Researchers have introduced AtlasNav, a persistent multi-view corpus-navigation framework designed to address "Evidence Blindness" in large language model agents interacting with external corpora. This condition occurs when required evidence fails to surface or remain usable within finite interaction budgets.
- AtlasNav organizes the corpus once into a Corpus Atlas, allowing queries to navigate adaptively rather than reconstructing shared structure for each query.
- On BrowseComp-Plus, it achieves 92.05% strict accuracy while reducing recorded online inference cost by 30.21% relative to prior dynamic-workspace methods.
- Under matched budgets, the framework realizes complete required evidence earlier and approaches the model's evidence-supplied empirical reference more rapidly.
- The approach remains effective under PhantomWiki's corpus organization across 10K-1M scaling and transfers competitively to heterogeneous enterprise knowledge.
The results demonstrate that agentic search effectiveness depends not only on accessible evidence but also on how the corpus is represented to ensure limited interaction becomes effective navigation.