { "control-automation": 1, "social-relations": 2, "world-generation": 1 }, "incoming": { "applied_in": [ "sim-instructor-coach" ] } }, "affects": [ { "effect": "enables-new", "id": "natural-language-interaction", "note": "Characters can hold spoken, interruptible, full-duplex conversations with response onset near human turn-taking, replacing dialogue wheels or typed input; Whispers from the Star is built around this loop." }, { "effect": "deepens", "id": "social-exchanges", "note": "Delivery carries social meaning (whispering, sighing, hesitation, warmth) generated from conversational context, so the same line can land as a flirt, apology, or threat." }, { "effect": "deepens", "id": "unit-commands", "note": "Squad orders can be spoken mid-action and parsed into function calls, as in PUBG Ally's voice-directed teammate, without menus." }, { "effect": "cuts-authoring-cost", "id": "procedural-characters", "note": "Generated characters can get distinct designed or cloned voices and speak unscripted lines without recording sessions, which removes the voice-acting cost that keeps procedural characters silent." } ], "aliases": [ "speech-to-speech", "full-duplex dialogue", "Realtime API", "Moshi", "expressive TTS", "AI NPC voices" ], "builds_on": [], "category": "speech-audio", "claims": [ { "confidence": "high", "sources": [ "src-elevenlabs-2025-eleven-v3" ], "text": "Eleven v3's expressive model was not recommended for real-time conversation at launch because of latency and reliability, so expressive range and low latency were still traded off in mid-2025." }, { "confidence": "low", "sources": [ "src-inworld-2025-death-by-ai" ], "text": "A vendor case study reports Death by AI reached 20 million players in three months with 1.2 billion tokens in its first month after moving LLM and TTS to a gaming-priced provider; the cost claims are unverified." }, { "confidence": "low", "sources": [ "src-inworld-2026-consumer-ai-cost" ], "text": "Inworld's June 2026 price list puts streaming TTS at about $5-25 per million characters and speech-to-text at about $0.10 per hour, and claims only about 3 in 100 consumer AI users ever pay." }, { "confidence": "high", "sources": [ "src-gdc-2026-state-of-industry", "src-gdc-2025-state-of-industry" ], "text": "In GDC's 2026 survey of over 2,300 professionals, 36% used generative AI at work but 52% thought it was harming the industry, up from 30% in 2025, with game design and narrative (63%) and art (64%) among the most negative groups, which bears on adoption of generated character voices and dialogue." } ], "deployed_in": [ "whispers-from-the-star", "where-winds-meet" ], "determinism_impact": "breaks", "horizon": "now", "id": "realtime-voice-models", "key_results": [ { "confidence": "high", "sources": [ "src-defossez-2024-moshi" ], "text": "Moshi reported 160 ms theoretical and 200 ms practical latency for full-duplex spoken dialogue." }, { "confidence": "high", "sources": [ "src-openai-2024-realtime-api", "src-openai-2025-gpt-realtime" ], "text": "OpenAI's Realtime API launched in October 2024 at about $0.06 per minute of audio input and $0.24 per minute of output; gpt-realtime (August 2025) priced audio at $32 and $64 per million input and output tokens and raised function-calling accuracy on ComplexFuncBench from 49.7% to 66.5%." }, { "confidence": "high", "sources": [ "src-sesame-2025-csm" ], "text": "Sesame's CSM reached parity with human recordings in listener preference without conversational context, but listeners preferred the human recordings when 90 seconds of prior context was provided." }, { "confidence": "medium", "sources": [ "src-inworld-2026-tts-1-5" ], "text": "Inworld's TTS-1.5 (January 2026) lists P90 time to first audio under 130 ms (Mini) and under 250 ms (Max) at $15 and $25 per million characters, about $0.015-0.025 per minute of speech." }, { "confidence": "high", "sources": [ "src-steam-2026-whispers-from-the-star", "src-steam-2025-where-winds-meet" ], "text": "Whispers from the Star (August 2025) is built around open-ended voice, video-call, and text conversation with an AI-voiced character created with an actor, and Where Winds Meet (global November 2025) discloses real-time AI NPC text and voice chat." }, { "confidence": "medium", "sources": [ "src-wikipedia-artificial-intelligence-in-video-games" ], "text": "ARC Raiders (November 2025) was criticized and marked down by Eurogamer