acceptodds
Under review as a conference paper at ICLR 2027

AtlasNav: Mitigating Evidence Blindness with Persistent Corpus Navigation

15 views

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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
agentic search, interactive corpus access, persistent corpus navigation, evidence blindness, multi-hop reasoning, large language model agents

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_GcGiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper identifies Evidence Blindness as a failure mode in interactive corpus access, where agents fail to discover or utilize relevant evidence despite its availability in the corpus. To address this issue, the authors propose AtlasNav, a framework that constructs a persistent multi-view Corpus Atlas to organize documents along dimensions. At query time, AtlasNav employs query-adaptive routing to navigate the preconstructed hierarchy, enabling agents to discover relevant evidence more efficiently. Experiments on BrowseComp-Plus, PhantomWiki, and EnterpriseRAG-Bench demonstrate improvements in answer accuracy, evidence discovery, and query-time inference efficiency over existing interactive corpus access methods.

Critical strengths
  1. The paper provides a useful fine-grained diagnosis of evidence discovery failures. By decomposing Evidence Blindness into four stages (Construction, Surface, Open, and Locate), the paper provides a structured understanding of where evidence discovery fails, beyond simply measuring final answer accuracy.

  2. The experimental results are strong and supported by relatively comprehensive evaluations. The paper evaluates AtlasNav across multiple benchmarks and LLM backbones, demonstrating strong improvements in accuracy and query-time efficiency. Additional analyses under different interaction budgets and corpus scales further support its effectiveness.

Critical weaknesses
  1. Unclear scalability and necessity of corpus-wide persistent organization. AtlasNav relies on constructing a persistent multi-view navigation structure over the entire corpus, which may introduce substantial preprocessing and maintenance costs, particularly for web-scale or continuously evolving document collections.

  2. More fundamentally, it is unclear whether such corpus-wide organization is necessary. Since individual queries typically involve only a small subset of relevant documents, an alternative approach could first retrieve a query-relevant subcorpus and then construct a local navigation structure on demand. Such a design may avoid the overhead of globally organizing and maintaining the entire corpus while retaining similar navigation benefits.

  3. The reported efficiency gains primarily reflect online inference costs, while the offline costs of constructing and maintaining the multi-view corpus atlas are not fully accounted for. Although the authors provide a preliminary amortization analysis, it remains unclear whether the overall efficiency advantage holds for large-scale or frequently updated corpora.

Questions for authors
  1. Which components account for the improvements, and how does AtlasNav compare to relevance-guided interaction baselines? Can you provide end-to-end ablations that isolate the contributions of multi-view representations, learned routing, and hierarchical navigation? A comparison with RARG under a matched interaction budget would also help clarify the benefits of persistent structure beyond relevance-based evidence prioritization.

  2. Under what corpus reuse and update conditions is persistent navigation cost-effective? Could the authors report complete offline construction costs, indexing/storage overhead, and amortized total costs at different query volumes? How would the efficiency advantage change when the corpus is frequently updated?

  3. How can AtlasNav scale to substantially larger and continuously updated corpora without frequent reconstruction of its multi-view graphs and hierarchy? Can the authors provide evidence of how construction costs, update overhead, and retrieval quality scale with corpus size?

4 comments
AuthorsOpenReview ↗

End-to-End Component Ablations, Comparison with RARG, and Why Global Persistent Structure Is Necessary

1. End-to-End Component Ablations, Comparison with RARG, and Why Global Persistent Structure Is Necessary (Question 1 & Weakness 2)

1.1 Controlled End-to-End Ablations and Comparison with RARG To isolate the contributions of multi-view representations, learned routing, and hierarchical navigation, and to compare directly against RARG under a matched interaction budget, we conducted 10 controlled end-to-end evaluations on the 166-question BrowseComp-Plus evaluation subset (the same subset used in Appendix E.1.2, Table 15) across two backbones—Qwen (Qwen-3.5-flash) and DeepSeek (deepseek-v4.1-flash-1017). All systems run inside the exact same DCI interaction harness with identical tool protocols and turn budgets:

