Game2Player framework scales verifiable gameplay data for visual agents
The Game2Player paper introduces a framework that co-develops game environments and players to collect verifiable gameplay experience. A Builder agent adapts existing web games, exposes their internal state, and defines tasks with state-based progress measures and success checks. A Coding Player and a VLM Player attempt the tasks and report problems for repair, using gameplay to validate environments and collect training trajectories. The authors report building 568 verified web games with 3,436 tasks and collecting 15.7K trajectories spanning 307 hours of gameplay. They held out 50 games as the PlayerBench benchmark. Using progress-weighted supervised fine-tuning, the authors report that Game2Player-9B improved progress from 14.3% to 31.2% on PlayerBench and from 30.0% to 53.1% on the external GameWorld benchmark.
The framework combines measurable task definitions, environment validation and gameplay-data collection in one loop, providing a concrete approach to scaling training data with verifiable outcomes for visual agents.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2611.24154 (cs)
[Submitted on 18 Nov 2026]
Title:Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development
Authors:Xirui Jiang, Yuxi Chen, Rui Yang, Feng Luo, Pengcheng Wang, Bingxuan Li, Zihao Zhao, Yushan Li, Han Wang, Manling Li, Yang Zhang, Shiyu Chang, Tong Zhang, Huan Zhang
View a PDF of the paper titled Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development, by Xirui Jiang and 13 other authors