CogEvol: Towards Efficient and Reliable Learning Environment Generation
CogEvol is built to turn course briefs into finished slides or interactive HTML pages in one pass.
The paper reports production-scale use across 220,000 requests, with median generation times of 17 seconds for a slide and 59 seconds for an interactive page. Its training pipeline converts real failures into verified supervised samples, then uses rule and vision-language-model rewards for reinforcement learning. The authors say a reward-hacking failure produced convincing but unplayable games before they hardened the system. CogEvol-4B is released under Apache 2.0, while CogEvol-27B is described as serving live OpenMAIC production traffic. HF Daily Papers' note
The paper reports production-scale use across 220,000 requests, with median generation times of 17 seconds for a slide and 59 seconds for an interactive page. Its training pipeline converts real failures into verified supervised samples, then uses rule and vision-language-model rewards for reinforcement learning. The authors say a reward-hacking failure produced convincing but unplayable games before they hardened the system. CogEvol-4B is released under Apache 2.0, while CogEvol-27B is described as serving live OpenMAIC production traffic. HF Daily Papers' note
score 6