AtlasNav: Mitigating Evidence Blindness with Persistent Corpus Navigation

Hongyu Guo; Zhiyu Zheng; Minghao Hu; Shuai Lei; Zhunchen Luo; Chao Ma; Yunbo Cao; Zhao Cao

cs.AIAug 25, 2026

AtlasNav: Mitigating Evidence Blindness with Persistent Corpus Navigation

Authors: Hongyu Guo, Zhiyu Zheng, Minghao Hu, Shuai Lei, Zhunchen Luo, Chao Ma, Yunbo Cao, Zhao Cao

Abstract

As language-model agents become more capable of iterative search, corpus access is shifting from retrieval toward interaction. Agents can explore the corpus, inspect documents, and use newly discovered evidence to decide what to examine next. Yet accessible evidence may still fail to become usable within a finite interaction budget. We call this progressive failure Evidence Blindness: supporting documents may never enter view, may remain unopened, or may fail to expose the decisive evidence even after being opened. A key reason is that agents often have to infer useful evidence directions during interaction, spending limited budget on deciding where to search next. Existing approaches either leave corpus structure largely implicit or reconstruct useful directions at query time. AtlasNav instead organizes reusable cross-document structure before any query arrives. It builds a persistent multi-view Corpus Atlas, which each query can navigate adaptively while still accessing the original documents directly. On BrowseComp-Plus, AtlasNav outperforms the previous state-of-the-art interactive corpus access method across different backbones. On DeepSeek, it improves strict accuracy by 7.47 points while reducing query-time inference cost by 30.22%.AtlasNav also reduces Evidence Blindness, realizes complete evidence earlier, remains robust to corpus-structure and scale shifts on PhantomWiki, and achieves leading performance on heterogeneous enterprise data.

Figures & tables

Figure 1: Evidence realization on BrowseComp-Plus with DeepSeek-V4-Flash. (a) Complete evidence realized per checkpoint. (b) Strict accuracy given complete vs. incomplete realization.

Figure 2: AtlasNav and Evidence Blindness. (a) Offline construction of the persistent multi-view Corpus Atlas. (b) Query-adaptive finite-budget navigation. (c) Evidence Blindness diagnosis across the four evidence-realization checkpoints.

	Strict Accuracy (%) $\uparrow$					Normalized Query-Time Agent Cost $\downarrow$
Backbone	DCI	DR-DCI	AtlasNav	Ref.	$\Delta$ Acc.	DCI	DR-DCI	AtlasNav	Cost save (%)
GPT-5.6	87.35	87.83	91.81	96.51	$+3.98$	1.718	1.173	1.000	14.74
DeepSeek	81.45	84.58	92.05	96.51	$+7.47$	2.347	1.433	1.000	30.22
Qwen	36.99	50.96	72.53	95.42	$+21.57$	2.243	1.271	1.000	21.31
MiMo	62.89	71.08	79.04	96.14	$+7.96$	2.579	1.006	1.000	0.62

Table 1: Accuracy and recorded query-time agent inference cost on BrowseComp-Plus. Costs are normalized within each backbone to AtlasNav

=1.000

; accuracy gains and cost savings are relative to DR-DCI.

		Surface $\downarrow$		Open $\downarrow$		Locate $\downarrow$
Backbone	Interface	$\mathrm{EB}_{\mathrm{Any}}^{S}$	$\mathrm{EB}_{\mathrm{All}}^{S}$	$\mathrm{EB}_{\mathrm{Any}}^{O}$	$\mathrm{EB}_{\mathrm{All}}^{O}$	$\mathrm{EB}_{\mathrm{Any}}^{L}$	$\mathrm{EB}_{\mathrm{All}}^{L}$
GPT-5.6	DCI	6.63	6.99	12.89	13.01	15.90	16.27
	DR-DCI	9.40	9.76	10.48	11.08	15.90	16.63
	AtlasNav	5.66	6.02	9.04	9.52	11.93	12.65
DeepSeek	DCI	13.13	13.37	14.70	15.42	17.83	18.80
	DR-DCI	13.49	14.10	16.51	17.11	23.61	24.46

Table 2: Evidence Blindness (EB) on BrowseComp-Plus (%); lower is better.

