MiMo-V2 paper: agentic groupwise grading redistributes RL advantages to curb reward hacks
Sauers_ · x · 2026-10-17
The MiMo-V2 paper details a groupwise agentic grading scheme for code-agent RL. (a) Groupwise Reward Synthesis builds task-specific rubrics (solution + behavior) offline from rollouts, grades trajectories online, and combines reward as test outcome × solution score × behavior score. (b) Groupwise Advantage Redistribution compares passing trajectories online, ranks their quality with a groupwise grader, resets confirmed reward-hack rewards to zero, and redistributes sequence-level positive advantage from low- to high-quality passing solutions — letting the agent choose which trajectories earn positive RL advantage.
More from coding & agent
- DeepSeek post-training lead Wu Yu declares the era of democratized wish-driven coding — teortaxesTex · 2026-10-17
- Vercel Labs ports tsc back to TypeScript — and it runs 2-4x faster than tsgo — zeeg · 2026-10-17
- Programmer meme: LGTM now only stands for GTM in the AI coding era — Paimaamu · 2026-10-17
- Dev pushes back on "vibes coding" critique: 100% AI doesn't mean shipping vibes — cjimti · 2026-10-17
- Castwerk: a local ComfyUI studio chaining consistent persona videos via frame anchoring — dizzle_69 · 2026-10-17
- Solo dev's day-35 MMO log: agent-first asset workflow for 2000-player battles — max_paperclips · 2026-10-17