Grounded Credit Assignment in Reinforcement Learning for Multi-Turn Search Agents
作者: Qirui Liu, Hao Chen, Yi Wen, Weijie Shi, Yao Zhao, Xianhe Chen, Hao Zhao, Xing Tang, Weihong Luo, Yuyu Luo
分类: cs.LG
发布日期: 2026-10-16
备注: 27 pages
💡 一句话要点
提出基于证据的信用分配方法以解决多轮搜索代理的奖励问题
🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture)
关键词: 强化学习 信用分配 多轮问答 证据检索 模型优化
📋 核心要点
- 现有的GRPO方法在信用分配上存在不足,未能有效区分搜索和答案的贡献,导致准确性下降。
- 论文提出的GCA方法通过引入证据检索和答案正确性两个可验证的子结果,分别调整搜索和答案的信用分配。
- 实验结果显示,GCA方法在多个问答基准上显著提升了准确性,尤其在证据支持的正确答案比例上从76.9%提升至93.4%。
📝 摘要(中文)
现有的结果奖励强化学习方法如GRPO对搜索代理的每个回合都给予相同的优势,未能有效区分其在证据查找和利用方面的能力。我们提出了基于证据的信用分配(GCA)方法,通过可验证的子结果(证据检索和答案正确性)分别纠正搜索和答案的信用。实验结果表明,GCA在不增加额外模型或训练阶段的情况下,显著提高了多个问答基准的准确性,尤其在Qwen2.5-3B和Qwen3-30B-A3B模型上,平均EM分别提升了5.7和6.3个百分点。
🔬 方法详解
问题定义:论文旨在解决现有GRPO方法在信用分配上的不足,导致搜索代理在多轮问答中未能有效利用证据,影响整体准确性。
核心思路:GCA方法通过引入可验证的子结果,分别对证据检索和答案的正确性进行信用分配,从而更准确地反映各自的贡献。
技术框架:GCA的整体架构包括两个主要模块:证据检索模块和答案评估模块。前者负责查找相关证据,后者则评估答案的正确性,并根据这两个模块的输出进行信用调整。
关键创新:GCA的核心创新在于其能够独立评估证据检索和答案正确性,避免了传统方法中对错误答案的过度惩罚,同时确保对有效证据检索的信用不被削弱。
关键设计:GCA方法的设计包括对每个token的优势进行限制,确保其信用调整不超过10%,并且不影响最终的结果信号。
🖼️ 关键图片
📊 实验亮点
实验结果显示,GCA方法在多个问答基准上显著提升了准确性,Qwen2.5-3B模型的平均EM提升了5.7个百分点,Qwen3-30B-A3B模型提升了6.3个百分点。同时,证据支持的正确答案比例从76.9%提升至93.4%。
🎯 应用场景
该研究的潜在应用场景包括智能问答系统、搜索引擎优化和对话系统等领域。通过更精确的信用分配,能够提升系统在复杂查询中的表现,增强用户体验,未来可能推动更智能的搜索代理和问答模型的发展。
📄 摘要(原文)
Outcome-reward reinforcement learning such as GRPO gives every token of a search-agent rollout the same advantage, although the rollout exercises two capabilities: finding evidence and using it. We show that accuracy improves without grounding: Search-R1 raises accuracy on all seven QA benchmarks, yet on Bamboogle the evidence-supported share of correct answers falls from 84.0% to 42.9%. We attribute this to GRPO's credit assignment: rollouts that retrieve answer-bearing evidence are favoured in aggregate only in proportion to their accuracy gain over the rest, a grounding gap that answering from memory or misusing evidence shrinks or reverses. To address this, we propose Grounded Credit Assignment (GCA), which uses verifiable sub-outcomes (evidence retrieval and answer correctness) to correct search and answer credit separately: a wrong answer is blamed more than the search that found the evidence, and a correct answer that bypasses retrievable evidence is credited less. Never negative on evidence-finding search, the correction does not shrink with the grounding gap; capped at 10% of each token's advantage, it never overrides the outcome signal. Without extra models, rollouts, or training stages, GCA improves average EM over Search-R1 by 5.7 (Qwen2.5-3B) and 6.3 points (Qwen3-30B-A3B) and raises this share from 76.9% to 93.4%.