Figure 3: Finite-budget empirical reference gaps

\Delta_{I}(B)

under turn and query-time inference-cost budgets for GPT-5.6 and DeepSeek; lower is better. Results for Qwen and MiMo are provided in Appendix B .

System	Overall $\uparrow$	Correct. (%) $\uparrow$	Complete. (%) $\uparrow$	Doc. Recall (%) $\uparrow$	Invalid Extra $\downarrow$
AtlasNav	87.00	94.40	89.14	87.05	0.66
Mixedbread (+ Opus 5)	86.58	89.80	90.62	94.28	0.39
Prism (aurait.ai)	80.46	85.60	86.26	87.81	3.31
metor.com	80.34	82.00	86.22	85.53	4.96
ZNV AgentCube	80.26	82.40	86.56	86.10	2.94
Skyller AI	79.30	81.60	87.39	86.50	8.74

Table 4: EnterpriseRAG-Bench metrics with representative public systems.

Appendix figures & tables20 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Interface	Surface	Open	Locate
GPT-5.6-Luna	DCI	6.81	12.93	16.02
	DR-DCI	9.58	10.78	16.27
	AtlasNav	5.84	9.28	12.29
DeepSeek	DCI	13.25	15.04	18.29
	DR-DCI	13.80	16.81	24.04
	AtlasNav	4.90	7.73	11.08

Appendix

Table 5: Mean endpoint Evidence Blindness on BrowseComp-Plus (%); lower is better. Any-slot and All-slot results are reported in the main text. Construction is omitted because it is saturated for all evaluated interfaces.

		Strict Accuracy (%) $\uparrow$			$\operatorname{EB}_{\mathrm{All}}^{S/O/L}$ (%) $\downarrow$
Backbone	Cost	DCI	DR-DCI	AtlasNav	DCI	DR-DCI	AtlasNav
GPT-5.6-Luna	0.025	45.66	65.42	66.39	24.70/36.14/40.60	17.47/ 20.12 /29.88	16.51 /23.61/ 26.75
	0.05	57.83	74.82	75.90	20.24/30.00/34.46	14.10/ 16.02 /23.73	13.37 /18.67/ 21.57
	0.15	69.52	81.20	82.05	13.49/22.77/26.51	12.77/ 14.46 /20.84	10.12 /15.30/ 18.07
	0.65	84.46	86.39	91.81	7.95/14.58/18.07	10.12/11.57/17.47	6.02 / 9.52 / 12.65
DeepSeek	0.15	49.16	50.36	70.96	43.73/49.88/54.46	29.40/36.02/45.18	23.25 / 30.36 / 35.90

Appendix

Table 7: BrowseComp-Plus results at within-backbone query-time agent-cost checkpoints. Each row is one cost budget for one backbone; the first block reports strict accuracy (%) and the second checkpoint-wise

\operatorname{EB}_{\mathrm{All}}^{S/O/L}

(%), with each EB entry as Surface / Open / Locate. Lower is better for EB.

Scale	Files	Interface	$\operatorname{EB}_{\mathrm{All}}^{O}$ $\downarrow$	Surface-to-Open (%) $\uparrow$
10K	10,059	DCI	38.0	93.94
		DR-DCI	76.5	95.92
		AtlasNav	45.5	80.15
50K	50,324	DCI	52.0	84.21
		DR-DCI	76.5	79.66
		AtlasNav	52.5	77.24

Appendix

Table 8: PhantomWiki evidence realization beyond the main text. Questions and supporting evidence are fixed across nested corpus sizes. Surface-to-Open is the fraction of complete-Surface questions that also reach complete Open.

Category	$n$	Overall $\uparrow$	Correct. $\uparrow$	Complete. $\uparrow$	Doc. Recall $\uparrow$	Invalid Extra $\downarrow$
Basic	175	95.16	98.86	95.16	93.71	0.26
Semantic	125	80.73	89.60	82.25	80.00	0.80
Intra-document reasoning	40	96.46	100.00	96.46	97.50	0.05
Project related	40	59.09	87.50	67.88	65.48	2.77
Constrained	30	94.57	96.67	96.79	95.00	0.70
Conflicting information	20	92.69	100.00	92.69	80.00	0.20

Appendix

Table 9: Complete AtlasNav metrics by EnterpriseRAG-Bench question category.

