How Specific Do I Need to Be? Effects of User Query Underspecification on Deep Research Agents
Underspecified queries lower answer quality even when citations stay sound. Broader searches lean on training priors instead of the query.
The claim
Underspecified user queries make deep research agents miss the user's real information need.
The problem
Deep research agents reason, search and read for a long time.
Training and benchmarks always give them fully specified queries.
Real people often leave scope, format and depth open.
How it works
The paper introduces UNDERSPEC-DR to control how vague a query is.
It runs three deep research agents on the same tasks.
It grades final answers against the user's information need.
It also traces the searches each agent performs mid-task.
Vague queries push searches wider and away from the query.
Figures
What changes
Underspecification substantially reduces answer quality across the three agents.
Answers still cite evidence, so grounding hides the intent failure.
Limits
The evidence covers three agents and the tasks the authors ran.
The link from broad searches to weak answers is suggested, not proven.
Only the level of detail is varied; other query errors are untested.
Source · arXiv