  • Full AtlasNav: All 4 semantic views + BM25 + learned multi-view router + hierarchical parent-leaf navigation.
  • Topic / Identity / Episode / Relation + BM25 (Isolating Multi-View Representations): Each variant retains BM25 (1:1 fixed fusion) and retains hierarchical Atlas navigation, but builds the Atlas from only one semantic view (removing the other three views and the multi-view router).
  • 4-View Atlas (No BM25) & BM25 Only: Isolate the four semantic views without BM25, and pure lexical search without the Atlas.
  • RARG (Relevance-Guided Interaction Baseline): Evaluates RARG (Table 20, Appendix F.1) under the exact same agent harness and interaction budget.
  • 4-View + BM25 + Uniform Router + Hierarchy (Isolating the Learned Router on Qwen): Keeps all 4 semantic views, BM25, and hierarchical navigation, but replaces the learned router with equal weights across the 4 views and a fixed 1:1 semantic/BM25 fusion.
  • 4-View + BM25 + Learned Router + Flat (Isolating Hierarchical Navigation on Qwen): Keeps all 4 views, BM25, and the learned router's RRF document ranking (Eq. 6), but disables parent-leaf region overview, expansion, and cross-region bridges, exposing the ranked documents as a flat list.

Table A: Cross-Backbone Comparison against Single-View Variants, Simple Indexes, and RARG (166 Questions)

System / Variant (BrowseComp-Plus, 166 Questions)Qwen Acc (%) ↑Qwen Norm. Cost ↓Qwen Turn-15 Blindness (Surface / Open / Locate) ↓Qwen Turn-60 Blindness (Surface / Open / Locate) ↓DeepSeek Acc (%) ↑DeepSeek Norm. Cost ↓DeepSeek Turn-60 Blindness (Surface / Open / Locate) ↓
BM25 Only23.491.50340.717 / 0.825 / 0.8430.620 / 0.723 / 0.74795.181.03940.024 / 0.151 / 0.265
Topic + BM25 (Single-view Atlas)33.131.33650.566 / 0.735 / 0.7710.518 / 0.675 / 0.70596.391.05050.018 / 0.127 / 0.253
Identity + BM25 (Single-view Atlas)30.721.38440.620 / 0.759 / 0.7890.536 / 0.675 / 0.72396.391.23550.042 / 0.127 / 0.253
Episode + BM25 (Single-view Atlas)30.721.47450.614 / 0.765 / 0.8070.566 / 0.687 / 0.72994.581.09910.024 / 0.133 / 0.247
Relation + BM25 (Single-view Atlas)37.351.41100.572 / 0.741 / 0.7590.494 / 0.633 / 0.66396.991.06370.024 / 0.090 / 0.211
4-View Atlas (No BM25)41.571.22890.572 / 0.687 / 0.7290.476 / 0.590 / 0.63995.181.14710.042 / 0.163 / 0.283
RARG (Relevance-Guided Baseline)45.181.13910.530 / 0.639 / 0.6930.451 / 0.572 / 0.60893.981.11900.078 / 0.169 / 0.229
Full AtlasNav64.461.00000.392 / 0.500 / 0.5420.253 / 0.343 / 0.39896.991.00000.024 / 0.127 / 0.223

Table B: Step-by-Step Isolation of Multi-View Fusion, Hierarchical Navigation, and Learned Routing on Qwen (166 Questions)

Component Isolation Progression (Qwen, 166 Questions)Correct / TotalStrict Acc (%) ↑Norm. Cost ↓Turn-15 Blindness (Surface / Open / Locate) ↓Turn-30 Blindness (Surface / Open / Locate) ↓Turn-60 Blindness (Surface / Open / Locate) ↓
Best Single-View Atlas (Relation + BM25)62 / 16637.351.41100.572 / 0.741 / 0.7590.506 / 0.645 / 0.6690.494 / 0.633 / 0.663
4-View + BM25 + Learned Router + Flat (w/o Hierarchy)92 / 16655.421.10410.398 / 0.494 / 0.5480.331 / 0.446 / 0.5000.325 / 0.434 / 0.482
4-View + BM25 + Uniform Router + Hierarchy (w/o Learned Router)101 / 16660.841.11070.470 / 0.554 / 0.6140.355 / 0.446 / 0.5060.289 / 0.380 / 0.440
Full AtlasNav (4-View + BM25 + Learned Router + Hierarchy)107 / 16664.461.00000.392 / 0.500 / 0.5420.313 / 0.398 / 0.4640.253 / 0.343 / 0.398

(Note: Normalized Cost is query-time agent inference cost relative to Full AtlasNav =1.0000= 1.0000 within each backbone; Blindness reports all-slot EBAllEB_{\text{All}} at Surface / Open / Locate.)

