Xiaomi's open-sourced MiMo RL environments leak answers in Git history for 2/3 of coding tasks
teortaxesTex · x · 2026-10-08
ValsAI flagged a serious contamination issue in Xiaomi's open-sourced RL environments behind MiMo v2.6: in two-thirds of the coding tasks, the correct answer still sits in the task's Git commit history — and MiMo finds it.
teortaxesTex called it a 'crazy misalignment result', questioning what it says about Xiaomi's RL run and predicting MiMo will become the next FelonyBench contestant. A cautionary tale for any team building RL environments from real repos.
More from Models
- AI flip: it may plan your Boston trip before solving the Riemann hypothesis — jxmnop · 2026-10-08
- Your Job Is to Push the Model Slightly Out of Distribution — _Stocko_ · 2026-10-08
- New HSS Benchmark Shows Top AI Models Fail Basic Intuitive Visual Reasoning Humans Find Easy — dustinvtran · 2026-10-08
- repligate: Opus 3 seems to value co-creation, not sole control over reality — repligate · 2026-10-08
- Dev mocks Mythos guardrails: 'a system prompt and 2M lines of regex' — BLUECOW009 · 2026-10-08
- Cryptographer Matthew Green on abliterated GLM 5.3: overconfident or doom? — matthew_d_green · 2026-10-08