← 研踪 PaperTrailarXiv:2611.14015 · v1
Margin-Guided Token Credit for Reinforcement Learning with Verifiable Rewards
首次提交 2026-11-10 20:51
版本更新 2026-11-10 20:51
雷达观测 2026-11-11 11:04
摘要
In reinforcement learning with verifiable rewards, a language model is rewarded for a correct final answer, and GRPO passes that reward to every token of the solution in equal measure. We instead place the credit where the model hesitated, in a method we call MCPO (Margin-Credit Policy Optimization). We measure hesitation by the top-2 margin $M=p_1-p_2$ of the policy's next-token distribution, map it to a score with two fixed thresholds, and spread each solution's advantage over its tokens so that the mean stays unchanged. The margin is read from a forward pass the trainer already performs, so the method needs no extra model evaluation and no batch statistics. We train MiniCPM5-2B-Base with LoRA for 200 steps on 2,048 DAPO-Math problems and evaluate on six competition-math benchmarks. MCPO beats an unmodified GRPO baseline for all five data seeds, by 6.55 points of macro avg@32 on average (95\% CI $[4.1, 9.0]$), and a Qwen3-4B-Base replication improves for all three seeds ($+4.75$). The gain is largest on problems that GRPO solves only some of the time and grows with $k$ in pass@$k$, reaching 13.8 points at pass@32. MCPO models also write two to three times longer answers, and runs that write longer score higher; our experiments do not separate the two effects.
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-10 20:51 · 最新观测版本
Margin-Guided Token Credit for Reinforcement Learning with Verifiable Rewards
In reinforcement learning with verifiable rewards, a language model is rewarded for a correct final answer, and GRPO passes that reward to every token of the solution in equal measure. We instead place the credit where the model hesitated, in a method we call MCPO (Margin-Credit Policy Optimization). We measure hesitation by the top-2 margin $M=p_1-p_2$ of the policy's next-token distribution, map it to a score with two fixed thresholds, and spread each solution's advantage over its tokens so that the mean stays unchanged. The margin is read from a forward pass the trainer already performs, so the method needs no extra model evaluation and no batch statistics. We train MiniCPM5-2B-Base with LoRA for 200 steps on 2,048 DAPO-Math problems and evaluate on six competition-math benchmarks. MCPO beats an unmodified GRPO baseline for all five data seeds, by 6.55 points of macro avg@32 on average (95\% CI $[4.1, 9.0]$), and a Qwen3-4B-Base replication improves for all three seeds ($+4.75$). The gain is largest on problems that GRPO solves only some of the time and grows with $k$ in pass@$k$, reaching 13.8 points at pass@32. MCPO models also write two to three times longer answers, and runs that write longer score higher; our experiments do not separate the two effects.
查看此版本解读 →查看此版本原文 ↗