PaperlayerView exact revision on arXiv

Source-linked AI summary

Evidence Blindness in Direct Corpus Interaction: Persistent Navigation with AtlasNav

Hongyu Guo, Zhiyu Zheng, Zhao Cao

arXiv:2608.24764v1cs.AI

TL;DR

Agentic search can fail even when required evidence exists, because finite interaction budgets leave evidence undiscovered, unopened, or unlocalized. The paper introduces AtlasNav, which organizes a corpus once into a persistent multi-view atlas for adaptive navigation; on BrowseComp-Plus, it reaches 92.05% strict accuracy with 30.21% lower recorded online cost than the prior dynamic-workspace state of the art.

  • Problem

    Evidence Blindness captures how required evidence can remain unusable under finite interaction despite being present and reachable.

  • Method

    AtlasNav organizes the corpus once into a persistent multi-view Corpus Atlas and combines query-adaptive routing with direct corpus interaction.

  • Results

    92.05% strict accuracy is achieved on BrowseComp-Plus with 30.21% lower recorded online cost than the prior dynamic-workspace state of the art.

  • Takeaways & Limitations

    Corpus representation is a first-class component of the agent interface, with reusable structure turning limited interaction from repeated exploration into effective navigation.

  • Takeaways & Limitations

    The router excludes BrowseComp-Plus evaluation questions and labels from its corpus-derived training tasks.

Abstract

from arXiv · show

Large language model agents are moving beyond conventional retrieval-augmented generation toward direct interaction with external corpora. Direct Corpus Interaction (DCI) keeps the full corpus accessible, yet reachable evidence can remain unusable under finite interaction budgets. Required evidence may fail to surface, a surfaced supporting document may remain unopened, or an opened document may fail to expose its decisive fragment. We call this progressive silent loss Evidence Blindness and quantify it through stage-wise evidence realization. Within the DCI paradigm, raw interaction adds little reusable corpus organization, while dynamic-workspace methods reconstruct a query-conditioned interaction space from each query and trajectory. In both cases, useful structure is recovered largely online. We instead formulate large-scale agentic search as finite-budget navigation over reusable corpus structure. We introduce AtlasNav, a persistent multi-view corpus-navigation framework that retains direct corpus interaction but organizes the corpus once into a Corpus Atlas, allowing each query to navigate adaptively rather than reconstruct shared structure. On BrowseComp-Plus, AtlasNav achieves 92.05% strict accuracy while reducing recorded online inference cost by 30.21% relative to the prior dynamic-workspace state of the art. Under matched budgets, it realizes the complete required evidence earlier and approaches the same model's evidence-supplied empirical reference more rapidly. The same representation principle remains effective under PhantomWiki's distinct corpus organization and controlled 10K-1M scaling, and transfers competitively to heterogeneous enterprise knowledge. These results show that agentic search depends not only on accessible evidence, but also on how the corpus is represented so that limited interaction becomes effective navigation.

1 INTRODUCTION

The paper reframes agentic search as finite-budget navigation over reusable corpus structure. AtlasNav retains direct corpus access while organizing the corpus once into a persistent, query-adaptable Corpus Atlas, improving accuracy and efficiency.

  • Direct Corpus Interaction lets agents search, read, and verify evidence over the full corpus instead of reasoning over a compressed Top-k context.
  • Evidence Blindness occurs when required evidence is not surfaced, opened, or localized despite remaining reachable under a finite interaction budget.The paper formalizes these stages as Construction, Surface, Open, and Locate.
  • AtlasNav learns a persistent multi-view Corpus Atlas once, then routes each query adaptively across Topic, Identity, Episode, and Relation views.Canonical documents remain directly accessible to the unchanged agent.
  • Under matched budgets, AtlasNav realizes complete required evidence earlier and remains effective across corpus organization, scale, and source heterogeneity.
  • AtlasNav formulates large-scale agentic search as finite-budget navigation over a shared representation rather than repeated query-time reconstruction.
  • 92.05% strict accuracy is achieved on BrowseComp-Plus, 7.47 points above the prior dynamic-workspace state of the art, with 30.21% lower recorded online cost.

2 RELATED WORK

Prior work progresses from bounded retrieval toward direct corpus interaction, dynamic workspaces, and reusable structured representations. Evidence-access diagnostics complement final-answer accuracy by measuring whether and how completely required evidence becomes usable.