for generative AI NPC voices, and the SAG-AFTRA video game strike (July 2024 to July 2025) ended with contract terms on consent for AI voice replicas." } ], "limitations": [ "Cost scales with talk time: at 2024 Realtime API rates, an hour split evenly between player and NPC speech is roughly $9 before context re-reads; gpt-realtime rates cut that to roughly $3 (estimates derived from list prices).", "Expressive delivery and low latency trade off: context-conditioned prosody still trails humans, and the most expressive TTS was not suited to real time at launch.", "Voice actor consent and compensation are now contract terms, and players have penalized AI voices in reviews.", "Speech-native models are hard to constrain to game-legal actions; most games keep a text layer between speech and game state.", "Microphone use and online connectivity raise privacy and access barriers (Whispers from the Star lists a broadband connection and an account)." ], "maturity": "shipped-niche", "mechanism": "Two pipelines. Cascaded: streaming speech recognition, then an LLM, then streaming TTS, each optimized for time to first output. Speech-native: the model reads and writes audio tokens from a neural codec. Moshi models the user's and its own speech as parallel streams (full duplex, so it can listen while speaking) and predicts time-aligned text tokens before audio tokens. Hosted realtime APIs keep a persistent WebSocket session with voice activity detection, interruption handling, and function calls that can trigger game actions. Expressive TTS conditions delivery on context: Sesame's CSM feeds prior conversation audio and text to a Llama-style backbone over codec tokens, and Eleven v3 accepts inline tags such as [whispers] or [sighs]. Small ASR and TTS models (hundreds of millions of parameters) can run on device.", "name": "Realtime speech-to-speech and expressive TTS for characters", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-review-fixes" ] }, "risks": [ "determinism", "compute-cost", "latency", "labor", "player-trust", "content-safety" ], "runtime": [ "server-realtime", "client-realtime" ], "sources": [ "src-defossez-2024-moshi", "src-openai-2024-realtime-api", "src-openai-2025-gpt-realtime", "src-sesame-2025-csm", "src-elevenlabs-2025-eleven-v3", "src-inworld-2026-tts-1-5", "src-inworld-2026-consumer-ai-cost", "src-inworld-2025-death-by-ai", "src-steam-2026-whispers-from-the-star", "src-steam-2025-where-winds-meet", "src-wikipedia-artificial-intelligence-in-video-games", "src-krafton-2026-pubg-ally-beta", "src-gdc-2026-state-of-industry", "src-gdc-2025-state-of-industry" ], "status": "sourced", "summary": "Streaming speech recognition, speech-native language models, and context-aware text-to-speech that let characters hold spoken, interruptible conversations with roughly 200-250 ms response onset and generated emotional delivery.", "trl": 8, "type": "technique", "year": 2024 }, { "_derived": { "affects_count": 3, "families_affected": { "control-automation": 1, "world-generation": 2 } }, "affects": [ { "effect": "replaces-substrate", "id": "procedural-terrain", "note": "The world is reconstructed from real imagery instead of generated from a seed and noise; the player-facing role of an explorable map is unchanged, but its content is the real planet, with procedural filling only where imagery is missing." }, { "effect": "deepens", "id": "embodied-skill-control", "note": "Recognizable real buildings and landmarks across the planet make visual navigation by reference to the ground practicable anywhere, extending hand-flying practice beyond hand-modeled airports and cities." }, { "effect": "deepens", "id": "real-world-data-feeds", "note": "Learned classification of satellite imagery turns real-world map data into simulated scenery at planetary scale (Microsoft Flight Simulator)." } ], "aliases": [ "Blackshark.ai", "satellite-to-3D", "planet-scale scenery generation" ], "category": "procedural-generation", "deployed_in": [ "microsoft-flight-simulator-2020", "microsoft-flight-simulator-2024" ], "determinism_impact": "preserves", "horizon": "now", "id": "satellite-world-reconstruction", "key_results": [ { "confidence": "medium", "sources": [ "src-wikipedia-microsoft-flight-simulator-2020" ], "text": "Microsoft Flight Simulator (2020) used Blackshark.ai's AI on Azure to generate 3D buildings, trees, and terrain from Bing Maps data and