DeepStress Benchmark: 12 Deep Search Agents Answer Confidently From Sources With Obvious Credibility Red Flags

Orange Research and Aix-Marseille researchers release DeepStress, a dynamic stress-testing framework that replaces search agents' retrieval modules with synthetic documents whose trustworthiness, relevance and factuality are independently controlled. Across 12 systems tested on HotpotQA and BrowseComp-Plus, several open-weight agents answer with high confidence from obviously flagged sources, and many HotpotQA questions get answered from parametric knowledge even when every returned document is factually wrong. The paper introduces reliability-aware RAS and CoPRAS metrics.