  • Agentic corpus interfaces: Classical RAG retrieves bounded context before generation, whereas agentic systems interleave retrieval with reasoning and query reformulation.
  • Agentic corpus interfaces: Direct Corpus Interaction exposes the raw corpus, while DR-DCI dynamically reconstructs a local workspace and RISE retrieves a query-specific bounded shell workspace.
  • Structured representations: Hierarchical, graph, and multi-relational methods replace flat retrieval with reusable structure for corpus-level and multi-hop retrieval.
  • Evidence-access diagnostics: RAGAS, ARES, and RAGChecker evaluate retrieval and generation beyond final-answer accuracy, while DCI adds trajectory-level Coverage and Localization.
  • Evidence-access diagnostics: Evidence Blindness extends evidence-access diagnosis through Construction, Surface, Open, and Locate, with fragment-level Qrel making Locate directly measurable.

3 EVIDENCE BLINDNESS

The paper shows that evidence can be present and reachable yet remain unusable during interaction. It formalizes this progressive loss with stage-wise evidence realization and a fragment-level measure of decisive evidence.

  • 96.51% strict accuracy with supplied gold documents versus 84.58% for the prior dynamic-workspace state of the art shows that interaction can fail to make present evidence usable.
  • Complete required evidence is surfaced for 85.90% of questions but decisive fragments are completely located for only 75.54%.
  • Evidence Blindness describes progressive silent loss when required evidence remains undiscovered, unopened, or unlocalized while the agent continues reasoning.
  • Formalization: The analysis tracks required evidence slots rather than merely whether a supporting document is reached.
  • Formalization: Evidence realization proceeds from Surface to Open to Locate: document entries appear, canonical bodies enter context, and decisive fragments become available.
  • Formalization: Construction is saturated in the main comparisons because all annotated supporting evidence is present in the canonical corpus and remains reachable.
  • Measurement: Locate is measured with a fragment-level Qrel that separates reaching a supporting document from reaching its decisive evidence.

4 ATLASNAV: PERSISTENT CORPUS NAVIGATION

AtlasNav treats agentic search as finite-budget navigation over a persistent corpus representation, preserving direct access while organizing reusable evidence directions once. Query-adaptive routing then prioritizes complementary atlas views so agents can expose and verify required evidence more efficiently.

  • Persistent corpus navigation: AtlasNav retains the DCI search–read–verify loop while changing the corpus-side representation to a persistent multi-view Corpus Atlas.The full corpus remains directly accessible rather than being replaced by a query-specific candidate pool.
  • Evaluation: The evaluation fixes the agent and budget protocol across interfaces and measures performance across budgets to assess earlier evidence realization and interaction efficiency.The empirical reference uses the same agent with benchmark-designated gold documents supplied directly, while the reference gap captures unrealized evidence-conditioned capability.
  • Evidence realization: Evidence Blindness diagnoses failures at successive stages: required evidence may not surface, supporting documents may remain unopened, or decisive fragments may remain undiscovered.The diagnosis distinguishes evidence reachability from evidence usability under a finite interaction budget.
  • Persistent corpus navigation: The Corpus Atlas organizes canonical documents through Topic, Identity, Episode, and Relation views, whose neighborhoods support a persistent navigable hierarchy.Topic and Identity define coarse regions, while Episode and Relation distinguish event stages and relational patterns.
  • Query-adaptive navigation: Each query changes the priority of reusable evidence directions rather than rebuilding the search space online.AtlasNav combines the four semantic channels with BM25, allowing query-suited and complementary signals to influence navigation.
  • Query-adaptive navigation: Document priorities are projected onto complementary Atlas regions, giving limited observations broader coverage across plausible parts of a multi-hop evidence chain.The fixed agent then searches, reads, compares, and verifies canonical evidence from those initial entries.

5 EXPERIMENTS

Experiments evaluate AtlasNav across evidence realization, efficiency, corpus organization, scale, and enterprise transfer. Across these settings, persistent multi-view navigation improves accuracy–efficiency trade-offs and reduces Evidence Blindness, enabling earlier use of complete evidence.

