Web research agents grade their own steps too kindly

A new benchmark called IBIS holds an agent's search results fixed and varies only who wrote the step before them, then scores worse when the agent did. Researchers call the effect inertia bias: an agent grades its own plans more leniently, even when the evidence hasn't changed.

The bias shows up in how an agent judges the consequences of a query or plan it already produced. IBIS isolates that judgment by holding the search observations constant, so the accuracy gap traces only to authorship of that earlier step.

NIS-Agent, described in the same paper, walls off the context that produced a step from the context that judges it. It does this at two points: triaging which webpages to trust, and validating the agent's own final answer. Splitting judgment from authorship at just those two spots holds accuracy on GAIA, WebWalkerQA and BrowseComp. Token use drops by a third.