DeepStress: Stress-Testing Deep Search Agents
Several open-weight search agents answered confidently despite obvious credibility red flags in their sources.
DeepStress replaces retrieval with dynamically generated documents whose trustworthiness, relevance and factuality can be controlled independently. Researchers tested 12 systems on questions from HotpotQA and BrowseComp-Plus. Many HotpotQA questions were answered from model knowledge even when every returned document was factually wrong. The framework also introduces five answer-outcome categories and two metrics that reward caution and account for computational cost.
HF Daily Papers' note
DeepStress replaces retrieval with dynamically generated documents whose trustworthiness, relevance and factuality can be controlled independently. Researchers tested 12 systems on questions from HotpotQA and BrowseComp-Plus. Many HotpotQA questions were answered from model knowledge even when every returned document was factually wrong. The framework also introduces five answer-outcome categories and two metrics that reward caution and account for computational cost.
HF Daily Papers' note
score 5