  • Experimental setup: AtlasNav is compared with raw DCI and DR-DCI across BrowseComp-Plus, PhantomWiki, and EnterpriseRAG-Bench, using strict accuracy, cost, Evidence Blindness, reference gaps, and official enterprise metrics.PhantomWiki isolates corpus-growth effects with nested corpora, while EnterpriseRAG-Bench tests heterogeneous enterprise transfer.
  • BrowseComp-Plus: accuracy–efficiency: AtlasNav improves strict accuracy by 3.98–21.57 percentage points over DR-DCI while reducing recorded online cost by 0.62–30.21% across four BrowseComp-Plus backbones.The result supports a backbone-robust interface improvement rather than a model-specific optimization.
  • BrowseComp-Plus: Evidence Blindness: AtlasNav reduces complete-evidence blindness on DeepSeek from 14.10% to 4.94% and on Qwen from 49.76% to 20.12%.The improvement first appears at Surface and carries through Open and Locate, indicating that exposed evidence becomes usable evidence.
  • BrowseComp-Plus: finite-budget convergence: At the lowest DeepSeek matched-cost budget, AtlasNav reduces EBL_All to 35.90% versus 45.18% for DR-DCI, with reference gaps of 25.55 and 46.15 points.On ChatGPT at 60 turns, EBL_All is 13.61% versus 20.72%, with gaps of 7.47 versus 15.55 points.
  • PhantomWiki: structure and scale: On PhantomWiki, AtlasNav has the lowest EBS at every scale and the highest strict accuracy, including 76.0–77.5% from 10K to 100K files.At roughly one million files, AtlasNav retains the lowest Surface blindness and highest observed strict accuracy despite all interfaces degrading.
  • EnterpriseRAG-Bench: transfer: AtlasNav reaches 73.72 Overall on EnterpriseRAG-Bench, with 79.80% Correctness, 78.53% Completeness, 66.98% Document Recall, and 0.66 Invalid Extra Documents.Performance varies by question type, with lower Overall scores on project-related and completeness-heavy questions.

6 CONCLUSION

The paper frames agentic search as finite-budget navigation over reusable corpus structure rather than merely access to existing evidence. AtlasNav retains direct interaction while using a persistent multi-view Corpus Atlas, and the reported results support this principle across corpus organization, scale, and enterprise knowledge.

  • Conclusion: Evidence Blindness captures whether required evidence becomes visible, accessible, and usable under finite interaction, not merely whether it exists.AtlasNav reduces this silent loss and closes the empirical reference gap faster on BrowseComp-Plus.
  • Conclusion: AtlasNav extends Direct Corpus Interaction with a persistent multi-view Corpus Atlas while retaining direct corpus access.The corpus is organized once so limited interaction can become effective navigation rather than repeated exploration.
  • Conclusion: Persistent navigation remains effective across changes in corpus organization, scale, and heterogeneous enterprise knowledge.The conclusion treats corpus representation as part of the agent interface within the evaluated settings.

A LIMITATIONS

The paper’s evaluation separates evidence realization from answer accuracy and uses a frozen, benchmark-aligned target. AtlasNav’s benefits are bounded by offline-construction amortization and by downstream interaction bottlenecks.

  • Limitations: AtlasNav requires reusable offline construction, whose per-query cost depends on shared-query volume, reuse duration, and refresh frequency.Construction resources are therefore reported separately from within-backbone online inference comparisons.
  • Limitations: Persistent navigation improves finite-budget corpus navigation but does not eliminate downstream reasoning or document-consumption constraints.At million-file scale, converting surfaced directions into opened documents becomes an additional bottleneck.
  • Evaluation scope: The evidence stages distinguish availability, document visibility, document opening, and decisive-fragment realization during model-visible interaction.Filenames, previews, failed tool outputs, and internal search state do not count as Open or Locate.
  • Evaluation scope: Evidence Blindness is measured against a frozen benchmark-aligned fragment-level Qrel, while accuracy remains an independent task-success measure.The Qrel is deterministic, auditable, and shared across interfaces; Locate uses exact normalized substring matching without an embedding or LLM judge.
  • AtlasNav design: AtlasNav’s corpus-side method combines an offline multi-view Atlas, a corpus-trained query-adaptive router, and the original DCI search–read–verify loop.The Atlas organizes rather than prunes the corpus, using Topic, Identity, Episode, and Relation views.
  • AtlasNav design: On 1,024 parent-disjoint routing tasks, learned routing improves All-positive Recall@30 from 0.6465 to 0.6680 and triple Recall@30 from 0.3647 to 0.4235 over the uniform safe baseline.It also improves pair Recall@30 from 0.5955 to 0.6147 and rare-task Recall@30 from 0.6378 to 0.6786.

