English

GTPO and GRPO-S: Token and Sequence-Level Reward Shaping with Policy Entropy

Computation and Language 2026-02-06 v6 Artificial Intelligence

Abstract

Reinforcement Learning (RL) is pivotal for enhancing Large Language Model (LLM) reasoning, yet mainstream algorithms such as GRPO and DAPO remain constrained by a coarse-grained credit assignment paradigm, where all tokens within the same response receive the identical reward. In this paper, we propose Dynamic Entropy Weighting, systematically define entropy-based weight ratios Hi,tk=1nHk,t\frac{H_{i,t}}{\sum_{k=1}^{n} H_{k,t}} and similar variants to redistribute rewards and get fine-grained rewards through two new algorithms: Group Token Policy Optimization (GTPO), which assigns an entropy-weighted reward to each token and synthesizes token-specific advantage function to drive the model toward optimal path, and the analogous algorithm Sequence-Level GRPO (GRPO-S), which extends this design to the sequence level and exhibits superior stability in long Chain-of-Thought (CoT) reasoning tasks.

Keywords

Cite

@article{arxiv.2508.04349,
  title  = {GTPO and GRPO-S: Token and Sequence-Level Reward Shaping with Policy Entropy},
  author = {Hongze Tan and Zihan Wang and Jianfei Pan and Jinghao Lin and Hao Wang and Yifan Wu and Tao Chen and Zhihang Zheng and Zhihao Tang and Haihua Yang},
  journal= {arXiv preprint arXiv:2508.04349},
  year   = {2026}
}