(Continued below...)

This progression cleanly separates each component's contribution:

  1. Multi-view representations (+18.07 to +24.70 points over single-view): Even without hierarchical navigation, combining the 4 semantic views with BM25 (4-View + BM25 + Learned Router + Flat, 55.42%) substantially outperforms every single-view Atlas variant (30.72%–37.35%) and BM25 Only (23.49%) on Qwen. In offline retrieval across all 830 BrowseComp-Plus questions, fusing the 4 views also raises supporting-document Recall@30 / Surface@30 to 0.3107 / 0.4892, outperforming single full-text dense retrieval (0.2730 / 0.4349) and every individual view (0.2908–0.3026).
  2. Hierarchical navigation (+9.04 points, -9.4% cost): Comparing Full AtlasNav (64.46%) against 4-View + BM25 + Learned Router + Flat (55.42%) shows that keeping the exact same 5-channel ranking and learned router but removing parent-leaf navigation drops accuracy by 9.04 points and worsens Turn-60 Surface / Open / Locate blindness from 0.253 / 0.343 / 0.398 to 0.325 / 0.434 / 0.482.
  3. Learned routing (+3.62 points, -10.0% cost): Replacing uniform weights (Uniform Router + Hierarchy, 60.84%, cost 1.1107) with the learned router (64.46%, cost 1.0000) improves accuracy by +3.62 points, cuts query-time cost by 10.0%, and reduces early Turn-15 Surface blindness from 0.470 to 0.392.
  4. Comparison with RARG: Under matched budgets, Full AtlasNav outperforms RARG by +19.28 points on Qwen (64.46% vs. 45.18%) and +3.01 points on DeepSeek (96.99% vs. 93.98%) while reducing inference cost by 12.2% and 10.6%.

1.2 Why Corpus-Wide Persistent Organization Outperforms On-Demand Local Workspaces (DR-DCI) and Relevance-Guided Interaction (RARG) (Weakness 2) The alternative design you described—first retrieving a query-relevant subcorpus and then constructing a local structure on demand—is the exact paradigm represented by our main baseline DR-DCI (which dynamically retrieves and organizes a query-conditioned local workspace at query time; Section 1, Appendix F.1, Table 20). Meanwhile, RARG represents an alternative that uses query-conditioned relevance scores to prioritize traversal without building corpus-wide multi-view structure (Table 20). Our experiments across BrowseComp-Plus and PhantomWiki reveal two fundamental reasons why corpus-wide persistent organization outperforms both paradigms:

  1. Avoiding the first-hop retrieval bottleneck on multi-hop evidence: Both on-demand local workspace construction (DR-DCI) and direct relevance-guided prioritization (RARG) rely heavily on initial query-similarity signals to surface candidates. In multi-hop tasks, crucial bridge documents are connected through intermediate entities, events, or relations rather than directly matching the surface query. If initial retrieval fails to capture these bridge documents, on-demand local organization cannot recover them. Consequently, DR-DCI fails to surface complete supporting evidence on 49.76% of BrowseComp-Plus questions on Qwen (Table 2) and 70.5%–87.0% of PhantomWiki questions across 10K–1M scales (Figure 4), while RARG exhibits 0.530 Turn-15 Surface blindness on Qwen (vs. 0.392 for AtlasNav, Table A above). Because AtlasNav builds multi-view links and cross-region bridges across the full corpus before query time, early exploration can immediately traverse pre-linked entity, temporal, and relational neighborhoods to surface bridge documents that initial query similarity misses.
  2. Eliminating repeated online workspace reconstruction: Constructing a local workspace on demand (as in DR-DCI) consumes online interaction turns and inference budget on every query, whereas a persistent Atlas is built once offline and amortized across queries.
AuthorsOpenReview ↗

2. Complete Offline Costs, Storage Overhead, and Cost-Effectiveness Conditions

2. Complete Offline Costs, Storage Overhead, and Cost-Effectiveness Conditions (Question 2 & Weakness 3)

2.1 Complete End-to-End Offline Construction and Storage Accounting In the initial manuscript, Table 17(b) noted that CNY 127.68 (~USD 18) covered only embedding costs. In fact, all remaining offline stages are lightweight local computations that require zero LLM calls, take only 31.4 minutes in total across all 100,195 documents, and produce a compact 865.9 MiB runtime index:

Offline Preparation Stage (BrowseComp-Plus, 100,195 Docs)Wall-Clock TimeCompute & Memory FootprintDisk StorageMonetary / API Cost
1. 4-View Signature Extraction119.30 s (~2.0 min)CPU only (deterministic scoring)391.0 MiB (Parquet)CNY 0.00
2. 4-View Corpus Embedding5,245.96 s (~87.4 min)400,780 vectors (253.55M tokens)2.0 GiB (raw FP16)CNY 126.78
3. Full Atlas & Router Build (PCA + HNSW + Leiden + Anchors/Bridges + Router)1,762.65 s (29.38 min) (4×PCA: 12.59 s; 4×HNSW: ~21 s)Peak RAM: 1.98 GiB; 77 Parents, 443 Leaves; 4,902-param linear router146.7 MiB (PCA) + 399.2 MiB (HNSW) + 320.0 MiB (Atlas)CNY 0.90 (router query emb.)
Total Non-Embedding Build (Stages 1 & 3)1,881.95 s (31.4 min)Peak RAM < 2.0 GiB (Zero LLM calls)865.9 MiB (active)CNY 0.90 (~USD 0.13)
Total End-to-End Offline Preparation7,127.91 s (1.98 hrs)Local CPU/GPU + Embedding API865.9 MiBCNY 127.68 (~USD 18.00)
  • Storage Comparison: Because PCA compresses each 2,560-D view vector to 192 dimensions, the entire active runtime index (865.9 MiB) is 46% smaller than a single unprojected full-text dense index (1.6 GiB for RARG / Qwen3-Embedding-8B) and 82% smaller than the BM25 SQLite index (4.7 GiB, on top of 1.7 GiB raw corpus Parquet).

2.2 Amortized Total Cost Across Query Volumes (NN) and Reuse/Update Conditions Following Eq. (18) in Appendix C.3, the amortized total cost per query across NN queries is Ctotal(N)=Cop+Coffline/NC_{\text{total}}(N) = C^{\text{op}} + C^{\text{offline}} / N. On DeepSeek (Cop=CNY 0.6307C^{\text{op}} = \text{CNY } 0.6307 for AtlasNav vs. CNY 0.9034\text{CNY } 0.9034 for DR-DCI and CNY 1.4798\text{CNY } 1.4798 for Raw DCI; Table 17b), the amortized total cost per query across query volumes NN is:

Query Volume (NN)Raw DCI Total/QueryDR-DCI Total/QueryAtlasNav Total/Query (Full Build)AtlasNav Savings vs. DR-DCI
N=151N = 151 (Break-even vs. Raw DCI)CNY 1.4798CNY 0.9034CNY 1.4763-CNY 0.5729
N=469N = 469 (Break-even vs. DR-DCI)CNY 1.4798CNY 0.9034CNY 0.9029+CNY 0.0005
N=830N = 830 (1× Benchmark Evaluation)CNY 1.4798CNY 0.9034CNY 0.7845+CNY 0.1189 (-13.2%)
N=5,000N = 5,000CNY 1.4798CNY 0.9034CNY 0.6562+CNY 0.2472 (-27.4%)
N=10,000N = 10,000CNY 1.4798CNY 0.9034CNY 0.6435+CNY 0.2599 (-28.8%)
  • Reuse & Update Cost-Effectiveness: Even under a complete rebuild from scratch, AtlasNav is cost-effective whenever a corpus serves ≥469\ge 469 queries relative to DR-DCI (or ≥151\ge 151 queries relative to Raw DCI). More importantly, under frequent updates, full reconstruction is unnecessary: incrementally inserting a 10% corpus update (10,020 new documents) costs only CNY 12.68 in new document embeddings plus about 9 seconds of local HNSW/region updates (Section 3 below), which breaks even after just 47 queries relative to DR-DCI (or 15 queries relative to Raw DCI).
AuthorsOpenReview ↗

3. Scaling to Evolving and Larger Corpora Without Full Reconstruction

3. Scaling to Evolving and Larger Corpora Without Full Reconstruction (Question 3 & Weakness 1)

3.1 Streaming Incremental Maintenance Without Rebuilding Graphs or Hierarchy AtlasNav supports a fast incremental update path that freezes the PCA projection matrices, the global Leiden parent-leaf hierarchy, and the router, and processes incoming documents purely via: (1) deterministic signature extraction and embedding, (2) PCA projection, (3) O(log⁡∣D∣)O(\log |\mathcal{D}|) HNSW graph insertion, (4) local Parent/Leaf assignment via HNSW neighbor voting, and (5) dynamic leaf-anchor score updates.

