What they did

The authors model direct corpus interaction as finite-budget navigation and measure evidence realization in stages: whether required evidence surfaces, whether the supporting document is opened, and whether its decisive fragment is exposed.

They built AtlasNav, which preprocesses a corpus into a persistent, multi-view Corpus Atlas that can be reused across queries. They evaluated it on BrowseComp-Plus, on PhantomWiki across controlled corpus sizes from 10,000 to 1 million items, and on heterogeneous enterprise knowledge.

Key findings

• AtlasNav achieved 92.05% strict accuracy on BrowseComp-Plus.
• It reduced recorded online inference cost by 30.21% relative to the prior dynamic-workspace state of the art.
• Under matched interaction budgets, AtlasNav assembled the complete required evidence earlier and approached the same model’s evidence-supplied empirical reference more quickly.
• The persistent-representation approach remained effective across PhantomWiki’s different corpus structure and 10K–1M scaling range, and transferred competitively to enterprise corpora.

Why it matters

The results suggest that giving an agent access to a corpus is not enough: corpus organization determines whether relevant evidence becomes usable before the interaction budget runs out. Reusable navigation structure may therefore improve agentic search while avoiding the repeated cost of rebuilding a query-specific workspace.

Caveats

The reported gains center on the evaluated benchmarks, model configuration, budgets, and cost accounting; the abstract does not establish how AtlasNav behaves under continuously changing corpora or what offline construction and maintenance costs the Corpus Atlas incurs. Evidence-supplied performance is an empirical reference rather than proof that all remaining errors arise from navigation.