photogrammetry, covering about 1.5 billion buildings and 2 trillion trees from over two petabytes of streamed map data." }, { "confidence": "high", "sources": [ "src-blackshark-2026-homepage" ], "text": "Blackshark.ai describes its products as turning satellite and aerial imagery into structured models of terrain, buildings, and infrastructure." }, { "confidence": "medium", "sources": [ "src-wikipedia-microsoft-flight-simulator-2024" ], "text": "Microsoft Flight Simulator 2024 carries over the satellite-derived scenery systems of the 2020 game." } ], "limitations": [ "Accuracy is bounded by imagery quality; blurred or pixelated areas fall back to procedural fill.", "Requires cloud streaming of petabyte-scale data and a licensed imagery source.", "Reconstructed buildings are generic from inferred class and height, not modeled individually, except in photogrammetry cities." ], "maturity": "shipped-niche", "mechanism": "Models trained on labeled imagery detect buildings, trees, and land cover in satellite and aerial imagery and map data and extrapolate approximate geometry; offline procedural generation then builds and dresses the buildings and places vegetation, while selected cities use photogrammetry meshes. The reconstruction runs offline in the cloud, and the result is streamed to players on demand as they fly, with procedural filling where imagery is poor.", "name": "ML reconstruction of the planet from aerial imagery", "provenance": { "created": "2026-10-24-techniques-world-models", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "consistency", "ip-provenance", "compute-cost" ], "runtime": [ "offline-authoring" ], "sources": [ "src-wikipedia-microsoft-flight-simulator-2020", "src-blackshark-2026-homepage", "src-wikipedia-microsoft-flight-simulator-2024" ], "status": "sourced", "summary": "Machine learning classifies buildings, vegetation, and terrain in satellite and aerial imagery and procedural rules rebuild them as 3D scenery, which Microsoft Flight Simulator streams for the whole Earth.", "trl": 8, "type": "technique", "year": 2020 }, { "_derived": { "affects_count": 6, "families_affected": { "agents-behavior": 1, "economy-finance": 1, "environment-ecology": 2, "logistics-networks": 1, "risk-failure": 1 }, "incoming": { "applied_in": [ "calibrated-serious-sims", "human-calibrated-customers", "participatory-planning-sandboxes" ] } }, "affects": [ { "effect": "tunes", "id": "contagion", "note": "Transmission, recovery and contact parameters can be inferred from real outbreak curves or target curves, with a posterior that shows which parameters the data cannot pin down." }, { "effect": "tunes", "id": "supply-demand-pricing", "note": "Economic ABM parameters can be calibrated to price and quantity time series (Dyer et al.), grounding market sims in real data instead of hand-set elasticities." }, { "effect": "tunes", "id": "traffic-congestion", "note": "Demand and route-choice parameters of an agent-based traffic model can be fitted to observed counts, the same likelihood-free ABM calibration applied to a city." }, { "effect": "tunes", "id": "attribute-match-engine", "note": "Match-engine parameters can be fitted so simulated score and statistic distributions match real league data, with uncertainty on each parameter." }, { "effect": "tunes", "id": "food-web", "note": "Predator-prey rates can be inferred from target population time series, so designers state desired oscillations and the posterior returns rates that produce them." }, { "effect": "tunes", "id": "adaptive-threat-scaling", "note": "Treating target survival-time or loss-rate distributions from playtest telemetry as the observation, the posterior gives scaling parameters that reproduce them and how sensitive each one is." } ], "aliases": [ "SBI", "neural posterior estimation", "likelihood-free inference", "ABM calibration" ], "category": "math-discovery", "deployed_in": [], "determinism_impact": "preserves", "horizon": "1-2y", "id": "simulation-based-inference", "key_results": [ { "confidence": "high", "sources": [ "src-cranmer-2020-sbi-frontier" ], "text": "Cranmer, Brehmer and Louppe reviewed simulation-based inference as a general framework for inverse problems in simulators with intractable likelihoods." }, { "confidence": "high", "sources": [ "src-greenberg-2019-apt" ], "text": "Automatic posterior transformation (APT) made sequential neural posterior