Agents
NIS-Agent isolates context at two decision points to fix agents grading their own prior search steps
The paper names inertia bias: once a web research agent has produced a query, plan or intermediate conclusion, it judges the consequences of that same action less objectively. The IBIS benchmark isolates the effect by holding search observations fixed while varying whether the model owns the preceding step, and models are substantially worse when they do. NIS-Agent applies context isolation at webpage triage and final-answer validation, staying competitive on GAIA, WebWalkerQA, BrowseComp and BrowseComp-zh while cutting token cost 33% against the authors' baseline.
Source
↳ Follow the thread