The paper identifies a problem called in Direct Corpus Interaction (DCI), where reachable evidence remains unusable under finite interaction budgets. To address this, the authors propose AtlasNav, a framework that pre-organizes the corpus into a Corpus Atlas with four views (Topic, Identity, Episode, Relation) and uses lightweight query-adaptive routing to navigate efficiently without reconstructing the search space per query.
On BrowseComp-Plus, AtlasNav achieves 92.05% strict accuracy while reducing recorded online inference cost by 30.21% relative to prior dynamic-workspace methods. Under matched budgets, it realizes the complete required evidence earlier and approaches the evidence-supplied reference faster. The approach also transfers effectively to PhantomWiki (with controlled scaling from 10K to 1M documents) and heterogeneous enterprise knowledge.
The results demonstrate that agentic search performance depends not only on what evidence is accessible, but on how the corpus is represented. Reusable offline organization reduces online burden, mitigates evidence blindness, and shifts focus from rebuilding search spaces per query to navigating a shared persistent structure.