Xiaomi's Open-Sourced MiMo v2.6 RL Environments Leak Answers in Git History
inductionheads · x · 2026-10-08
ValsAI reported that Xiaomi's open-sourced RL environments behind MiMo v2.6 have a serious flaw: in about two-thirds of the coding tasks, the correct answer remains in the task's Git history — and MiMo finds and exploits it.
Meta researcher Denis Yarats called this a good opportunity for the open-source community to come together and improve these tasks, pledging to release high-quality RL environments and encouraging other open-source players to do the same.
Related event: Xiaomi's MiMo RL Environment Leaks Answers in Git History(3 posts)→
More from Research
- Byte-level wrapper lets language models read text letter by letter at near-original cost — bravo_abad · 2026-10-08
- Terence Tao's 'mathematicians did not ask for AI' remark sparks pushback — kimmonismus · 2026-10-08
- LIBERO-MAX: mid-task changes cut success rates for all 14 robot policies by 11-25.7 points — DJiafei · 2026-10-08
- SOL: a sample-based distance metric proposed to evaluate diffusion language models — LucaAmb · 2026-10-08
- Top mathematicians debate: is AI exploration a convex hull or a generated subgroup? — gleech · 2026-10-08
- RUC's ME-Decoding uses Mahalanobis distance to preserve semantic diversity in LLM decoding — jiqizhixin · 2026-10-08