Search agents encounter a severe reliability crisis during reinforcement‑learning fine‑tuning. Conventional Top‑K retrieval often discards crucial evidence or admits noise, while progressive RL induces over‑confidence, leading to hallucinated answers and redundant searches.
To address these issues, we introduce Conformal Prediction (CP) and propose the Conformalized Agentic Search (CAS) framework. CAS provides reliability guarantees on both the retrieval and training sides.
On the retrieval side, CAS employs an Adaptive Prediction Set (APS) that translates statistical coverage $1-\alpha$ into dynamic document truncation. APS automatically adjusts the number of returned documents, forming prediction sets whose size adapts to the desired coverage, thus preserving essential evidence while avoiding unnecessary items.
On the training side, CAS utilizes Adaptive Conformal Inference (ACI) to construct confidence sets for answers on the fly. These confidence intervals are incorporated into the Group Relative Policy Optimization (GRPO) objective, penalizing low‑confidence trajectories so that the model learns only from reliable interactions.
Experiments on single‑hop and multi‑hop QA benchmarks demonstrate that CAS markedly improves reasoning accuracy and cuts redundant tool invocations by over 60%, establishing a highly reliable and efficient agent paradigm. The code is released at https://github.com/S1llyBird/CAS.
Blogger's Review: CAS ingeniously integrates conformal prediction into both retrieval and RL pipelines, offering quantifiable reliability for intelligent agents and promising broader applicability in tool‑augmented AI systems.