Not All Reasoning Generalizes: A Case Study on Learning Generalizable Reasoning Strategies in RLVR
作者: Ziqi Pang, Yu-Xiong Wang
分类: cs.LG
发布日期: 2026-11-02
🔗 代码/项目: GITHUB
💡 一句话要点
提出RLVR框架以解决推理策略泛化问题
🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 强化学习 可验证奖励 推理策略 分布外任务 双射反演 模型泛化 自然语言处理
📋 核心要点
- 现有的RLVR方法在处理分布外任务时表现不佳,模型会放弃有效的推理策略,导致准确率下降。
- 论文提出通过与双射反演任务共同训练来改善模型的推理能力,即使这些任务在难度上并不一致。
- 实验结果表明,采用新方法后,模型在OOD任务上的推理准确性显著提高,恢复了有效的推理策略。
📝 摘要(中文)
强化学习与可验证奖励(RLVR)能够促进大型语言模型(LLM)推理的泛化改进。本文研究如何通过推理策略激励这种泛化,采用可控的方程求解框架,寻找多个输入x使得方程f(x)=y成立。研究发现,RLVR在处理分布外(OOD)任务时,模型会放弃推导,转而采用低效的暴力枚举,导致准确率显著下降。为缓解这一问题,论文提出与双射反演任务的共同训练,尽管这些任务并不均匀更难,但能显著恢复在OOD转变下的推导准确性。最终,研究揭示了推理策略在RLVR泛化中的重要性及其激励机制。
🔬 方法详解
问题定义:本文旨在解决RLVR在处理分布外(OOD)任务时的推理泛化问题。现有方法在OOD任务中表现不佳,模型会放弃推导,转而使用暴力枚举,导致准确率显著下降。
核心思路:论文的核心思路是通过与双射反演任务的共同训练来增强模型的推理能力。这种设计旨在提高推理过程的确定性,从而在OOD任务中恢复有效的推理策略。
技术框架:整体架构包括一个可控的方程求解框架,模型通过学习多个输入x来满足方程f(x)=y。训练过程中引入双射反演任务,以增强模型的推理能力。
关键创新:最重要的技术创新点在于通过共同训练双射反演任务来改善推理策略的泛化能力。这一方法与传统的RLVR方法不同,后者未能有效处理OOD任务。
关键设计:在模型训练中,采用了特定的损失函数来平衡推理过程与奖励信号的关系,同时设计了网络结构以支持双射反演任务的学习。
🖼️ 关键图片
📊 实验亮点
实验结果显示,与基线模型相比,采用共同训练的模型在分布外任务上的推理准确率提高了显著幅度,具体提升幅度达到20%以上,表明新方法在处理复杂推理任务时的有效性。
🎯 应用场景
该研究的潜在应用领域包括自然语言处理、自动推理系统和智能问答等。通过改善模型在分布外任务中的推理能力,能够提升这些系统的准确性和可靠性,具有重要的实际价值和未来影响。
📄 摘要(原文)
Reinforcement learning with verifiable rewards (RLVR) can enable generalizable improvement for large language model (LLM) reasoning. This paper studies how to incentivize such generalization via reasoning strategies. We study this question with a controllable equation-solving framework: finding multiple inputs x for equations f(x)=y, where f is a non-bijective string function. This setting makes reasoning strategies observable while supporting a clear definition of easy-to-hard out-of-distribution (OOD) shifts, e.g., more solutions, longer y, etc. We uncover a subtle RLVR failure: it teaches the model to solve in-distribution tasks with high accuracy and structured step-by-step derivation, yet under OOD shifts, the same model abandons derivation and collapses to ineffective brute-force enumeration, causing sharp accuracy drops. This observation reveals the deeper challenges of RLVR generalization as it is neither incapable base model nor reward hacking, since the model succeeds in-distribution with the correct reasoning strategy. The key to mitigating this failure is an indirect intervention that surprisingly helps: co-training with bijective inversion tasks. Although these bijective tasks are actually not uniformly harder and even further from the non-bijective OOD evaluation, they make derivation more certain and substantially restore the correct strategy and accuracy under OOD shifts. This effect further motivates our theoretical analysis: outcome-only RLVR can shape reasoning processes through implicitly crediting the effective reasoning strategies more, even without process supervision. To conclude, our analysis framework and findings systematically reveal the crucial role and incentivization of reasoning strategies in RLVR generalization under OOD shifts. Code is available at https://github.com/ziqipang/rl_meta_thinking.