Deep Research Agent
An AI agent that synthesizes information into analytical reports through multi-step web retrieval. It possesses autonomous research and report generation capabilities.
Timeline (last 90 days)
Deep research agents typically consist of three parts: retrieval tools, a reasoning loop, and answer synthesis.
The research team evaluated three DRAs and found that when user queries are underspecified, the degree to which agent answers satisfy users' information needs drops significantly.
When facing ambiguous queries, DRAs conduct broader searches, but these searches rely less on the original user query and more on prior knowledge the model learned during training.
Deep research agents (DRAs) are typically trained and evaluated on fully specified user queries.
Most benchmarks construct complete queries through human completion or templating, thereby systematically overestimating agents' actual reliability.
Future work needs to build DRAs that can proactively identify when additional user input is needed, requesting clarification from users when necessary or narrowing scope through interactive questioning.
When uncertain, agents tend to improvise freely rather than actively clarifying or narrowing the search scope.
All Facts (7)
Deep research agents typically consist of three parts: retrieval tools, a reasoning loop, and answer synthesis.
50%UnverifiedThe research team evaluated three DRAs and found that when user queries are underspecified, the degree to which agent answers satisfy users' information needs drops significantly.
50%UnverifiedWhen facing ambiguous queries, DRAs conduct broader searches, but these searches rely less on the original user query and more on prior knowledge the model learned during training.
50%UnverifiedDeep research agents (DRAs) are typically trained and evaluated on fully specified user queries.
50%UnverifiedMost benchmarks construct complete queries through human completion or templating, thereby systematically overestimating agents' actual reliability.
50%UnverifiedWhen uncertain, agents tend to improvise freely rather than actively clarifying or narrowing the search scope.
50%UnverifiedFuture work needs to build DRAs that can proactively identify when additional user input is needed, requesting clarification from users when necessary or narrowing scope through interactive questioning.
50%