estimation work with arbitrary, dynamically updated proposals and flow-based estimators, including on high-dimensional time series and images." }, { "confidence": "high", "sources": [ "src-dyer-2024-bayesian-inference-economic-abm" ], "text": "Neural posterior and neural ratio estimation gave state-of-the-art parameter inference for economic agent-based models from multivariate time series, where older likelihood-free methods needed hundreds of thousands of simulations." }, { "confidence": "medium", "sources": [ "src-dyer-2022-abm-microdata-gnn" ], "text": "Temporal graph neural networks learned ABM parameter posteriors directly from microdata, avoiding information loss from hand-chosen summary statistics." }, { "confidence": "medium", "sources": [ "src-boelts-2025-sbi-reloaded" ], "text": "The open-source sbi toolkit packages these estimators with diagnostics and defaults, needing only simulator samples." } ], "limitations": [ "Needs many simulator runs, from thousands to hundreds of thousands depending on method and dimensionality, which is costly for slow, full-scale game simulations.", "A misspecified simulator yields confident posteriors over wrong parameters; calibration fits the data and does not validate the model.", "Posterior quality depends on the summary statistics or learned embeddings of noisy, high-dimensional outputs.", "Matching real data or telemetry targets does not make a mechanic fun; it only removes guesswork about parameter values.", "No game studio use was found in this run's sources." ], "maturity": "research", "mechanism": "Draw parameters theta from a prior, run the simulator to get outputs x, and train a conditional density estimator, usually a normalizing flow, to approximate p(theta | x) from the (theta, x) pairs (neural posterior estimation). Conditioning on observed data x_o gives the posterior; sequential variants such as APT draw new simulations from proposals concentrated near x_o. Related variants learn the likelihood or a likelihood ratio instead. Summary statistics can be hand-chosen or learned, for example with a temporal graph neural network over agent-level microstates. The method needs only forward runs, no likelihood and no gradients, and once trained it is amortized: a new observation gets its posterior in one forward pass.", "name": "Simulation-based inference for calibrating simulators", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "evaluation", "compute-cost" ], "runtime": [ "analytics", "offline-authoring" ], "sources": [ "src-cranmer-2020-sbi-frontier", "src-greenberg-2019-apt", "src-dyer-2024-bayesian-inference-economic-abm", "src-dyer-2022-abm-microdata-gnn", "src-boelts-2025-sbi-reloaded" ], "status": "sourced", "summary": "Neural density estimators trained on simulator runs return a posterior over a simulator's parameters given observed data, calibrating agent-based and economic models whose likelihoods cannot be computed.", "trl": 3, "type": "technique", "year": 2019 }, { "_derived": { "affects_count": 2, "families_affected": { "control-automation": 1, "physics-materials": 1 } }, "affects": [ { "effect": "accelerates", "id": "soft-body-deformation", "note": "Deformation learned from an offline solver runs 300-5000x faster than that solver, putting near-offline-quality cloth and flesh response within a real-time budget." }, { "effect": "deepens", "id": "embodied-skill-control", "note": "Hands-on tasks such as surgery or handling soft objects could show tissue and cloth reacting plausibly to tools in real time, adding feedback the player reads to perform the skill." } ], "aliases": [ "Ubisoft subspace neural physics", "Unreal ML Deformer", "learned cloth deformation", "neural morph model" ], "category": "neural-physics", "deployed_in": [], "determinism_impact": "preserves", "horizon": "now", "id": "subspace-neural-physics", "key_results": [ { "confidence": "medium", "sources": [ "src-holden-2019-subspace-neural-physics", "src-ubisoft-2019-data-driven-physics", "src-holden-2020-gdc-squishy-bunnies" ], "text": "Ubisoft La Forge's subspace neural physics ran interactive cloth and deformable simulation 300 to 5000 times faster than conventional simulation at a small additional memory cost (the project page states 300 to 4000)." }, { "confidence": "high", "sources": [ "src-epic-2026-ml-deformer" ], "text": "Unreal Engine's ML Deformer trains models on externally generated