Source tag	$n$	Overall $\uparrow$	Correct. $\uparrow$	Complete. $\uparrow$	Doc. Recall $\uparrow$	Invalid Extra $\downarrow$
Confluence	115	78.00	91.30	81.68	80.92	1.37
Fireflies	24	87.22	95.83	87.22	74.58	0.38
GitHub	60	82.80	93.33	86.07	82.67	0.98
Gmail	55	86.74	94.55	89.58	86.00	0.45
Google Drive	59	88.70	96.61	89.53	83.92	1.05
HubSpot	33	92.42	96.97	92.93	96.97	0.03

Appendix

Table 10: Complete AtlasNav metrics by overlapping EnterpriseRAG-Bench source tag. Counts do not sum to the benchmark size because questions may have multiple source tags.

Item	Count
Questions	830
Required evidence slots	889
Accepted canonical fragments	1,753
Canonical documents represented	1,445
Questions passing automatic validation	828
Questions requiring manual adjudication	2

Appendix

Table 11: Frozen BrowseComp-Plus fragment-level Qrel used to operationalize Answer-Evidence Locate.

Component	Setting
View signatures	Topic (6,144 characters); Identity, Episode, Relation (4,096 characters each)
Encoding	2,560-dimensional unit vectors; independent 192-dimensional PCA projections fitted on 50,000 deterministic samples
Sparse graphs	cosine/IP HNSW; $k=48$ , $M=32$ , efConstruction $=160$ , efSearch $=128$
Coarse hierarchy	multiplex Topic (1.0) + Identity (0.75); 77 parent regions
Fine hierarchy	conditional Episode + Relation within each parent; 443 leaves
Addressing	100,195/100,195 documents receive one parent–leaf address; canonical files remain directly searchable and openable

Appendix

Table 12: Frozen Atlas construction configuration on BrowseComp-Plus.

Item	Setting
Tasks	1,913 single; 4,300 pair; 950 triple; 7,163 total
Parent-disjoint split	4,972 train; 1,167 validation; 1,024 test
Input/model	816 features; 4,902-parameter linear three-head router
Output	Topic, Identity, Episode, Relation, and BM25 channel weights
Objective	all-positive ranking + worst-positive bottleneck and calibration losses
Safety calibration	0.05 floor on each view’s share of semantic mass; low confidence shrinks toward uniform semantic weights and a 1:1 semantic/BM25 mass

Appendix

Table 13: Corpus-derived router data and frozen model configuration.

Figure 5: BrowseComp-Plus Locate Evidence Blindness

\operatorname{EB}_{\mathrm{All}}^{L}

over interaction turns for DCI, DR-DCI, and AtlasNav; lower is better. Exact checkpoint values are reported in Appendix B .

Backbone	AtlasNav Acc.	DR-DCI Acc.	Difference	Paired bootstrap 95% CI	Exact McNemar
DeepSeek	92.05%	84.58%	$+7.47$ pp	$[4.94,10.00]$ pp	$p=1.22\times 10^{-8}$
GPT-5.6-Luna	91.81%	87.83%	$+3.98$ pp	$[1.69,6.39]$ pp	$p=0.00119$

Appendix

Table 14: Paired uncertainty for the AtlasNav–DR-DCI endpoint comparison on the two backbones with retained per-question records. Confidence intervals are fixed-seed paired bootstrap intervals;

p

-values are from exact McNemar tests.

Figure 6: Finite-budget empirical reference gaps

\Delta_{I}(B)

for Qwen and MiMo under turn and query-time inference-cost budgets. These panels complement Figure 3 in the main text; lower is better.