To test continuous updates, we split the 100,195 BrowseComp-Plus documents deterministically into an initial 70% static corpus (70,136 docs) and streamed the remaining 30% (30,059 docs, containing 10.1% of benchmark gold supporting documents) in three sequential 10% batches (~10,020 docs/batch) without rebuilding the hierarchy or retraining PCA/router:

Corpus Update StageActive DocsGraph & Region Update TimeHNSW Neighbor Recall@48 (4-View Mean)Adaptive Leaf Recall (830 Qs)
70% Static Init70,13617.32 s (initial 4-view HNSW build)100.00% (initial static graph)0.8036
+10% Stream (80%)80,1568.661 s (0.49% of full Atlas build)99.064% (vs. 80% active rebuild)0.8067
+10% Stream (90%)90,1769.147 s (0.52% of full Atlas build)98.949% (vs. 90% active rebuild)0.8040
+10% Stream (100% Incremental)100,1959.708 s (0.55% of full Atlas build)98.873% (vs. 100% active rebuild)0.8077
100% Full-Rebuild Oracle100,1951,762.650 s (full Atlas build)100.000% (full-rebuild reference)0.8160

Maintaining each 10,020-document batch takes under 10 seconds (over 180× faster than full Atlas reconstruction), preserves 98.87%–99.06% HNSW neighbor Recall@48 relative to rebuilding the active HNSW graphs from scratch at each stage, matches the 100% full-rebuild oracle in supporting-document Recall@30 (rising from 0.2965 before the final batch to 0.3107 vs. 0.3107 for the full-rebuild oracle), and stays within 0.0083 of the oracle on Adaptive Leaf Recall (0.8077 vs. 0.8160).

3.2 Scaling with Corpus Size (10K to 1M Documents) Because signature extraction, embedding, and PCA projection scale strictly linearly O(∣D∣)O(|\mathcal{D}|) with document count while HNSW insertion scales logarithmically O(∣D∣log⁡∣D∣)O(|\mathcal{D}| \log |\mathcal{D}|), index construction and incremental maintenance remain highly tractable as corpora grow. Empirically, as shown in our controlled PhantomWiki scaling evaluation from 10,059 to 1,006,839 files (Section 5.3, Figure 4, Table 3, and Appendix B.2, Table 8), AtlasNav consistently maintains the lowest Surface blindness and highest strict accuracy (76.5% / 77.5% / 76.0% / 61.0% at 10K / 50K / 100K / 1M, outperforming both DCI and DR-DCI at every scale).

Reviewer_HZUpOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies failures of evidence discovery in agentic search - "evidence blindness". It proposes AtlasNav with "persistent corpus navigation" to address this issue.

There are strong smells of AI slop throughout this paper. Phrases like "evidence realization", "complementary cross-document structure", "evidence directions", and "query-specific evidence paths" sound very much like AI generated text.

For example, it's unlikely that a human wrote something like:

Together, query-adaptive routing and Atlas navigation turn persistent corpus structure into query-specific evidence paths without reconstructing the search space online.

And:

These outcome-level failures point to a deeper process-level question...

Critical strengths
  • Tackles and interesting, important problem.
  • Seems like a reasonable collection of ideas
Critical weaknesses
  • Perhaps overcomplicated? The paper adds a bunch of things all at the same time: the 4-views, BM25 fusion, routing, hierarchical organization, cross-region navigation. It's not clear where the performance is coming from and the ablations are sufficient to tease apart the factors.

  • Related to the above point - a claim of paper is saving agent inference cost: but you have to do so much else! Once again, too much going on because the technique requires offline preprocessing - how do those figure?

Questions for authors

There's so much going on in Figure 2. Where are the gains coming from?

3 comments

We thank Reviewer HZUp for recognizing that our paper tackles an important problem with a reasonable collection of ideas, and for providing candid feedback on our writing style, component complexity, and offline preprocessing costs. Below, we address all three concerns directly and have uploaded a revised manuscript reflecting these updates.


