CRISP: Critical Step Perception for Training Efficient Deep Search Agents
作者: Haosi Mo, Zihao Yan, Ruiqing Zhang, Zhongli Li, Hexuan Deng, Xuebo Liu, Min Zhang
分类: cs.CL
发布日期: 2026-08-03
备注: 15 pages, 6 figures
💡 一句话要点
提出CRISP框架以提高深度搜索代理的训练效率
🎯 匹配领域: 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 深度搜索代理 关键步骤感知 效率优化 多步交互 反向证据归纳 训练奖励机制 信息检索
📋 核心要点
- 现有深度搜索代理在多步交互中常产生冗余查询和低效探索,导致高计算和交互成本。
- CRISP框架通过关键步骤感知,区分必要证据的收集与冗余步骤,优化训练奖励以提高效率。
- 实验表明,CRISP在BrowseComp和HLE-Verified数据集上,平均交互轮次分别减少15.1%和33.2%,同时保持了最终答案的准确性。
📝 摘要(中文)
大型语言模型(LLMs)逐渐被扩展为深度搜索代理,通过与外部搜索和浏览工具的多步交互来解决复杂问题。然而,现有代理通常会产生大量计算和交互成本,生成冗长的轨迹,包含冗余查询和低效探索。本文提出CRISP框架,通过关键步骤感知来训练高效的深度搜索代理,区别必要证据的收集与冗余步骤,优化训练奖励,从而在不牺牲正确答案所需证据的情况下提高效率。实验结果表明,CRISP在保持竞争性最终答案准确率的同时,平均交互轮次分别减少了15.1%和33.2%。
🔬 方法详解
问题定义:本文旨在解决现有深度搜索代理在多步交互中产生的冗余查询和低效探索问题,导致计算和交互成本高昂。现有方法通常均匀惩罚工具使用,可能抑制必要证据的收集。
核心思路:CRISP框架通过关键步骤感知,区分收集必要证据的交互与冗余交互,优化训练奖励以保留必要步骤,从而提高搜索代理的效率。
技术框架:CRISP的整体架构包括两个主要模块:首先,通过反向证据归纳构建关键步骤标签;其次,利用小型关键步骤识别器进行全轨迹分析,最后在策略优化中应用效率感知奖励。
关键创新:CRISP的核心创新在于其能够有效区分必要证据的收集与冗余步骤,优化训练过程中的奖励机制,与现有方法的均匀惩罚策略形成鲜明对比。
关键设计:在关键步骤识别中,使用反向证据归纳技术,强模型从最终答案出发,判断每一步工具交互是否提供或保留证据。训练过程中,效率感知奖励仅应用于成功的回合,确保有效性。
🖼️ 关键图片
📊 实验亮点
在BrowseComp和HLE-Verified数据集上的实验结果显示,CRISP框架在保持最终答案准确率的同时,平均交互轮次分别减少了15.1%和33.2%。这一显著提升表明CRISP在提高深度搜索代理效率方面的有效性。
🎯 应用场景
CRISP框架具有广泛的应用潜力,尤其在需要高效信息检索和决策支持的领域,如智能客服、在线搜索引擎和自动化问答系统。通过提高搜索代理的交互效率,能够显著降低计算成本并提升用户体验,未来可能推动更智能的搜索工具的发展。
📄 摘要(原文)
Large language models (LLMs) are increasingly extended into deep search agents that solve complex questions through multi-step interaction with external search and browsing tools. However, existing agents often incur substantial computational and interaction costs, generating lengthy trajectories that contain redundant queries, inefficient exploration, and irrelevant observations. Existing efficiency-oriented methods usually encourage agents to use tools less frequently, but treating all tool interactions uniformly may also suppress steps that gather necessary evidence. In this paper, we propose CRISP, a framework for training efficient deep search agents through critical step perception. Unlike prior efficiency methods that uniformly penalize tool use, CRISP distinguishes interactions that gather necessary evidence from redundant ones and shapes the training reward to preserve the former while pruning the latter, improving efficiency without sacrificing the evidence needed for correct answers. Specifically, CRISP first constructs critical-step labels with Backward Evidence Induction: starting from the final answer, a strong model traverses a completed search trajectory backward and judges whether each tool-interaction step provides or preserves evidence for the final answer. We then distill these step-wise judgments into a smaller critical-step recognizer, enabling full-trajectory analysis in a single pass. During policy optimization, an efficiency-aware reward is applied only to successful rollouts. Experiments on BrowseComp and HLE-Verified show that CRISP maintains competitive final-answer accuracy while reducing average interaction turns by 15.1% and 33.2%, respectively, demonstrating substantial improvements in interaction efficiency.