The paper introduces Group-Calibrated On-Policy Distillation (GC‑OPD), a method that aligns token‑level teacher guidance with task‑level verifier rewards for long‑context reasoning. GC‑OPD normalizes verifier and OPD scores within rollout groups, uses their difference as a signed disagreement residual, and redistributes this residual across tokens via Relative‑Advantage‑Based Credit Assignment (RACA). Experiments on five long‑context benchmarks show that GC‑OPD improves Qwen3‑4B and Qwen3‑8B checkpoints from 29.08/35.12 to 40.47/44.65, outperforming vanilla OPD and demonstrating the effectiveness of group‑relative residual calibration.
GR-OPD introduces Group-Relative On-Policy Distillation, which separates feedback computation from credit assignment by aggregating token feedback into calibrated sequence-level credit. The method shows that broadcasting summed feedback to all response tokens provides an unbiased estimate of the negative gradient of the sequence-level reverse‑KL objective under ideal on‑policy assumptions. In practice, GR‑OPD discounts later feedback to mitigate distribution shift, uses same‑prompt rollout groups for calibration, and retains a controlled fraction of collective teacher pressure, leading to superior performance over vanilla OPD across multiple Qwen3 teacher‑student settings and mathematical reasoning benchmarks.
By Yushu Wang, Renkai Yang, Jingcao Xu, Zili Geng, Haihong Tang, Yaohui Jin, Hao He
The paper introduces UECR-GRPO, a method that unifies on‑policy distillation and verifier‑based reinforcement learning for mathematical reasoning. It combines verifier rewards and teacher‑derived log‑ratios into a single KL‑regularized objective (Path‑Utility Unification) and then redistributes credit at the token level using entropy‑calibrated redistribution, preserving total task credit. Experiments on five benchmarks show that UECR‑GRPO improves average accuracy by up to 0.89 percentage points over the best baseline for both Qwen3‑1.7B and Qwen3‑4B students.
By Jie Zhang, Jingxiao Yang, Zhehao Huang, Yuhang Liu, Xiaolin Huang
The paper introduces Residual Advantage (RA), a new guidance method for reinforcement learning with verifiable rewards (RLVR) that treats the teacher–student probability residual as a bounded one‑step reward, subtracts the student’s state value to form an advantage, and centers it within each response. RA preserves the verifier’s mean label while redistributing credit among response steps, and its extension CoRA updates the teacher model using verifier advantages to adapt guidance to the student’s attempts. Experiments with Qwen3 models show that RA combined with GRPO or REINFORCE++ consistently outperforms teacher‑only on‑policy distillation across three mathematical benchmarks, improving macro Avg@8 and Pass@8 scores, with CoRA providing additional gains.
By Xiaobing Chen, Zhiqi Pang
arXiv:2604. 10688v2 Announce Type: replace-cross Abstract: On-policy reinforcement learning has become the dominant paradigm for reasoning alignment in large language models, yet its sparse, outcome-level rewards make token-level credit assignment notoriously difficult.
By Binbin Zheng, Xing Ma, Yiheng Liang, Jingqing Ruan, Xiaoliang Fu, Kepeng Lin, Benchang Zhu, Ke Zeng, Xunliang Cai
The paper investigates on‑policy distillation (OPD), showing that teacher supervision during OPD contains significant noise that grows with teacher size, yet the student policy remains largely unaffected by this noise. It finds that OPD’s gains stem mainly from suppressing low‑log‑probability tokens, a process that can be replicated without a teacher. Building on this insight, the authors propose On‑Policy Self‑Adaptation (OPSA), a supervision‑free method that uses entropy‑adaptive negative advantages to improve performance on several benchmarks, outperforming both the base model and OPD.
By Yi Ding, Ruqi Zhang