Gamer: A Generalist VLA Foundation Model Trained on Tens of Thousands of Hours of Keyboard-Mouse Gameplay
pmttyji · reddit · 2026-11-16
THUSI-Lab and collaborators release Gamer, a generalist vision-language-action foundation model for long-horizon, real-time game agents.
- Training: continued pre-training on tens of thousands of hours of device-level keyboard/mouse trajectories spanning 2D interfaces and 3D simulated worlds; sparse-thinking instruction tuning cuts redundant computation; a stable multi-turn RL method handles finite context, accumulated rollout errors, and sparse rewards.
- Serving: a cloud-edge engine concurrently serves many simulated game environments with unbounded context and low latency.
- Results: generalization and efficiency comparable to skilled players across 3D open-world games, plus competitive results on 2D benchmarks.
Paper, code, weights, and project page are all public.
Related event: Gamer: A VLA Foundation Framework for Long-Horizon Game Agents(2 posts)→
More from Research
- AI paper authors swarmed by predatory journal spam that smells like a pay-to-publish scam — letandrewcook · 2026-11-16
- Can we fix AI? Peter Bloem's new essay examines data, copyright and synthetic data paths — pbloemesquire · 2026-11-16
- Noumenal Labs open-sources hybrid-aif, a JAX active inference library for robotics — mjdramstead · 2026-11-16
- ICLR 2027 author gets "you didn't address my concern" one day before the rebuttal deadline — 3scorciav · 2026-11-16
- Realtime LLM vector steering demo shared on X — SaleemUsama · 2026-11-16
- HDM paper: "zero-or-T" routing cuts agentic token cost by 80% with no accuracy loss — pmttyji · 2026-11-16