DeepStress stress-tests deep search agents with degraded evidence
DeepStress is a dynamic stress-testing framework that replaces an agent’s retrieval module with dynamically generated documents whose trustworthiness, relevance, and factuality can be controlled independently. The paper also introduces RAS and CoPRAS to assess cautious answers and computational cost, and evaluates 12 systems on HotpotQA and BrowseComp-Plus. The authors report that several open-weight agents remained confident despite obvious credibility red flags, while some HotpotQA answers came from parametric knowledge even when every retrieved document was factually wrong.
DeepStress provides a concrete way to measure whether AI search agents remain cautious when their retrieved evidence is unreliable, while also revealing overconfidence among several tested open-weight systems.
Computer Science > Computation and Language
arXiv:2607.13920 (cs)
[Submitted on 15 Jul 2026 (v1), last revised 12 Nov 2026 (this version, v2)]
Title:DeepStress: Stress-Testing Deep Search Agents
Authors:Ismael Rousseau, Geraldine Damnati, Frederic Bechet
View a PDF of the paper titled DeepStress: Stress-Testing Deep Search Agents, by Ismael Rousseau and 2 other authors