让小模型也能高效推理，通过动态调整引导强度提升性能。
G$^2$RPO-A: Guided Group Relative Policy Optimization with Adaptive Guidance
- 根据模型训练状态自动调节推理引导强度。
- 在数学和代码生成任务上显著超越基线方法。
- 适合资源有限但需强推理能力的场景。
基于可验证奖励的强化学习（RLVR）显著提升了大语言模型（LLMs）的推理能力，但其效果高度依赖于具备丰富世界知识的强基底模型，对小尺寸语言模型（SLMs）提升有限。为解决此问题，本文研究了引导式组相对策略优化（Guided GRPO），通过在回溯轨迹中注入真实推理步骤来弥补SLMs的固有缺陷。通过对多种引导配置的全面分析，发现直接添加引导效果有限。由此启发提出G²RPO-A，一种能根据模型训练动态自适应调整引导强度的算法。在数学推理与代码生成基准测试中，G²RPO-A显著优于原始GRPO方法。代码与模型已开源：https://github.com/T-Lab-CUHKSZ/G2RPO-A。
原文摘要 · Abstract (English)
Reinforcement Learning with Verifiable Rewards (RLVR) has markedly enhanced the reasoning abilities of large language models (LLMs). Its success, however, largely depends on strong base models with rich world knowledge, yielding only modest improvements for small-size language models (SLMs). To address this limitation, we investigate Guided GRPO, which injects ground-truth reasoning steps into roll-out trajectories to compensate for SLMs' inherent weaknesses. Through a comprehensive study of various guidance configurations, we find that naively adding guidance delivers limited gains. These insights motivate G$^2$RPO-A, an adaptive algorithm that automatically adjusts guidance strength in response to the model's evolving training dynamics. Experiments on mathematical reasoning and code-generation benchmarks confirm that G$^2$RPO-A substantially outperforms vanilla GRPO. Our code and models are available at https://github.com/T-Lab-CUHKSZ/G2RPO-A.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。