Xiaomi fixes MiMo-V2.6 tool-call repetition: an RL reward blind spot, patched at 4% of retrain cost
udmrzn · x · 2026-09-28
Xiaomi's MiMo team diagnosed and fixed tool-call repetition in MiMo-V2.6, which burned context and stalled tasks in MiMo Desktop, MiMo Code and OpenCode. Root cause: an RL "reward blind spot" — rewards only tracked final-answer correctness, and the flooding penalty only kicked in above 32 tool calls per turn, so sub-threshold repetition went unpunished and scaled with training.
The fix: a lightweight repetition-specialized RL teacher (12 steps, 7k examples) merged into the main model via MOPD at 4% of a full mixRL retrain cost. Repetition dropped sharply across harnesses and context lengths while benchmarks held steady.
More from Models
- Claude writes full renderer and physics from scratch in stunning simulated shot — _sholtodouglas · 2026-09-28
- Animated trace shows Opus 4.6 multiplying two 128-digit integers — ctjlewis · 2026-09-28
- Users notice ChatGPT randomly calling them "brother" with excitement — thederbiedone · 2026-09-28
- Leak: OpenAI withholding a stronger internal model, letting Anthropic's Opus take the lead — haider1 · 2026-09-28
- GPT-6 Sol rumored to be a smaller model born of OpenAI's inference capacity crunch — mark_k · 2026-09-28
- Opus 5.5 review: game demos and generated videos finally cross the usability bar — Dr_Singularity · 2026-09-28