Parent views	Child views	Accuracy $\uparrow$	Mean turns $\downarrow$
Topic + Identity	Episode + Relation	95.78	27.43
Topic + Episode	Identity + Relation	89.16	37.33
Topic + Relation	Identity + Episode	91.57	35.10
Identity + Episode	Topic + Relation	91.57	35.88
Identity + Relation	Topic + Episode	89.76	31.89
Episode + Relation	Topic + Identity	88.55	36.02

Appendix

Table 15: Hierarchical view-assignment ablation on a shared 166-question subset. All variants use the same corpus, view graphs, router, agent, and observation budget and differ only in parent–child view assignment.

Interface	P@10	L@10	P@20	L@20	P@50	L@50
DCI	7.77	8.64	12.96	15.77	22.83	33.84
DR-DCI	3.95	5.50	5.97	8.85	9.89	15.80
AtlasNav	9.00	9.99	16.54	18.64	23.53	31.31

Appendix

Table 16: Mean number of distinct Atlas parent (P) and leaf (L) regions under matched surfaced-file budgets on BrowseComp-Plus.

Case	Evidence state	Outcome	Diagnostic
8 ( Lush Life )	AtlasNav and DCI realize the complete annotated evidence; DR-DCI does not realize the annotated chain.	AtlasNav correct; DCI and DR-DCI wrong.	Evidence access and downstream reasoning are distinct failure modes.
394 (wing three-quarter)	AtlasNav reaches only part of the annotated supporting-document set but realizes the complete decision-relevant slot-level evidence; the baselines do not.	AtlasNav correct; baselines wrong.	Complete support-document coverage is not required once the decisive evidence slots are realized.
417 (Owlman)	AtlasNav realizes the complete annotated evidence; DR-DCI realizes only part of it.	AtlasNav wrong; DR-DCI correct.	Complete annotated evidence does not guarantee correct synthesis, and partial annotated realization does not preclude a correct answer.

Appendix

Table 18: Representative BrowseComp-Plus trajectories illustrating the distinction between evidence realization and final-answer correctness.

Question category	Overall	Source tag	Overall
Miscellaneous	96.08	HubSpot	92.42
Intra-document	96.46	Gmail	86.74
Constrained	94.57	Jira	81.21
Basic	95.16	Google Drive	88.70
Conflicting information	92.69	Fireflies	87.22
Information not found	100.00	Linear	84.23

Appendix

Table 19: EnterpriseRAG-Bench Overall score by question category and overlapping source tag. Source tags are not mutually exclusive; the full-set Overall score is 87.00. Complete metric matrices are reported in Appendix B .

Paradigm	Representative methods	Persistent structure	Query-specific structure	Continued interaction	Primary role of structure
Sparse/dense retrieval	BM25; dense retrievers	Index only	No	Limited	Select top- $k$ evidence
Retrieval + reranking	ReasonRank ( Liu et al., 2026b ) ; neural rerankers	Index	No	Limited	Refine evidence ranking
Hierarchical/ graph RAG	RAPTOR; HippoRAG; SiReRAG; LinearRAG ( Zhuang et al., 2026a )	Yes	Usually no	Retrieval-mediated	Structured evidence selection
Trained retrieval agents	Search-R1; R1-Searcher ( Song et al., 2025 ) ; ASearcher ( Gao et al., 2025 )	Retrieval index	Adaptive querying	Via search API	Improve search policy
Relevance-guided interaction	RARG ( Li et al., 2026a )	Retrieval index	Query-conditioned priority	Yes	Guide traversal and local match visibility
Raw direct interaction	DCI	No	No	Yes	Direct corpus operations

Appendix

Table 20: Conceptual corpus-access design space. The taxonomy positions methods by interface semantics rather than comparing performance across heterogeneous experimental stacks.

Interface	Strict Acc. $\uparrow$	Cost $\downarrow$	Turns $\downarrow$
DCI	95.50	22.59	4,117
DR-DCI	94.50	28.01	5,540
AtlasNav	94.50	10.89	3,135

Appendix

Table 21: Exact 2Wiki-Global results on 400 questions over a shared 56,684-document corpus. Cost is total recorded query-time agent inference cost in CNY.

