Tandem Reinforcement Learning with Verifiable Rewards

📄 arXiv: 2606.28166v1 📥 PDF

作者: Difan Jiao, Raghav Singhal, Robert West, Ashton Anderson

分类: cs.AI

发布日期: 2026-06-26

备注: 21 pages,7 figures,8 tables


💡 一句话要点

提出串联强化学习以解决可验证奖励问题

🎯 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: 可验证奖励 强化学习 串联训练 推理能力 多模型通信

📋 核心要点

  1. 现有的可验证奖励强化学习方法在推理能力上取得了显著进展,但在弱代理和人类的应用中存在不确定性。
  2. 论文提出的串联强化学习(TRL)通过强弱代理的交替合作生成推理,旨在提升推理的可理解性和鲁棒性。
  3. 实验表明,TRL在竞争数学任务上与传统方法相当,同时在交接鲁棒性和推理可读性上有显著提升。

📝 摘要(中文)

可验证奖励的强化学习(RLVR)显著提升了大型语言模型的推理能力,在竞争数学等领域达到了专家甚至超人类的表现。然而,较弱的代理和人类是否能够有效利用这一能力仍然不确定,RLVR已被记录为导致推理向特定模式漂移,如可读性差和语言混合。为了解决这一兼容性问题,论文提出了串联强化学习(TRL),该方法通过强大的高级代理与冻结的弱级代理交替共同生成推理,确保高级代理的推理方式更易于弱级代理理解。实验结果表明,TRL在竞争数学任务上与传统的GRPO方法在单独推理能力上相当,同时在交接鲁棒性、分布漂移和推理链的可读性等方面表现出显著优势。

🔬 方法详解

问题定义:论文要解决的问题是如何使较弱的代理和人类能够有效利用可验证奖励的强化学习(RLVR)能力。现有方法存在推理模式漂移和可读性差的问题。

核心思路:论文的核心思路是引入串联强化学习(TRL),通过强大的高级代理与冻结的弱级代理交替共同生成推理,确保高级代理的推理方式更易于弱级代理理解。

技术框架:TRL的整体架构包括两个主要模块:一个是训练好的强代理,另一个是被冻结的弱代理。两者交替生成推理,并共同获得奖励。

关键创新:TRL的最重要技术创新在于将串联训练范式引入到RLVR中,解决了传统方法中推理可读性和鲁棒性不足的问题。

关键设计:在TRL中,采用标准的GRPO损失函数来训练强代理,同时确保弱代理的生成过程不被干扰,从而实现更好的推理交接和可读性。

🖼️ 关键图片

fig_0
fig_1
fig_2

📊 实验亮点

实验结果显示,TRL在竞争数学任务上与传统的GRPO方法在单独推理能力上相当,同时在交接鲁棒性、分布漂移和推理链的可读性等方面表现出显著优势,展示了TRL在多模型通信和人类兼容性方面的潜力。

🎯 应用场景

该研究的潜在应用场景包括教育、自动化推理系统以及人机交互等领域。通过提升推理的可读性和鲁棒性,TRL能够帮助弱代理和人类更好地理解和利用复杂的推理过程,进而推动智能系统的普及和应用。

📄 摘要(原文)

Reinforcement learning with verifiable rewards (RLVR) has significantly improved the reasoning capability of large language models, reaching expert or even superhuman performance in domains such as competition math. However, whether weaker agents and humans can actually harness this capability is far less certain, with RLVR documented to drift reasoning toward idiosyncratic patterns such as poor readability and language mixing. Tandem training is a recently introduced paradigm that targets this compatibility problem: a trained, stronger senior co-generates each rollout with a frozen, weaker junior, and the two are rewarded as a team, so the senior is pushed to reason in ways the junior can follow. Yet this paradigm has so far been demonstrated only in proof-of-concept settings, leaving open whether it scales to the long chains of thought of the modern RLVR pipeline. In this work, we propose Tandem Reinforcement Learning (TRL), which carries the tandem training paradigm into RLVR. In TRL, the senior and a frozen junior alternate stochastically to co-generate the reasoning, the resulting generation is rewarded, and the standard GRPO loss is applied to the senior. Training Qwen3-4B-Instruct on competition math, we find that TRL matches vanilla GRPO on solo reasoning capability while three properties emerge together from the same rollout structure: stronger handoff robustness with the junior, reduced distributional drift from the junior, and a chain-of-thought more legible to the junior. Our results demonstrate a promising route for RLVR with practical payoffs in multi-model communication and human compatibility.