deformation caches from Maya or Houdini and applies them at runtime as GPU morph targets; the Neural Morph Model is the recommended default and the cloth-oriented Nearest Neighbor Model is marked experimental." } ], "limitations": [ "Quality holds only near the training distribution of poses and interactions; unusual poses or collisions produce artifacts, so coverage of the offline dataset sets the limit.", "Training is heavy: Unreal's documentation notes long training times and Alembic training caches of 50-100 GB.", "GPU memory for morph targets is the main runtime cost, and compression settings trade memory and speed for quality.", "Output is usually cosmetic; it does not feed back into gameplay physics unless the game reads the deformed state, which brings determinism and consistency issues.", "No commercial title using these models at runtime was confirmed in this run's sources." ], "maturity": "prototype", "mechanism": "Developers run an expensive offline simulation (cloth, soft tissue, muscle, volume-preserving flesh) over many randomized poses or interactions. Vertex positions are compressed, for example with principal component analysis, into a low-dimensional subspace. A small network is trained to predict the next subspace coordinates from the current coordinates plus external inputs such as bone transforms or collider positions (Holden et al. 2019), or to predict per-pose corrective morph-target weights applied on top of linear skinning (Unreal Engine ML Deformer, Neural Morph Model and Nearest Neighbor Model). At runtime only the network and the decode from subspace to full mesh run, on the GPU.", "name": "Subspace neural physics and learned deformers", "provenance": { "created": "2026-10-24-techniques-sim-math" }, "risks": [ "consistency", "controllability" ], "runtime": [ "client-realtime", "offline-authoring" ], "sources": [ "src-holden-2019-subspace-neural-physics", "src-ubisoft-2019-data-driven-physics", "src-holden-2020-gdc-squishy-bunnies", "src-epic-2026-ml-deformer" ], "status": "sourced", "summary": "Small networks trained on offline cloth, flesh, or muscle simulation predict deformation in a compressed subspace or as corrective morph targets, running hundreds to thousands of times faster than the source simulation inside a game engine.", "trl": 6, "type": "technique", "year": 2019 }, { "_derived": { "affects_count": 3, "families_affected": { "narrative-generation": 1, "space-construction": 2 }, "incoming": { "applied_in": [ "generative-decoration-catalogs", "illustrated-chronicles" ] } }, "affects": [ { "effect": "cuts-authoring-cost", "id": "decoration-beauty", "note": "Decor and furniture catalogs that beauty and environment scores consume can be produced from prompts or reference images in seconds per asset, so catalog size stops being bounded by modeling time." }, { "effect": "deepens", "id": "generative-content", "note": "Runtime generation extends from text and textures to placeable 3D props: inZOI players convert their own images into in-game objects." }, { "effect": "enables-new", "id": "modular-assembly", "note": "Part-controllable generation (Roblox CubePart) produces assets split into parts intended for physics and scripts, which could supply new components for build-from-parts mechanics on request; not yet shown in a shipped assembly game." } ], "aliases": [ "TRELLIS", "Hunyuan3D", "Roblox Cube", "image-to-3D" ], "category": "neural-rendering-3d", "deployed_in": [ "inzoi" ], "determinism_impact": "mitigable", "horizon": "now", "id": "text-to-3d-assets", "key_results": [ { "confidence": "high", "sources": [ "src-xiang-2024-trellis" ], "text": "TRELLIS (Microsoft, CVPR 2025) trains models up to 2B parameters on about 500K 3D objects and decodes one latent to radiance fields, 3D Gaussians, or meshes; it is MIT-licensed and needs a GPU with at least 16 GB." }, { "confidence": "high", "sources": [ "src-microsoft-2025-trellis-2-repo" ], "text": "TRELLIS.2 (4B parameters) outputs PBR-textured GLB assets in about 3 seconds at 512^3, 17 seconds at 1024^3, and 60 seconds at 1536^3 on an H100." }, { "confidence": "high", "sources": [ "src-tencent-2025-hunyuan3d-2", "src-tencent-2025-hunyuan3d-2-1-repo" ], "text": "Hunyuan3D 2.0 (January 2025) released open shape and texture models; Hunyuan3D 2.1 (June 2025) added PBR texturing and training code and needs about 29 GB of VRAM for shape plus