Interface	Loose $\uparrow$	Strict $\uparrow$	Cost $\downarrow$	Turns $\downarrow$
DCI	79.65	39.35	73.35	6,596
DR-DCI	81.42	44.19	43.24	6,493
AtlasNav	80.10	40.32	36.49	4,218

Appendix

Table 22: Exact FanOutQA endpoint results on 310 questions. Loose and Strict are the official answer-coverage metrics; cost is recorded query-time agent inference cost in CNY.

Figure 7: Deliverable nDCG@10 over frozen, data-driven turn checkpoints on TREC-COVID. Unfinished queries score zero at each checkpoint. The evidence-supplied reference receives the positive-document pool but not graded labels or the ideal ranking.

Symbol	Meaning	Symbol	Meaning
$q$	Question or query.	$\mathcal{Q}$	Set of evaluation questions.
$\mathcal{E}_{q}=\{e_{1},\ldots,e_{m_{q}}\}$	Required evidence slots for $q$ ; $m_{q}$ is their number.	$\tau\in\mathcal{T}=\{C,S,O,L\}$	Construction, Surface, Open, or Locate checkpoint.
$\mathcal{E}_{q}^{\tau}$	Evidence slots realized by checkpoint $\tau$ .	$r_{q}^{\tau}$	Fraction of required slots realized at $\tau$ .
$\operatorname{EB}_{\mathrm{Any/Mean/All}}^{\tau}$	Evidence Blindness metrics for total failure, mean loss, and incomplete realization.	$\mathcal{C}$	Original accessible corpus.
$M$	Language-model agent.	$I$	Corpus interface.
$B$	Interaction budget.	$A_{I}(B)$	Strict task performance of $I$ under budget $B$ .

Appendix

Table 23: Principal notation used in Sections 3–4.

Explore similar work

Follow the Entities: A Corpus Map for Agentic Search
Sep 29, 2026Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang +1Agentic Search Text Corpora
KAIST · Microsoft
GrepSeek: Training Search Agents for Direct Corpus Interaction
May 28, 2026Alireza Salemi, Chang Zeng, Atharva Nijasure +4Search Agents Text Corpora
University of Massachusetts Amherst · Princeton University · Carnegie Mellon University
A New Role for Relevance: Guiding Corpus Interaction in Agentic Search
Jul 27, 2026Jiangnan Li, Yuqing Li, Mo Yu +2Agentic Search Relevance
1Tencent · 2IIE-CAS

Sep 29, 2026, cs.CLJ/K move · Enter open · S save

Follow the Entities: A Corpus Map for Agentic Search

Soyeong Jeong, Sujay Kumar Jauhar, Sung Ju Hwang, Andrew Joohun Nam

KAIST · Microsoft

Agentic Search Text Corpora Entities Document Navigation

Open paper AskPDF arXiv

Answering questions and completing tasks over large document collections often requires connecting evidence spread across multiple documents, such as a project's approval recorded in one, its requirements in another, and its latest status in a third. Recent LLM agents approach this by iteratively searching the full corpus rather than reading only a fixed set of top-ranked documents. However, when the corpus is exposed only as a flat collection of files, a relevant document gives no indication of how it relates to others, so the agent must rediscover these relationships for every query, often missing complementary evidence while simultaneously consuming substantial additional tokens. To address this, we introduce CorpusMap, a navigation layer that organizes the corpus around its recurring entities, which are identifiable from the documents themselves and can link a single document to many others across sources. Specifically, CorpusMap represents each recurring entity as an Entity Page that aggregates information about it and links to every document that refers to it, forming a graph between entities and documents that the agent can traverse to gather otherwise disconnected evidence. Moreover, since CorpusMap is constructed offline by resolving mentions of the same entity across documents, its links are shared across queries rather than rediscovered repeatedly at inference time. Using 7 different models with 3 benchmark datasets, we show that CorpusMap improves both evidence discovery and answer quality over raw-corpus agentic search while using fewer tokens on average, and further outperforms 4 alternative navigation layers, suggesting that entities serve as effective anchors for navigating large document collections.

Abstract

Explore similar work

Follow the Entities: A Corpus Map for Agentic Search

GrepSeek: Training Search Agents for Direct Corpus Interaction

A New Role for Relevance: Guiding Corpus Interaction in Agentic Search