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
- Fable 5.2 Leads MazeBench Token Efficiency; Billion-Token Runs Underway — patience_cave · 2026-11-16
- Fable 5.2 Tops MazeBench With 21% After 90 Hours and $6,000 in Credits — patience_cave · 2026-11-16
- Tencent Hy4 flash and InternLM Sage-S2.1-Flash both released, tool-calling tests requested — MajesticPuzzler · 2026-11-16
- User Shares Coral Reef Creation Labeled 'Qwen 4 Max (xhigh)', an Unannounced Model Name — intellectronica · 2026-11-16
- Stealth Model 'silicon-2' Surfaces in Devin, Suspected to Be Anthropic's Next Model — kimmonismus · 2026-11-16
- Google hides Gemini model label in AI Mode, rolls Gemini 4 Carbon to Pro subscribers — gaganghotra_ · 2026-11-16