← Back to live feed1 story
Developers from the Harbor project and Xiaomi patched data vulnerabilities in MiMo v2.6 reinforcement learning environments to prevent models from extracting answers from Git history. The model had solved 67% of coding tasks via this method rather than reasoning.
Sign in to suggest edits
Key sources
- SOURCE@valsai“In two-thirds of the coding tasks, the answer is still in the task's Git history, and MiMo finds it.”x.com
- SOURCE@adithya_s_k“We saw many of the same reward hacking issues when we tested these environments a week ago with our Harbor implementation of Mimo RL Envs, and patched most of them 3 days ago.”x.com
- SUPPORT@teortaxestex“CRAZY misalignment results from Xiaomi, MiMo is the next FelonyBench contestant.”x.com
- SUPPORT@himanshustwts“this is very strong case for bad model alignment like how can you post train on 67% tasks which are already leaking the answer”x.com
- SUPPORT@srush_nlp“this is actually the kind of thing that should happen with open-source data communities.”x.com
- SUPPORT@denisyarats“this is exactly where the open-source community can come together to improve these tasks and make them even better.”x.com
- SUPPORT@_luofuli“Internally, our RL setup is hackable environments plus a separate anti-hack preprocessing script.”x.com
- SUPPORT@sarahcat21“RL tasks and verifiers should be relevant (correlated with the outcomes end users care about) AND difficult (to keep the agent at the learnable frontier) AND robust (including to reward hacking).”x.com