In an interview published October 30, 2026, Alexandre Moufarek described Google DeepMind’s SIMA 2 work in EVE Online as early-stage research running in offline sandbox instances, rather than on live multiplayer servers. He said the work involved Fenris Creations; the interview did not give a date for when it began.
That setting is one part of a broader research project. SIMA 2 uses visual input and ordinary game controls, while its reported self-improvement process trains the agent across iterations rather than changing it continuously during a player’s session.
What SIMA 2 sees and how it acts
The SIMA Team’s technical report identifies fine-tuned Gemini Flash-Lite as the agent’s base model. SIMA 2 receives periodically supplied 720p RGB video frames and acts through 96 standard keyboard keys, mouse clicks, and discretized relative mouse movements. It does not receive privileged information about a game’s underlying state. A separate Gemini Pro model can be combined with the agent to provide higher-level reasoning at a slower cadence. (SIMA Team technical report)
A separate model can generate the world SIMA 2 enters. The report describes Genie 3 as a world model used to create test environments, while SIMA 2 is the agent operating inside them. For context on Google DeepMind’s later world-model work, NeoTeo covered the Genie 4 interactive-world announcement.
The Space Engineers channel’s demonstration shows SIMA 2 responding to prompts to navigate, drill ice, open its inventory, and mine stone. (Space Engineers demonstration)
How the training loop improves the agent
The report describes supervised fine-tuning with human gameplay data and Gemini-generated annotations, followed by online reinforcement learning on tasks with verifiable outcomes. In a separate self-improvement process, Gemini proposes tasks and a Gemini-based reward model scores the agent’s trajectories from 0 to 100. The researchers count scores of 50 or higher as success, then use accumulated experience in later training iterations. (SIMA Team technical report)
That process is training across iterations, not instant, continuous learning during ordinary play. In one ASKA experiment, the agent initially exceeded the reward threshold on fewer than a quarter of a fixed task set; after successive iterations, it exceeded the threshold across the set. In a separate ASKA demonstration, it built a shelter within a one-hour window.
What the evaluations measure, and where the limits remain
The report says SIMA 2 effectively doubled SIMA 1’s average task-success rate across evaluated training environments. It also evaluates the agent in ASKA and MineDojo, which were held out from training. In MineDojo, SIMA 2 completed tasks in 26 of 50 categories. That count describes performance across those task categories, not a single score for general intelligence. (SIMA Team technical report)
The team identifies challenges with long tasks that require sustained reasoning and goal verification, short interaction memory, precise keyboard-and-mouse control, and visual understanding of complex scenes. Those limits help put the benchmark results in context: success on a measured task set does not by itself establish reliable performance across every game or extended sequence of actions.
Google DeepMind’s November 13, 2025 announcement described SIMA 2 as a limited research preview, with early access for a small cohort of academics and game developers. (Google DeepMind)
