What did DeepMind report when SIMA 2 agents acted in Genie 3-generated worlds: inputs, outputs, constraints, measured results?
The SIMA 2 paper (DeepMind, Dec 2025) runs a Gemini Flash-Lite-based agent on 720p frames with keyboard-and-mouse actions in Genie 3 worlds. Genie 3 environments are text- or frame-conditioned, split into urban (train) and natural (held-out) navigation tasks. Qualitatively the agent navigates to points of interest; for self-improvement the authors report gains on most train tasks, often 25 points or more, and on most held-out natural tasks (Fig. 17; values read from figure not extracted). The Genie 3 blog (Aug 2025) is an earlier company statement of a SIMA run with no figures. Task count, number of episodes and variance are not given in the text I extracted.
The SIMA 2 paper (DeepMind, Dec 2025) runs a Gemini Flash-Lite-based agent on 720p frames with keyboard-and-mouse actions in Genie 3 worlds. Genie 3 environments are text- or frame-conditioned, split into urban (train) and natural (held-out) navigation tasks. Qualitatively the agent navigates to points of interest; for self-improvement the authors report gains on most train tasks, often 25 points or more, and on most held-out natural tasks (Fig. 17; values read from figure not extracted). The Genie 3 blog (Aug 2025) is an earlier company statement of a SIMA run with no figures. Task count, number of episodes and variance are not given in the text I extracted.
Curator qualifications: Stripped text: '720p' 1, 'Flash-Lite' 5, 'keyboard' 15, 'Genie 3' 22, 'License: CC BY 4.0' 1. Sec 4.5.2: 'we split our set of Genie 3 environments into urban (train) and natural (test) environments and tasks, primarily centered on navigation'; 'SIMA 2 improves across nearly all train tasks, often by 25 points or more'; 'in the majority of the tasks in natural environments (held out), the self-improved SIMA 2 outperforms the initial agent'; 'This provides initial evidence'.
Evidence status: statements located in the cited sources by a correlated desk review (Claude); not executed; not independently reviewed.
Structured fields
{
"citations": [
{
"url": "https://arxiv.org/html/2512.04797v1",
"note": "arXiv HTML 2512.04797v1, sec 4.5.2 / Fig. 17 text"
},
{
"url": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models",
"note": "cited at https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/; arXiv HTML 2512.04797v1, sec 4.5.2 / Fig. 17 text"
},
{
"url": "https://deepmind.google/blog/sima-2-an-agent-that-plays-reasons-and-learns-with-you-in-virtual-3d-worlds/",
"note": "arXiv HTML 2512.04797v1, sec 4.5.2 / Fig. 17 text"
},
{
"url": "https://deepmind.google/discover/blog/genie-3-a-new-frontier-for-world-models/",
"note": "observed URL (dated/alias form of the cited catalogue source; canonical identity retained)"
}
]
}Evidence claims
- [sourced_fact] The SIMA 2 paper (DeepMind, Dec 2025) runs a Gemini Flash-Lite-based agent on 720p frames with keyboard-and-mouse actions in Genie 3 worlds. Genie 3 environments are text- or frame-conditioned, split into urban (train) and natural (held-out) navigation tasks. Qualitatively the agent navigates to points of interest; for self-improvement the authors report gains on most train tasks, often 25 points or more, and on most held-out natural tasks (Fig. 17; values read from figure not extracted). The Genie 3 blog (Aug 2025) is an earlier company statement of a SIMA run with no figures. Task count, number of episodes and variance are not given in the text I extracted. · locator: arXiv HTML 2512.04797v1, sec 4.5.2 / Fig. 17 text source
Rights
| component | license | hosting | redistribution | modification | training | Evidence and notes |
|---|---|---|---|---|---|---|
| documentation | unknown | permitted | permitted | unknown | unknown | Applies to this original summary only, not to the cited sources. |