Saved in:
Bibliographic Details
Main Authors: Wang, Jiaxuan, Ouyang, Xuan, Chen, Zhiyu, Hu, Yulan, Pan, Zheng, Li, Xin, Guo, Lan-Zhe
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.10194
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917479902609408
author Wang, Jiaxuan
Ouyang, Xuan
Chen, Zhiyu
Hu, Yulan
Pan, Zheng
Li, Xin
Guo, Lan-Zhe
author_facet Wang, Jiaxuan
Ouyang, Xuan
Chen, Zhiyu
Hu, Yulan
Pan, Zheng
Li, Xin
Guo, Lan-Zhe
contents On-policy self-distillation (self-OPD) densifies reinforcement learning with verifiable rewards (RLVR) by letting a policy teach itself under privileged context. We find that when this guidance spans the full response, all-token KL spends gradients on mostly redundant positions and amplifies privileged-information leakage, causing entropy rise, shortened reasoning, and out-of-distribution degradation in long-horizon math training. We propose Token-Routed Alignment for Critical rEasoning (TRACE), which distills only on annotator-marked critical spans: forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on all remaining tokens, with the KL channel annealed away after a short warm-up. Our analysis explains TRACE through two effects: forward KL provides non-vanishing lift to teacher-supported tokens that the student under-allocates, while span masking and decay keep cumulative privileged-gradient exposure finite. On four held-out math benchmarks plus GPQA-Diamond, TRACE improves over GRPO by 2.76 percentage points on average and preserves the Qwen3-8B base OOD score on GPQA-Diamond, where GRPO and all-token self-OPD baselines degrade. Gains persist under online self-annotation (+1.90 percentage points, about 69% of the strong-API gain), reducing the concern that TRACE merely imports external annotator capability. Across scales, the best routed action is base-dependent: on Qwen3-8B it is forward KL on key spans, while on Qwen3-1.7B it shifts to reverse KL on error spans.
format Preprint
id arxiv_https___arxiv_org_abs_2605_10194
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle TRACE: Distilling Where It Matters via Token-Routed Self On-Policy Alignment
Wang, Jiaxuan
Ouyang, Xuan
Chen, Zhiyu
Hu, Yulan
Pan, Zheng
Li, Xin
Guo, Lan-Zhe
Artificial Intelligence
Machine Learning
On-policy self-distillation (self-OPD) densifies reinforcement learning with verifiable rewards (RLVR) by letting a policy teach itself under privileged context. We find that when this guidance spans the full response, all-token KL spends gradients on mostly redundant positions and amplifies privileged-information leakage, causing entropy rise, shortened reasoning, and out-of-distribution degradation in long-horizon math training. We propose Token-Routed Alignment for Critical rEasoning (TRACE), which distills only on annotator-marked critical spans: forward KL on key spans of correct rollouts, optional reverse KL on localized error spans, and GRPO on all remaining tokens, with the KL channel annealed away after a short warm-up. Our analysis explains TRACE through two effects: forward KL provides non-vanishing lift to teacher-supported tokens that the student under-allocates, while span masking and decay keep cumulative privileged-gradient exposure finite. On four held-out math benchmarks plus GPQA-Diamond, TRACE improves over GRPO by 2.76 percentage points on average and preserves the Qwen3-8B base OOD score on GPQA-Diamond, where GRPO and all-token self-OPD baselines degrade. Gains persist under online self-annotation (+1.90 percentage points, about 69% of the strong-API gain), reducing the concern that TRACE merely imports external annotator capability. Across scales, the best routed action is base-dependent: on Qwen3-8B it is forward KL on key spans, while on Qwen3-1.7B it shifts to reverse KL on error spans.
title TRACE: Distilling Where It Matters via Token-Routed Self On-Policy Alignment
topic Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2605.10194