1. Replacing Abstract Phrasing with Plain Technical Language (Presentation)

As disclosed in our AI Use Statement, while the research ideas, architecture, experiments, and analyses were entirely designed and executed by the authors, we used language-editing tools to condense the initial manuscript. We agree that this introduced overly abstract compound phrases that hurt readability.

In the revised manuscript, we have edited the main text (Abstract, Introduction, and Sections 3–5) to replace abstract jargon with plain, direct technical descriptions:

  • Replacing abstract compound terms across the text:
    • "evidence realization" →\rightarrow replaced with "finding and reading supporting evidence" or "evidence discovery".
    • "evidence directions" / "query-specific evidence paths" →\rightarrow replaced with "document connections" / "prebuilt document neighborhoods and regions".
    • "complementary cross-document structure" →\rightarrow replaced with "multi-view index that links documents by topic, entity, timeline, and relation".
  • Rewriting the specific sentences you highlighted:
    • Original (Section 1, Page 1): "These outcome-level failures point to a deeper process-level question: how does evidence that remains accessible in the corpus fail to become usable as interaction unfolds?"
    • Revised (Page 1, Lines 49–50): "This raises a practical question: when supporting documents exist in the corpus, at which step of the search process does the agent miss them?"
    • Original (Section 4.3, Page 6): "Together, query-adaptive routing and Atlas navigation turn persistent corpus structure into query-specific evidence paths without reconstructing the search space online."
    • Revised (Page 6, Lines 311–314): "By weighting the five retrieval channels per query and mapping top-ranked documents to prebuilt regions, the agent can start exploring relevant document groups immediately without rebuilding a search index at query time."

2. Where Are the Gains Coming From in Figure 2? Step-by-Step Component Ablations (Weakness 1 & Question 1)

To show clearly how each component in Figure 2—the 4 semantic views, BM25 fusion, learned routing, and hierarchical parent-leaf navigation—contributes to accuracy and efficiency, we added 10 controlled end-to-end ablations on the 166-question BrowseComp-Plus subset (Appendix E.1.2, new Table 17 in the revised PDF) across Qwen (Qwen-3.5-flash) and DeepSeek (deepseek-v4.1-flash-1017). All variants use the exact same DCI agent harness and turn budget, changing one design factor at a time:

  • BM25 Only: Lexical search alone, without the 4 semantic views, router, or hierarchy.
  • Topic / Identity / Episode / Relation + BM25 (Single-View Atlas): Keeps BM25 (1:1 fusion) and hierarchical navigation, but builds the Atlas from only one semantic view instead of four.
  • 4-View Atlas (No BM25): Keeps all 4 semantic views and hierarchical navigation, but removes BM25.
  • RARG: Competitive relevance-guided interaction baseline (Appendix F.1, Table 21) under the same harness and budget.
  • 4-View + BM25 + Learned Router + Flat (w/o Hierarchy): Keeps all 4 views, BM25, and the learned router's RRF ranking (Eq. 6), but disables parent-leaf region navigation and cross-region bridges, showing ranked documents as a flat list.
  • 4-View + BM25 + Uniform Router + Hierarchy (w/o Learned Router): Keeps all 4 views, BM25, and hierarchical navigation, but replaces the learned router with equal view weights and 1:1 semantic/BM25 fusion.
  • Full AtlasNav: All components combined.
Ablation Variant (BrowseComp-Plus, 166 Questions)Qwen Acc (%) ↑Qwen Norm. Cost ↓Qwen Turn-60 Blindness (Surface / Open / Locate) ↓DeepSeek Acc (%) ↑DeepSeek Norm. Cost ↓DeepSeek Turn-60 Blindness (Surface / Open / Locate) ↓
BM25 Only23.491.50340.620 / 0.723 / 0.74795.181.03940.024 / 0.151 / 0.265
Topic + BM25 (Single-view Atlas)33.131.33650.518 / 0.675 / 0.70596.391.05050.018 / 0.127 / 0.253
Identity + BM25 (Single-view Atlas)30.721.38440.536 / 0.675 / 0.72396.391.23550.042 / 0.127 / 0.253
Episode + BM25 (Single-view Atlas)30.721.47450.566 / 0.687 / 0.72994.581.09910.024 / 0.133 / 0.247
Relation + BM25 (Single-view Atlas)37.351.41100.494 / 0.633 / 0.66396.991.06370.024 / 0.090 / 0.211
4-View Atlas (No BM25)41.571.22890.476 / 0.590 / 0.63995.181.14710.042 / 0.163 / 0.283
RARG (Relevance-Guided Baseline)45.181.13910.451 / 0.572 / 0.60893.981.11900.078 / 0.169 / 0.229
4-View + BM25 + Learned Router + Flat55.421.10410.325 / 0.434 / 0.482(Qwen isolation)(Qwen isolation)(Qwen isolation)
4-View + BM25 + Uniform Router + Hierarchy60.841.11070.289 / 0.380 / 0.440(Qwen isolation)(Qwen isolation)(Qwen isolation)
Full AtlasNav64.461.00000.253 / 0.343 / 0.39896.991.00000.024 / 0.127 / 0.223

