DeepStress: A Dynamic Stress Testing Framework for Search Agents

Written by

in

DeepStress is a dynamic stress testing framework that generates low-quality evidence for search agents at controlled ratios while independently controlling the credibility, relevance, and authenticity of documents. The paper evaluates 12 systems on HotpotQA and BrowseComp-Plus, introduces five categories for classifying final answers, and proposes two reliability metrics—RAS and CoPRAS—that account for cautious behavior and computational cost.

Source:Hugging Face 社区论文精选 · Read original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *