Adaptive-Boundary-Clipping GRPO: Ensuring Bounded Ratios for Stable and Generalizable Training

Chi Liu, Xin Chen
1/7/2026
cs.LGcs.AIcs.CL

Official code: chi2liu/ABC-GRPO (442 stars, Apache-2.0, last commit 2026-03-26).

Abstract

Group Relative Policy Optimization (GRPO) has emerged as a popular algorithm for reinforcement learning with large language models (LLMs). However, upon analyzing its clipping mechanism, we argue that it is suboptimal in certain scenarios. With appropriate modifications, GRPO can be significantly enhanced to improve both flexibility and generalization. To this end, we propose Adaptive-Boundary-Clipping GRPO (ABC-GRPO), an asymmetric and adaptive refinement of the original GRPO framework. We demonstrate that ABC-GRPO achieves superior performance over standard GRPO on mathematical reasoning tasks using the Qwen3 LLMs. Moreover, ABC-GRPO maintains substantially higher entropy throughout training, thereby preserving the model's exploration capacity and mitigating premature convergence. The implementation code is available online to ease reproducibility https://github.com/chi2liu/ABC-GRPO.

View on arXivView PDF

Code Implementations(1)

chi2liu/ABC-GRPOOfficial100%
Apache-2.0

Code For Adaptive-Boundary-Clipping GRPO. arxiv.org/pdf/2601.03895

44242Shell, PythonJan 7, 20266 months agoApache-2.0
grpollmreinforcement-learningrlhf

Similar Papers

GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Models

88% match1 repo382025

GRPO-RM: Fine-Tuning Representation Models via GRPO-Driven Reinforcement Learning

88% match2025

Prompt Augmentation Scales up GRPO Training on Mathematical Reasoning

88% match1 repo132026

M-GRPO: Stabilizing Self-Supervised Reinforcement Learning for Large Language Models with Momentum-Anchored Policy Optimization

88% match2025

BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping

88% match1 repo942025

On the Theory and Practice of GRPO: A Trajectory-Corrected Approach with Fast Convergence

88% match2025

Cite this paper

@article{liu2026adaptiveboundaryclipping,
  title  = {Adaptive-Boundary-Clipping GRPO: Ensuring Bounded Ratios for Stable and Generalizable Training},
  author = {Chi Liu and Xin Chen},
  year   = {2026},
  eprint = {2601.03895},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2601.03895v1}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2601.03895.svg)](https://codeofpaper.com/paper/2601.03895)

Discussion