2606.10968v1 Jun 09, 2026 cs.LG

Beyond Uniform Token-Level Trust Region in LLM Reinforcement Learning

Yujia Liu
Yujia Liu
Citations: 0
h-index: 0
Xiangxin Zhou
Xiangxin Zhou
Citations: 69
h-index: 3
Renjie Mao
Renjie Mao
Citations: 35
h-index: 3
Lvfang Tao
Lvfang Tao
Citations: 16
h-index: 2
Yi Ding
Yi Ding
Citations: 56
h-index: 2
Yucheng Shi
Yucheng Shi
Citations: 2
h-index: 1
Yong Lin
Yong Lin
Citations: 762
h-index: 9
Honglin Zhu
Honglin Zhu
Citations: 7
h-index: 1
Qiang Qiu
Qiang Qiu
Citations: 21
h-index: 2
Wenxi Zhu
Wenxi Zhu
Citations: 26
h-index: 3

Reinforcement learning with verifiable rewards (RLVR) has become standard for improving LLM reasoning. However, existing PPO-style trust-region mechanisms remain position-agnostic by enforcing uniform thresholds across all tokens independently. This pointwise treatment conflicts with autoregressive generation in two critical ways. First, uniform thresholds ignore autoregressive asymmetry. Early-stage deviations produce compounding sequence-level drift, causing static thresholds to under-regulate early divergence and excessively constrain late-stage exploration. Second, evaluating token-level divergence in isolation overlooks cumulative prefix drift, granting the same divergence allowance regardless of how far the conditioning history has already deviated from the rollout policy. To address this limitation, we propose CPPO (Cumulative Prefix-divergence Policy Optimization), a token-level masking rule that aligns updates with a finite-horizon policy-improvement bound via two coupled mechanisms. First, a position-weighted threshold imposes stricter limits at early positions whose effects persist longer, relaxing constraints for late-stage tokens. Second, a cumulative prefix budget tracks historical deviations, dynamically restricting further token-level deviation to prevent compounding errors along the prefix. Empirically, CPPO enhances training stability and significantly improves reasoning accuracy across various model scales.

0 Citations
0 Influential
4.5 Altmetric
22.5 Score
Original PDF

No Analysis Report Yet

This paper hasn't been analyzed by Gemini yet.

Log in to request an AI analysis.

댓글

댓글을 작성하려면 로그인하세요.

아직 댓글이 없습니다. 첫 번째 댓글을 남겨보세요!