SIMA 2
SIMA 2 puts Gemini inside a 3D game agent that follows instructions, reasons about goals, converses and improves itself; also plays Genie 3-generated worlds.
Upgrade from SIMA 1's instruction following to a reasoning agent that generalizes to unseen games and to worlds generated on the fly by Genie 3, closing much of the gap to human task success on an expanded eval set.
- Date
- Thursday, 13 November 2025
- Lab
- Google DeepMind
- Kind
- model
- Access
- research preview
The SIMA 1 baseline was re-scored on a harder, wider eval set; the team lists short memory and imprecise low-level control as open problems.
Sources
- deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-
- blog.google/technology/ai/google-ai-updates-november-2025/
This record was checked against its sources on 6 October 2026. How we check