跳转至

CAS：通过自适应检索与策略加权的保形智能体搜索

文章背景与核心概要

在强化学习（RL）微调阶段，搜索智能体（Search Agents）常常面临严重的可靠性问题。传统的启发式 Top-K 检索方法经常导致关键证据丢失或引入噪声，而强化学习进程中逐渐累积的过度自信则会导致答案幻觉和冗余搜索。

为了克服这些局限性，本文提出了 CAS（Conformalized Agentic Search，保形智能体搜索） 框架。该框架引入了保形预测（Conformal Prediction, CP），在检索和训练两个层面上提供严格的可靠性保障：一方面，它通过自适应预测集（APS）实现动态文档截断，确保检索结果的规模适应性；另一方面，它利用自适应保形推断（ACI）量化答案置信度，并在群组相对策略优化（GRPO）目标中对低置信度轨迹进行惩罚，从而确保模型仅从可靠的路径中进行学习。

📋 Summary

Search agents often encounter reliability issues during reinforcement learning (RL) fine-tuning. Standard heuristic Top-K retrieval frequently introduces critical evidence loss or noise, while progressive RL-induced over-confidence leads to answer hallucinations and redundant searches. To overcome these limitations, CAS (Conformalized Agentic Search) introduces Conformal Prediction (CP) to provide strict reliability guarantees on both the retrieval and training fronts: * Adaptive Retrieval via APS: Uses Adaptive Prediction Sets (APS) to translate statistical coverage into dynamic document truncation, ensuring prediction sets are appropriately sized. * Reliable Policy Training via ACI: Employs Adaptive Conformal Inference (ACI) to dynamically construct prediction sets with controllable coverage to quantify answer confidence. This confidence is applied to penalize low-confidence trajectories within the Group Relative Policy Optimization (GRPO) objective, ensuring the model trains exclusively on reliable paths.

📌 Metadata

  • arXiv ID: arXiv:2608.20771 [cs.AI]
  • Subject Area: Artificial Intelligence (cs.AI)
  • Publication Date: August 21, 2026
  • Authors: Zixi Zhu, Jiayuan Su, Jian Zhang, Yu Lin, Hongwei Wang
  • Official Code Repository: GitHub - S1llyBird/CAS

🔍 Abstract

Search Agents face a severe reliability crisis during reinforcement learning (RL) fine-tuning. Heuristic Top-K retrieval often causes critical evidence loss or noise inclusion, while over-confidence induced by progressive RL leads to hallucinated answers and redundant searches.

To build highly reliable agents, we introduce Conformal Prediction (CP) and propose Conformalized Agentic Search (CAS). This framework establishes reliability guarantees on both the retrieval and training sides: 1. On the retrieval side: an Adaptive Prediction Set (APS), a specific CP realization, translates statistical coverage into dynamic document truncation to construct prediction sets that are adaptive in size. 2. On the training side: Adaptive Conformal Inference (ACI), a dynamic CP algorithm, dynamically constructs prediction sets with controllable coverage to quantify answer confidence, which is then used to penalize low-confidence trajectories within the Group Relative Policy Optimization (GRPO) objective, ensuring the model learns only from reliable ones.

Experiments across single-hop and multi-hop QA datasets demonstrate that our framework significantly improves reasoning accuracy while drastically reducing redundant tool invocations, establishing a highly reliable and efficient agent paradigm.

🔗 Full-Text & References