Evidence Blindness in Direct Corpus Interaction: Persistent Navigation with AtlasNav
AtlasNav is pitched as a fix for evidence that agents can reach but still fail to use.
The paper names that failure “Evidence Blindness”: evidence can be missed, left unopened, or opened without exposing the key fragment. AtlasNav organizes a corpus once into a reusable “Corpus Atlas,” so each query navigates existing structure instead of rebuilding a workspace from scratch. On BrowseComp-Plus, the authors report 92.05% strict accuracy and a 30.21% reduction in recorded online inference cost versus the prior dynamic-workspace state of the art. ArXiv · AI/CL/LG's note
The paper names that failure “Evidence Blindness”: evidence can be missed, left unopened, or opened without exposing the key fragment. AtlasNav organizes a corpus once into a reusable “Corpus Atlas,” so each query navigates existing structure instead of rebuilding a workspace from scratch. On BrowseComp-Plus, the authors report 92.05% strict accuracy and a 30.21% reduction in recorded online inference cost versus the prior dynamic-workspace state of the art. ArXiv · AI/CL/LG's note
score 4