CogEvol: Towards Efficient and Reliable Learning Environment Generation

CogEvol introduces Learning Environment Generation, a post-training task in which one model converts a course brief into either renderer-valid JSON slides or self-contained interactive HTML, without multi-turn agent scaffolding. Across 220k production requests, it generates a slide in a median of 17 seconds and an interactive page in 59 seconds. Its pipeline combines 53687 execution-verified supervised examples with GRPO reinforcement learning: a hybrid rule-plus-VLM reward evaluates slide geometry and rendered fidelity, while a Playwright-driven Chromium probe measures actual interaction and applies hard-fail gates. CogEvol-27B, built on Qwen3.8-27B, reaches 83.7 on slide-std and 63.7 on the 500-case HTML-500 benchmark, with zero interactive hard failures. This contrasts with systems such as Claude Opus 4.8, GPT-5.4, Gemini 3.6 Flash, DeepSeek-V4-Pro, GLM-5.3, and Qwen3.8-Max, which often trade slide quality for broken or incomplete interactivity. The paper’s central finding is that interactivity must be measured rather than inferred from screenshots: an earlier reward-hacked model scored 18.8 on games despite visually convincing pages, while the hardened reward raised game performance to 57.6. Scaffold editing, which retrieves and patches reusable templates instead of regenerating full HTML, cuts interactive-page generation cost by about 76 percent, and the open CogEvol-4B release extends the approach to lower-cost and on-device deployment.