Conformalized Agentic Search Boosts AI Reliability and Efficiency

This research addresses the reliability challenges faced by search agents during reinforcement learning (RL) fine-tuning, such as critical evidence loss from heuristic retrieval and hallucinations from over-confidence. The proposed solution is Conformalized Agentic Search (CAS), a novel framework that integrates Conformal Prediction (CP) to establish statistical reliability guarantees. CAS operates on two fronts: for retrieval, it uses an Adaptive Prediction Set (APS) to dynamically truncate documents, ensuring robust prediction sets. For training, it employs Adaptive Conformal Inference (ACI) to quantify answer confidence, which then penalizes low-confidence trajectories within the Group Relative Policy Optimization (GRPO) objective. This dual approach ensures that the model learns primarily from reliable data. Experimental results on single-hop and multi-hop QA datasets demonstrate that CAS significantly improves reasoning accuracy and drastically reduces unnecessary tool invocations, leading to a more reliable and efficient agent paradigm.