texture." }, { "confidence": "high", "sources": [ "src-roblox-2025-cube", "src-roblox-2026-cube-repo" ], "text": "Roblox released Cube 3D text-to-shape (v0.1 March 2025, v0.5 July 2025) and CubePart part-controllable generation (May 2026) aimed at assets usable with animation, physics, and scripts." }, { "confidence": "high", "sources": [ "src-steam-2025-inzoi" ], "text": "inZOI's store page states that players can turn 2D images into in-game 3D objects, with AI tools described as auxiliary and generated content reviewed by the team." }, { "confidence": "medium", "sources": [ "src-wikipedia-inzoi" ], "text": "inZOI received negative user reviews in May 2025 after Krafton announced its use of generative AI." } ], "limitations": [ "Hardware: 16 GB (TRELLIS) to about 29 GB (Hunyuan3D 2.1 shape plus texture) of GPU memory, so player-side generation usually runs in the cloud or on high-end PCs.", "Output topology and UVs are not authored for animation or LOD; production use still needs cleanup and review, which inZOI states it applies.", "Generated assets are single objects without gameplay semantics; the game must assign function, footprint, and value.", "Training data provenance and licenses vary by model, and player reception of generative AI can be negative (inZOI reviews in 2025)." ], "maturity": "shipped-niche", "mechanism": "Generation is split into shape and appearance. A flow-matching or diffusion transformer generates a 3D latent conditioned on an image or text: TRELLIS uses a Structured LATent (sparse voxels carrying features from a vision foundation model), Hunyuan3D-DiT a shape latent, and Roblox Cube discrete shape tokens predicted autoregressively. Decoders turn the latent into meshes, 3D Gaussians, or radiance fields. Texture is painted by a multi-view diffusion model (Hunyuan3D-Paint) or decoded directly, with 2025 versions emitting PBR maps (base color, roughness, metallic). Output is a GLB or OBJ file that drops into an engine like any other asset.", "name": "Text- and image-to-3D asset generation", "provenance": { "created": "2026-10-24-techniques-world-models" }, "risks": [ "ip-provenance", "player-trust", "labor", "content-safety", "evaluation" ], "runtime": [ "offline-authoring", "design-assist" ], "sources": [ "src-xiang-2024-trellis", "src-microsoft-2025-trellis-2-repo", "src-tencent-2025-hunyuan3d-2", "src-tencent-2025-hunyuan3d-2-1-repo", "src-roblox-2025-cube", "src-roblox-2026-cube-repo", "src-steam-2025-inzoi", "src-wikipedia-inzoi" ], "status": "sourced", "summary": "Open 3D generative models (TRELLIS, Hunyuan3D 2.x, Roblox Cube) produce textured meshes from a text prompt or image in seconds to a minute, and inZOI exposes image-to-3D object creation to players.", "trl": 7, "type": "technique", "year": 2025 }, { "_derived": { "affects_count": 5, "families_affected": { "agents-behavior": 1, "control-automation": 3, "progression-goals": 1 }, "incoming": { "applied_in": [ "language-modding", "sims-as-agent-benchmarks", "teachable-sim-companion" ] } }, "affects": [ { "effect": "enables-new", "id": "player-programming", "note": "The player can describe a behavior in language and the model writes, tests against the live game, and saves the program, so player code becomes a growing library of named skills rather than hand-typed scripts." }, { "effect": "enables-new", "id": "agent-skills", "note": "An NPC's competence can be an accumulating library of composable, verified routines it wrote itself, so skill growth adds new behaviors rather than raising a numeric speed or quality level." }, { "effect": "enables-new", "id": "delegation-to-ai", "note": "A delegated controller can acquire routines for situations its designers did not code by writing and verifying new functions, instead of running the fixed opponent-AI logic restricted to a subsystem." }, { "effect": "tunes", "id": "tech-tree", "note": "Exploration agents that unlock milestones 15.3x faster than earlier agents can be run as automated playtesters to measure how quickly and in what order a progression graph gets traversed." }, { "effect": "deepens", "id": "programming-by-demonstration", "note": "Demonstrated routines can be stored as verified, reusable code skills that generalize to new locations." } ], "aliases": [ "Voyager", "code-as-skill", "lifelong learning agent" ], "builds_on": [], "category": "llm-agents", "deployed_in": [], "determinism_impact": "mitigable", "horizon": "1-2y", "id": "voyager-skill-libraries", "key_results": [ { "confidence": "high", "sources": [ "src-wang-2023-voyager" ], "text": "Against prior LLM agents, Voyager obtained 3.3x more unique items, traveled 2.3x longer distances, and unlocked tech-tree milestones up to 15.3x faster, discovering 63 unique items in 160 prompting iterations." }, { "confidence": "high", "sources": [ "src-wang-2023-voyager" ], "text": "The learned skill library let a fresh agent solve novel tasks in a new Minecraft world from scratch, where baselines struggled." }, { "confidence": "high", "sources": [ "src-wang-2023-voyager" ], "text": "Code generation quality was the bottleneck: replacing GPT-4 with GPT-3.5 for code produced 5.7x fewer unique items, while GPT-4 cost 15x more per token." }, { "confidence": "medium", "sources": [ "src-hopkins-2025-factorio-learning-environment" ], "text": "In the 2025 Factorio Learning Environment, LLM agents writing code found growth strategies such as electric drilling but failed at complex automation such as electronic-circuit production, with weak spatial reasoning and error recovery." } ], "limitations": [ "Depends on a frontier code model; GPT-3.5 produced 5.7x fewer items and GPT-4 cost 15x more per token, so per-agent cost is high.", "Hallucinated tasks and calls: the curriculum proposed items that do not exist (copper sword), and generated code used invalid inputs or functions absent from the API, causing execution errors.", "Self-verification misjudges success (for example not recognizing a dropped string as proof a spider was defeated).", "Needs a scripting API that exposes game actions and state as text; it does not act from pixels.", "Long-horizon factory building in Factorio still fails at multi-step automation and spatial layout." ], "maturity": "research", "mechanism": "Voyager drives a Minecraft bot through a JavaScript control API. An automatic curriculum asks GPT-4 for the next task given inventory, surroundings, and completed tasks, aiming to discover as many diverse things as possible. For each task GPT-4 writes a function that may call existing skills; the function runs, and the agent feeds back environment observations, execution errors, and a GPT-4 self-verification verdict, retrying up to four rounds before moving on. Verified functions are stored in a skill library indexed by an embedding of their description and retrieved (top-k) when writing later functions, so complex behaviors are composed from earlier ones. No model weights are trained; all learning is in the growing code library, which can be reused in a new world.", "name": "LLM-written skill libraries (Voyager)", "provenance": { "created": "2026-10-24-techniques-llm-agents", "reviewed": [ "2026-10-24-technique-links" ] }, "risks": [ "compute-cost", "consistency", "evaluation", "controllability", "latency" ], "runtime": [ "server-realtime", "analytics" ], "sources": [ "src-wang-2023-voyager", "src-hopkins-2025-factorio-learning-environment" ], "status": "sourced", "summary": "An LLM agent that proposes its own next task, writes executable code to do it, refines the code from execution errors and self-checks, and stores verified programs in a retrievable skill library.", "trl": 4, "type": "technique", "year": 2023 }, { "_derived": { "affects_count": 3, "families_affected": { "space-construction": 2, "world-generation": 1 }, "incoming": { "applied_in": [ "learned-build-assist" ] } }, "affects": [ { "effect": "deepens", "id": "procedural-build-assist", "note": "Rough player input becomes consistent detailed structure through constraint propagation over hundreds of modules: in Townscaper a click toggles one cell and the solver re-chooses compatible modules around it, so each placement reshapes its neighbors." }, { "effect": "deepens", "id": "procedural-terrain", "note": "Tile maps can be generated from a small example or adjacency table instead of noise thresholds, giving locally coherent coastlines, cliffs, and interiors, as in Bad North's islands and Caves of Qud's areas; large-scale structure still needs another layer." }, { "effect": "cuts-authoring-cost", "id": "zoning", "note": "Buildings on painted zones could be assembled from modular tile sets constrained by neighbors and lot shape rather than authored per lot, varying silhouettes without new art for every building." } ], "aliases": [ "WFC", "WaveFunctionCollapse", "model