Conformalized Agentic Search Mitigates Hallucinations via Adaptive Retrieval and Policy Weighting

Conformalized Agentic Search (CAS) applies Conformal Prediction to agentic workflows to reduce hallucination and evidence loss. The framework uses Adaptive Prediction Sets for dynamic document truncation and Adaptive Conformal Inference to penalize low-confidence trajectories during reinforcement learning fine-tuning.