src-deepmind-2025-sima-2-paper source
SIMA 2: A Generalist Embodied Agent for Virtual Worlds
locator
notes
Roughly doubles SIMA 1 average success, approaching humans; over 10% better than SIMA 1 on held-out ASKA and MineDojo; 26 of 50 MineDojo categories. Task setter plus reward model scoring 0-100 (>=50 success). Gemini Flash-Lite base for latency; SFT cut AIME about 25.5%; weak combat; short memory.