文章背景与核心概要
群组相对策略优化(Group Relative Policy Optimization, GRPO)是大语言模型(LLMs)中一种备受瞩目的强化学习算法。然而,它存在一个结构性不稳定问题:通过将 PPO 风格的词元级裁剪(token-level clipping)与序列级优势(sequence-level advantages)相结合,它在(似然比,优势)空间中留下了一个“盲区”。具体而言,在第四象限(负优势伴随似然比增加)中,模型会接收到过大的抑制性更新,从而导致熵坍塌(entropy collapse)和推理边界的缩小。
为此,本文作者推出了全象限有界裁剪 GRPO(All-Quadrant Bounded Clipping GRPO, ABC-GRPO)。该方法在所有四个象限中应用了依赖符号的无条件裁剪。通过在第二象限实现信任域下限并在第四象限实现上限,ABC-GRPO 确保了有界的策略位移。在 Qwen3 基础模型上的实验结果表明,ABC-GRPO 优于标准的 GRPO、SAPO 以及双裁剪 PPO(dual-clip PPO),在数学推理(Avg@64 和 Pass@64)上表现出卓越的性能,并对代码生成(HumanEval)展现出很强的泛化能力。
全象限有界裁剪 GRPO:消除无界盲区以实现稳定且泛化的训练
作者: Chi Liu, Xin Chen
arXiv: 2601.03895 [cs.LG]
提交时间: 2026年1月7日 (v1), 2026年8月6日 (v2)
摘要 (Summary)
Group Relative Policy Optimization (GRPO) is a prominent reinforcement learning algorithm for Large Language Models (LLMs). However, it suffers from a structural instability: by combining PPO-style token-level clipping with sequence-level advantages, it leaves a "blind spot" in the (likelihood-ratio, advantage) space. Specifically, in the fourth quadrant (negative advantage with an increased likelihood ratio), the model can receive excessively large suppressive updates, leading to entropy collapse and narrowed reasoning boundaries.
群组相对策略优化(GRPO)是大语言模型(LLMs)中一种备受瞩目的强化学习算法。然而,它存在一个结构性不稳定问题:通过将 PPO 风格的词元级裁剪与序列级优势相结合,它在(似然比,优势)空间中留下了一个“盲区”。具体而言,在第四象限(负优势伴随似然比增加)中,模型会接收到过大的抑制性更新,从而导致熵坍塌和推理边界的缩小。
The authors introduce All-Quadrant Bounded Clipping GRPO (ABC-GRPO), which applies sign-dependent, unconditional clipping across all four quadrants. By implementing a trust-region floor in the second quadrant and a cap in the fourth, ABC-GRPO ensures bounded policy displacement. Empirical results on Qwen3 base models demonstrate that ABC-GRPO outperforms standard GRPO, SAPO, and dual-clip PPO, showing superior performance on mathematical reasoning (Avg@64 and Pass@64) and strong transferability to code generation (HumanEval).
为此,作者推出了全象限有界裁剪 GRPO(ABC-GRPO),该算法在所有四个象限中应用了依赖符号的无条件裁剪。通过在第二象限实现信任域下限并在第四象限实现上限,ABC-GRPO 确保了有界的策略位移。在 Qwen3 基础模型上的实验结果表明,ABC-GRPO 优于标准的 GRPO、SAPO 以及双裁剪 PPO,在数学推理(Avg@64 和 Pass@64)上表现出卓越的性能,并对代码生成(HumanEval)展现出很强的泛化能力。
核心贡献 (Key Contributions)
- Four-Quadrant Analysis: Identifies that the combination of token-level clipping and sequence-level advantages creates an unbounded update mechanism in the Q4 quadrant.
- ABC-GRPO Algorithm: Proposes a novel clipping mechanism that applies constraints before multiplying by the advantage, effectively capping suppressive updates and preventing policy collapse.
- Mathematical & Generalization Gains: Achieves state-of-the-art performance on benchmarks like MATH-500 and HumanEval while maintaining higher policy entropy compared to baseline methods.
- Ablation Studies: Confirms that the Q4 quadrant is the primary source of instability in existing GRPO implementations.
- 四象限分析: 发现词元级裁剪与序列级优势的组合在 Q4 象限中产生了一个无界的更新机制。
- ABC-GRPO 算法: 提出了一种新型裁剪机制,在与优势相乘之前应用约束,从而有效地限制了抑制性更新并防止了策略坍塌。
- 数学与泛化增益: 在 MATH-500 和 HumanEval 等基准测试中达到了 SOTA(最先进的)性能,同时与基线方法相比保持了更高的策略熵。
- 消融实验: 证实了 Q4 象限是现有 GRPO 实现中不稳定的主要来源。
访问与元数据 (Access & Metadata)
- Full-text Links: View PDF | HTML (Experimental) | TeX Source
License: Creative Commons Attribution 4.0 International
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
- DOI: https://doi.org/10.48550/arXiv.2601.03895
- 全文链接: 查看 PDF | HTML(实验性) | TeX 源码
-
许可证: 知识共享署名 4.0 国际许可协议
-
学科分类: 机器学习 (cs.LG);人工智能 (cs.AI);计算与语言 (cs.CL)
- DOI: https://doi.org/10.48550/arXiv.2601.03895
