Learning to Reason by Analogy via Retrieval-Augmented Reinforcement Fine-Tuning
作者: Zilin Xiao, Qi Ma, Chun-cheng Jason Chen, Xintao Chen, Avinash Atreya, Hanjie Chen, Vicente Ordonez
分类: cs.CL, cs.AI
发布日期: 2026-06-12
💡 一句话要点
提出基于检索增强的强化微调方法以解决复杂推理问题
🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture)
关键词: 检索增强生成 强化微调 推理能力 类比推理 数学推理 机器学习 自然语言处理
📋 核心要点
- 现有的基于检索的推理方法在处理复杂推理任务时存在局限性,无法有效区分语义相似但解决策略不同的问题。
- 本文提出的RA-RFT框架通过黄金相关性蒸馏训练检索器,优化上下文排名,进而通过强化微调提升模型的推理能力。
- 在多个数学推理基准测试中,RA-RFT相较于标准强化微调方法表现更佳,AIME 2025平均准确率提升显著。
📝 摘要(中文)
检索增强生成(RAG)已成为将语言模型与外部知识结合的标准机制,但传统的基于词汇或语义相似性的检索方法不适合复杂推理任务。本文提出了检索增强强化微调(RA-RFT),该框架通过黄金相关性蒸馏训练检索器,依据预期推理收益而非语义重叠来排名上下文,并通过强化微调方法对策略模型进行微调。实验表明,RA-RFT在数学推理基准测试中表现优异,相较于标准强化微调方法,显著提升了模型的推理能力。
🔬 方法详解
问题定义:本文旨在解决传统检索方法在复杂推理任务中的不足,尤其是无法有效利用语义相似性进行推理的局限性。
核心思路:RA-RFT框架通过黄金相关性蒸馏训练检索器,依据预期推理收益进行上下文排名,从而使模型能够通过类比推理来学习解决问题的策略。
技术框架:该框架主要包括两个阶段:首先是训练检索器以优化上下文的选择,其次是通过强化微调方法对策略模型进行微调,利用检索到的类比示例进行学习。
关键创新:RA-RFT的核心创新在于引入了推理意识的检索机制,使得模型能够在不同问题间找到潜在的类比关系,从而提升推理能力。
关键设计:在技术细节上,采用了黄金相关性蒸馏作为损失函数,优化检索器的上下文排名,同时强化微调过程中引入了可验证的结果奖励机制。
🖼️ 关键图片
📊 实验亮点
在多个数学推理基准测试中,RA-RFT相较于标准强化微调方法表现出色,AIME 2025平均准确率分别提升了7.1和2.8个百分点,显示出推理意识检索作为改进的补充方向的重要性。
🎯 应用场景
该研究的潜在应用领域包括教育、智能问答系统和复杂决策支持系统等。通过提升模型的推理能力,RA-RFT可以帮助用户更好地理解和解决复杂问题,具有重要的实际价值和广泛的应用前景。
📄 摘要(原文)
Retrieval-augmented generation (RAG) has become a standard mechanism for grounding language models in external knowledge, yet conventional retrieval based on lexical or semantic similarity is poorly suited for complex reasoning tasks: a semantically similar problem may demand an entirely different solution strategy, while a superficially different problem may share the same underlying reasoning pattern. We propose Retrieval-Augmented Reinforcement Fine-Tuning (RA-RFT), a post-training framework that teaches language models to reason by analogy. RA-RFT uses gold-relevance distillation to train a retriever that ranks contexts by expected reasoning benefit rather than semantic overlap, and then fine-tunes the policy model via reinforcement fine-tuning methods with retrieved analogous demonstrations, so the model learns to leverage reasoning traces under verifiable outcome rewards. We further analyze the diversity of retrieved contexts and find that reasoning-aware retrieval surfaces complementary solution strategies that provide distinct reasoning scaffolds for individual problems. Across challenging mathematical reasoning benchmarks, RA-RFT consistently outperforms standard reinforcement fine-tuning methods. For example, it improves AIME 2025 average@32 accuracy by 7.1 and 2.8 points over GRPO for Qwen3-1.7B and Qwen3-4B respectively -- suggesting that reasoning-aware retrieval is a complementary axis of improvement and orthogonal to advances in reward design or training curricula.