CAS: Conformalized Agentic Search via Adaptive Retrieval and Policy Weighting

TL;DR - CAS applies conformal prediction to agentic search, adapting retrieval depth and weighting reinforcement-learning trajectories by answer confidence. It aims to improve QA reasoning reliability while reducing unnecessary tool calls.

• Adaptive Prediction Sets dynamically truncate retrieved documents instead of relying on a fixed Top-K cutoff.
• Adaptive Conformal Inference estimates answer confidence with controllable coverage during training.
• CAS penalizes low-confidence trajectories within the Group Relative Policy Optimization objective.
• Experiments on single-hop and multi-hop QA report higher reasoning accuracy and substantially fewer redundant tool invocations.