synthesis" ], "category": "procedural-generation", "deployed_in": [ "townscaper", "caves-of-qud" ], "determinism_impact": "preserves", "horizon": "now", "id": "wave-function-collapse", "key_results": [ { "confidence": "high", "sources": [ "src-gumin-2016-wavefunctioncollapse", "src-karth-2017-wfc-constraint-solving" ], "text": "Maxim Gumin released WaveFunctionCollapse publicly on 2016-09-30 with overlapping and simple tiled models, and developers began experimenting with it within a day." }, { "confidence": "high", "sources": [ "src-gumin-2016-wavefunctioncollapse" ], "text": "The WFC repository lists Bad North, Caves of Qud, Dead Static Drive, Townscaper, and Matrix Awakens as users, with adaptations for Unity, Unreal Engine 5, Godot 4, and Houdini." }, { "confidence": "high", "sources": [ "src-karth-2017-wfc-constraint-solving" ], "text": "Karth and Smith (FDG 2017) showed WFC is constraint solving with MRV ordering, frequency-weighted value choice, arc-consistency propagation, and global restarts; with an added global constraint restarts failed to finish within a minute while a backtracking solver resolved it quickly." }, { "confidence": "high", "sources": [ "src-thompson-2022-how-townscaper-works" ], "text": "Townscaper runs WFC in real time on an irregular quad grid to choose building tiles from the surrounding shape and may fail silently where cells cannot fit, whereas Bad North restarts generation; Bad North uses over 300 tiles." }, { "confidence": "high", "sources": [ "src-karth-2017-wfc-constraint-solving", "src-steam-2024-caves-of-qud" ], "text": "Caves of Qud, an early user of WFC alongside multipass generation, reached 1.0 on 2024-12-05 after early access began in 2015." }, { "confidence": "medium", "sources": [ "src-boristhebrave-2020-wfc-explained" ], "text": "WFC constrains only nearby tiles, so it rarely produces large-scale structure without added constraints or other techniques." } ], "limitations": [ "Only local adjacency is enforced; reachability, room counts, or path lengths need extra constraints, and Karth and Smith found a global constraint caused hundreds of conflicts that restarts could not clear within a minute.", "Contradictions are possible because the underlying problem is NP-hard; shipped systems restart (Bad North) or accept silent local failures (Townscaper).", "Interactive generation delay grows with structure size, and the overlapping model is slower and can copy the example too closely.", "Quality depends on an authored module set, often hundreds of tiles (Bad North uses over 300), so art cost moves into modules rather than disappearing." ], "maturity": "shipped-niche", "mechanism": "Each output cell starts with a domain of all tiles (simple tiled model) or of all N x N patterns extracted from one example image (overlapping model, typically N = 3). The solver repeatedly picks the undecided cell with the lowest Shannon entropy over its remaining options, a minimum-remaining-values heuristic, collapses it to one option sampled by that option's frequency in the example, and propagates: neighbors drop options that no longer have a compatible partner, which is arc consistency as in AC-3. If a domain empties, the attempt fails and the original algorithm restarts; later implementations add backtracking or local repair. Designers steer output with tile sets, adjacency rules, weights, pre-set cells, and extra global constraints. Townscaper runs it in real time on an irregular quad grid: the player toggles a cell's occupancy and the solver chooses compatible building modules from the surrounding shape, with decoration applied in a separate pass.", "name": "Wave Function Collapse and constraint-based tile generation", "provenance": { "created": "2026-10-24-techniques-rl-design" }, "risks": [ "controllability", "latency" ], "runtime": [ "client-realtime", "offline-authoring", "design-assist" ], "sources": [ "src-gumin-2016-wavefunctioncollapse", "src-karth-2017-wfc-constraint-solving", "src-thompson-2022-how-townscaper-works", "src-steam-2024-caves-of-qud", "src-boristhebrave-2020-wfc-explained" ], "status": "sourced", "summary": "An example-driven constraint solver that fills a grid by repeatedly collapsing the lowest-entropy cell to one tile and propagating adjacency constraints, shipped in games including Townscaper, Bad North, and Caves of Qud.", "trl": 8, "type": "technique", "year": 2016 } ]