VeriSearch: A Verification-Enhanced Search Agent with Proactive Evidence Filtering and Conflict Resolution

Search agents built on large language models (LLMs) answer knowledge-intensive questions by retrieving and reading webpages during inference. This allows them to supplement static parametric memory with external evidence. Existing agents primarily optimize search execution, while retrieved webpages may not provide sufficient information for the current question, and different sources may present conflicting claims. These issues can contribute to error propagation across subsequent reasoning steps. We propose VeriSearch, a verification-enhanced search agent that introduces explicit evidence control into the search workflow. Before full-page crawling, we design AnswerabilityPruner, which estimates the answerability of each search result from its search snippet and triggers a reflective retry when no candidate passes the answerability threshold. After crawling, ConflictResolver detects cross-source factual conflicts among retrieved evidence and resolves them through weighted evidence arbitration. With Qwen3.5-9B as the backbone, VeriSearch outperforms Search-o1 by an average of 10.3 points across three benchmarks and exceeds Claude-4-Sonnet's reported results by an average of 5.0 points on xbench-DeepSearch and SEAL-0. Code is included in the supplementary material and will be made publicly available upon publication.