AI 代理與 RAG
HALT Uses Evidence Coverage to Decide When RAG Agents Should Stop, Cutting Search Loops by Up to 45%
HALT no longer relies on a generative model’s confidence to determine whether a search is complete. Instead, it verifies each claim required to answer a multi-hop question against the available evidence. It can be added to frozen search agents, but the savings depend heavily on whether accurate intermediate claims can be generated in advance.
Retrieval-augmented search agents typically issue queries repeatedly until the model determines that it has enough information. The problem is that agents may continue searching even after the necessary evidence has appeared, increasing latency, inference costs, and distracting content. Submitted on August 3, HALT reframes the stopping condition as a verifiable “evidence coverage” problem. The system first decomposes a question into the expected multi-hop claims, then uses a small supervised verifier to classify the relationship between each claim and the accumulated documents as Match, Partial, or Null. It stops only when every claim is matched; otherwise, retrieval continues.
This control layer does not modify the original agent’s policy, retriever, or backbone model. The public implementation trains a three-class verifier using Qwen2.5-3B-Instruct with LoRA, runs inference with 4-bit NF4 quantization, and evaluates it on HotpotQA, 2WikiMultihopQA, and MuSiQue. With manually annotated gold hop claims, it reduces average search loops by 19% to 45%. When the claims are generated automatically from the questions and transferred to 7B agents, the savings fall to roughly 6% to 9%. The authors evaluate exact match using paired bootstrap tests, Holm correction, and a two-percentage-point non-inferiority margin, with most test configurations maintaining non-inferior results.
HALT’s engineering value lies in its ability to serve as a runtime gate outside an existing RAG agent, eliminating the need to retrain the entire agent. However, the headline results still come from closed datasets with predefined supporting facts. On the open web, conflicting documents, uneven source quality, and incomplete indexing could cause the verifier to abstain. The repository also does not directly include the datasets or some artifacts, so full reproduction requires downloading the corpora and indexes separately. The next questions are how often automatic claim decomposition fails and whether the latency added by the extra verification model is truly offset by the reduction in search calls.