Game2Player: builders and players co-develop 568 games to train a 9B visual agent
ManlingLi_ · x · 2026-11-02
Researchers (UIUC-led) present Game2Player, a pipeline where AI builds its own training grounds and learns from them — no human gameplay data needed.
- Builder agents turn web games into tasks with ground truth read from game state, so every step is auto-verifiable
- Player agents include coding policies (play by writing code) and VLM policies (play from pixels); broken games get reported back to the builder for repair, so environments and players improve together
- Scale: 568 games, 3,436 verifiable tasks, 15.7K trajectories, 1.56M gameplay steps, 307 hours
- PlayerBench: a held-out benchmark of 50 games / 287 tasks where even GPT-6.1 Sol reaches only 62.6% progress
- Result: a 9B VLM trained on this data plays from screenshots alone, lifting progress on GameWorld from 30.0 to 53.1 (roughly doubling, ahead of Qwen3.8-Flash baselines)
The key idea is verifiable experience at scale: programmable environments auto-generate tasks, auto-grade, and auto-repair, replacing costly human demonstrations. Paper, code, dataset, and models are open.
More from coding & agent
- Claude codes an animated pixel-art world in pure TypeScript — full prompt shared — AIandDesign · 2026-11-02
- Opus 5.5 builds a live painting studio from NVIDIA's free SimReady assets — petewoodbridge · 2026-11-02
- AI agents write 800% more code for 30% better delivery: a 2027 AI engineer roadmap — MaryamMiradi · 2026-11-02
- Anthropic's David Hoang: agentic coding pulled me back into iOS development — davidhoang · 2026-11-02
- User gets safety warning after letting Antigravity scan his own servers for exploits — antiriad76 · 2026-11-02
- psnek: A modern reimagining of the Python progress bar with metrics and semantic coloring — minimaxir · 2026-11-02