Automated generation of game rules and mechanics

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.

• 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.