D.1 BENCHMARKS AND INTERFACES

The experiments evaluate AtlasNav against raw and dynamic-workspace DCI interfaces across benchmark, scaling, auxiliary, and enterprise settings. Protocols keep comparisons controlled and report online inference cost separately from reusable construction.

  • Benchmarks: AtlasNav is evaluated on BrowseComp-Plus, PhantomWiki, and EnterpriseRAG-Bench, alongside auxiliary settings probing distinct reasoning, aggregation, and ranking regimes.BrowseComp-Plus contains 830 questions over 100,195 documents; PhantomWiki fixes 200 questions while scaling from 10,059 to 1,006,839 documents.
  • Interfaces: The main comparison includes raw DCI, query-conditioned DR-DCI, and AtlasNav, with AtlasNav retaining full-corpus access while adding persistent organization and query-adaptive routing.The language-model backbone, tool schema, answer judge, and nominal interaction ceiling are held constant within controlled comparisons.
  • Evaluation protocol: Every question contributes one adopted trajectory, and malformed, failed, empty, or missing legal answers remain in the benchmark denominator as incorrect.The protocol excludes best-of-N selection, correctness-conditioned replacement, and error-targeted reruns.
  • Evaluation protocol: The evidence-supplied empirical reference gives the same agent benchmark-designated gold documents without the reference answer, bypassing much evidence discovery while preserving answer formation.It is used as an empirical reference rather than a deployable system or theoretical upper bound.
  • Evaluation protocol: Turn checkpoints measure naturally available evidence at fixed interaction depths, while cost checkpoints compare the same process under fixed within-backbone inference budgets.Evidence or answers appearing after a checkpoint are not backfilled into earlier budgets.
  • Evaluation protocol: Recorded online cost includes adopted trajectories, answer judges, and recorded online query embeddings, while reusable construction is reported separately.Costs are compared only within the same backbone and currency because deployment-dependent amortization varies.

E.1 BROWSECOMP-PLUS: EVIDENCE DYNAMICS ACROSS BACKBONES

Across backbones and corpus scales, AtlasNav makes complete evidence available earlier, while its scaling behavior reveals a shift from surface discovery to document opening as the dominant bottleneck. Ablations and diversity analyses support an early-navigation advantage rather than uniformly greater global diversity.

  • Evidence dynamics: Across all four backbones, AtlasNav reaches lower Locate blindness earlier than alternatives and preserves the advantage throughout trajectories.The separation is largest for Qwen and smallest for ChatGPT, while the direction is consistent across capability levels.
  • Evidence dynamics: On DeepSeek, AtlasNav reduces Locate blindness from 38.19% at 15 turns to 22.77% at 30 turns, versus 50.48% and 39.64% for DR-DCI.On ChatGPT at 60 turns, AtlasNav reaches 13.61% Locate blindness versus 20.72% for DR-DCI.
  • Evidence dynamics: Within-backbone cost checkpoints preserve the same ordering, indicating that persistent navigation makes complete evidence usable earlier rather than merely extending trajectories.Cost budgets are not comparable across backbones because provider prices and currencies differ.
  • PhantomWiki scaling: At 1M PhantomWiki files, AtlasNav has 49.5% Surface blindness versus 59.0% for DCI and 87.0% for DR-DCI, but 71.0% Open blindness.Its complete Surface-to-Open conversion is 57.4%, showing that the largest-scale bottleneck follows surface exposure.
  • PhantomWiki scaling: PhantomWiki scaling analysis is restricted to Surface and Open because it lacks fragment-level annotations comparable to the BrowseComp-Plus Qrel.Answer-string matching is not treated as the same Locate criterion.
  • Enterprise transfer: EnterpriseRAG-Bench scores vary more by question type than source tag, with overall source-tag scores ranging from 64.76 on Confluence to 80.88 on HubSpot.Source tags overlap and have unequal sample sizes, so these are descriptive slices rather than independent test sets.
  • Hierarchical view assignment: On the 166-question hierarchy ablation, Topic+Identity parents with Episode+Relation children reach 95.78% accuracy at the lowest recorded cost and fewest turns.The selected hierarchy answers 159/166 questions versus 147/166 for the role reversal, costing CNY 22.02 less.
  • Early region diversity: Within the first ten surfaced files, AtlasNav covers 9.00 parent and 9.99 leaf regions on average, versus 7.77/8.64 for DCI and 3.95/5.50 for DR-DCI.This advantage remains at 20 files, but at 50 files DCI slightly exceeds AtlasNav in leaf diversity, 33.84 versus 31.31.

