DRNOISE: Evaluating deep research agents in misleading-evidence environments

Researchers propose DRNOISE, a benchmark containing 100 tasks for evaluating the ability of deep research agents to recover answers in misleading-evidence environments. Each task has a single correct answer, and two mutually corroborating indirect chains of records support that answer; in addition, a noise condition is set up by adding one document that appears plausible but directly gives a conflicting answer. The benchmark covers ten categories of evidence operations. In a shared agent framework, models that perform strongly on clean tasks drop in accuracy by 66 to 88 percentage points when only the aforementioned distractor document is added. Trajectory analysis shows that agents often exhibit verification laziness: they retrieve the real records, but adopt the document that directly gives the answer without adequately checking whether those records are consistent with it.