Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search
2026-04-09 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors address problems with how reinforcement learning helps language models use search engines, noting that current methods explore too randomly and train unstably. They introduce Hierarchical Experience (HiExp), a system that organizes past search attempts into helpful layers using clustering and comparison. This organized experience guides the model to explore more strategically and learn more steadily. Testing shows their method improves performance and works well across different tasks and algorithms.
reinforcement learninglarge language modelssearch engineshierarchical experiencecontrastive analysisclusteringstochastic explorationtraining stabilityagentic searchmathematical reasoning
Authors
Chuzhan Hao, Wenfeng Feng, Guochao Jiang, Guofeng Quan, Guohua Liu, Yuewei Zhang
Abstract
Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.