F.3 WHEN MULTIPLE VIEWS MATTER

The benefit of persistent multi-view navigation varies with task structure: exploratory evidence suggests larger gains when views disagree, while other regimes reveal boundaries on accuracy, aggregation, and cost. Across graded retrieval, AtlasNav improves ranking quality but not universally online efficiency.

  • When multiple views matter: AtlasNav’s relative accuracy gain rises from 0.71 to 12.86 points over DR-DCI and from 1.43 to 12.86 points over DCI across disagreement quartiles.Higher disagreement means Topic, Identity, Episode, and Relation induce less consistent evidence geometry.
  • When multiple views matter: 12.14 points relative to DR-DCI and 11.43 points relative to DCI separate the highest- from lowest-disagreement quartiles.The associations remain significant after correction, but the analysis is exploratory and does not establish generic dispersion as a predictor.
  • Cross-regime boundaries: 94.5% endpoint accuracy matches DR-DCI but trails DCI’s 95.5% on 2Wiki-Global, where AtlasNav instead reduces cost to CNY 11.03.AtlasNav also uses 3,135 turns, versus 4,117 for DCI and 5,540 for DR-DCI.
  • Cross-regime boundaries: Persistent navigation reduces search effort in FanOutQA, using 35% fewer turns than DR-DCI, but DR-DCI achieves the highest Loose and Strict scores.AtlasNav’s Strict differences are not significant against either DCI or DR-DCI.
  • Cross-regime boundaries: Persistent multi-view organization improves graded retrieval quality, but additional navigation does not universally reduce inference cost.Its benefit depends on whether evidence quality, interaction depth, or monetary cost is the dominant objective.

H QUALITATIVE ANALYSIS

Qualitative trajectories distinguish evidence realization from final-answer correctness and motivate reporting Evidence Blindness alongside accuracy. The appendix provides numerical definitions, frozen-checkpoint conventions, and scaling-related measurement boundaries.

  • Qualitative analysis: Evidence Blindness and final-answer accuracy are related but not interchangeable measurements.Evidence can be fully realized without correct synthesis, while a correct answer can occasionally arise from partial annotated evidence.
  • Qualitative analysis: Question 8 separates access failure from reasoning failure, while Question 417 shows that complete annotated evidence does not ensure correct synthesis.Question 394 shows that support-document coverage can exceed what is needed to realize decisive evidence slots.
  • Qualitative analysis: EB localizes failures in benchmark-annotated evidence realization but is not intended to identify the causal source of every answer.It should therefore be reported alongside, rather than instead of, task accuracy.
  • Measurement conventions: Any, Mean, and All summarize complete absence, average missing evidence mass, and failure to realize the complete evidence set.The small Any–All gaps in several systems indicate that required slots are often realized or missed together.
  • Measurement conventions: Frozen trajectory prefixes prevent evidence appearing after a checkpoint from being backfilled into an earlier budget.Turn and cost checkpoint tables report stage-wise Locate-All values under these conventions.
  • Measurement conventions: PhantomWiki scaling fixes questions and supporting evidence across nested corpus sizes while reporting Surface, Open, and conditional Surface-to-Open measures.The scaling table covers 10K–1M-style nested corpus conditions through its stated stage-wise metrics.

I.4 ENTERPRISERAG-BENCH METRIC MATRICES

EnterpriseRAG-Bench reports AtlasNav metrics by question category and overlapping document-source tag. Because source tags overlap and sample sizes differ, source-level rows are descriptive rather than independent test sets.

  • Metric matrices: AtlasNav’s official EnterpriseRAG-Bench metrics are organized by question category and overlapping document-source tag.The category and source matrices provide the exact values summarized in the transfer analysis.
  • Metric matrices: Source-level rows are descriptive rather than independent test sets because source tags overlap and have unequal sample sizes.Questions may carry multiple tags, so total tag counts exceed 500.
  • Metric matrices: Reusable construction workloads are reported as construction resources rather than deployment-independent per-query costs.The corpus representation is built once and reused across queries.
Loading 2608.24764v1…