Deep Research agents, which extend LLM-based assistants into long-horizon workflows involving planning, retrieval, evidence synthesis, and report generation, are increasingly used for complex information tasks. However, a new study from Hugging Face researchers highlights a critical reliability gap: these agents can be misled by apparently credible but factually false information encountered during their research process.
The researchers introduce MisKnow-Agent, a framework for constructing and validating misleading knowledge for Deep Research tasks. It generates misleading instances with controllable authority levels and styles, yielding 5,933 quality-controlled instances built on DeepResearch Benchmark tasks.
Experiments across open-source and closed-source Deep Research agents, including DeerFlow, WebThinker, and Gemini Deep Research, show that introducing just one misleading document increases the mean false-conclusion adoption rate (FCAR) from 0% in the control to 54.7%. FCAR varies with lifecycle stage, framework design, source authority, and presentation style, while search-result rank and additional documents have limited influence.
Notably, while search-enabled verifier models consistently identify the retained instances as misleading during focused corpus validation, the same instances can still be adopted during long-horizon research, revealing a disconnect between focused verification and workflow-level evidence use.
The study evaluates pre- and post-research defenses, both individually and in combination, finding that all three configurations mitigate but do not fully prevent false-conclusion adoption. The authors conclude that reliable Deep Research requires evidence verification and correction capabilities at both the model and framework levels, beyond improvements in planning, retrieval, evidence integration, or report-generation abilities.