HALT: A Verification-Aware Stopping Policy for Retrieval-Augmented Search Agents

Researchers have unveiled HALT, a streamlined verification-aware stopping policy aimed at enhancing retrieval-augmented search agents that tackle multi-hop inquiries by progressively submitting search queries and gathering evidence. The primary challenge tackled is the decision to stop: once enough evidence is collected, further retrieval can incur costs, delays, and irrelevant context. HALT conceptualizes stopping in terms of evidence coverage instead of generator confidence, maintaining the original search agent's functionality. The policy functions based on anticipated hop claims, ceasing operations only when the accumulated evidence substantiates each necessary claim. Tested on three multi-hop question-answering benchmarks, HALT minimizes unnecessary searches while largely maintaining exact match accuracy. The research distinguishes between a deployable scenario, where hop claims derive from the question, and a diagnostic upper limit using gold supporting-fact annotations. Generated claims provide modest yet exact-match-preserving savings, whereas gold claims show more significant savings. The paper can be accessed on arXiv with the identifier 2608.02009.