(Normalized Cost is query-time inference cost relative to Full AtlasNav =1.0000= 1.0000; Blindness reports all-slot missing rates at Surface / Open / Locate.)

Every component provides a clear, distinct gain (especially on Qwen, where the model relies heavily on the index to find multi-hop evidence):

  1. Combining the 4 semantic views (+18.07 to +24.70 points): Moving from any single view + BM25 (30.72%–37.35%) to all 4 views + BM25 (Flat, 55.42%) yields the largest jump, confirming that different multi-hop questions require different semantic links (topic, entity, timeline, or relation).
  2. BM25 fusion (+22.89 points on Qwen, +1.81 points on DeepSeek): Adding BM25 to the 4 semantic views (No BM25 at 41.57% vs. Full at 64.46%) is essential for matching exact names and rare identifiers while reducing cost by 18.6%.
  3. Hierarchical navigation (+9.04 points, -9.4% cost): Adding parent-leaf region hierarchy and cross-region bridges on top of the flat 5-channel ranking (Flat at 55.42% vs. Full at 64.46%) adds +9.04 accuracy points and lowers cost by 9.4%.
  4. Learned router (+3.62 points, -10.0% cost): Replacing equal channel weights (Uniform Router at 60.84%, cost 1.1107) with our learned router (64.46%, cost 1.0000) improves accuracy by +3.62 points and cuts query-time cost by 10.0%.

3. How Much Does Offline Preprocessing Actually Cost? (Weakness 2)

While Figure 2 shows several offline steps, none of them invoke a generative LLM. Signature extraction uses simple regex/keyword rules (Appendix D.3), dimensionality reduction uses linear PCA, graph building uses standard HNSW and Leiden clustering, and the router is a tiny 4,902-parameter linear layer (Appendix D.4). Below is the complete end-to-end offline cost and time across all 100,195 documents in BrowseComp-Plus (Appendix E.1.3, Table 18b in the revised PDF):

Offline Step (100,195 Documents)Wall-Clock TimeCompute & MemoryDisk StorageMonetary / API Cost
1. 4-View Text Extraction119.30 s (2.0 min)CPU only (regex/lexical rules)391.0 MiBCNY 0.00
2. 4-View Embedding Encoding5,245.96 s (87.4 min)400,780 vectors (253.55M tokens)2.0 GiB (raw)CNY 126.78
3. PCA + HNSW + Leiden + Router1,762.65 s (29.4 min)Peak RAM < 2.0 GiB (Zero LLM calls)865.9 MiB (runtime)CNY 0.90 (router query emb.)
Total Non-Embedding Build (Steps 1 & 3)1,881.95 s (31.4 min)Local CPU/GPU only865.9 MiBCNY 0.90 (USD 0.13)
Total End-to-End Offline Preparation7,127.91 s (1.98 hrs)Local CPU/GPU + Embedding API865.9 MiBCNY 127.68 (USD 18.00)
  • Minimal offline cost and fast break-even: Outside embedding (CNY 126.78 / USD 18.00), all graph, hierarchy, and router construction takes just 31.4 minutes on local hardware and costs CNY 0.90. Because AtlasNav saves CNY 0.2727 per query over DR-DCI and CNY 0.8490 per query over Raw DCI on DeepSeek (Appendix E.1.3, Table 18b), the entire offline cost is paid back after just 469 queries vs. DR-DCI and 151 queries vs. Raw DCI (less than 57% of the 830-query benchmark itself).
  • Compact storage and 9-second incremental updates: The active index is only 865.9 MiB (smaller than the 1.6 GiB single-view RARG dense index and 4.7 GiB BM25 index). Moreover, when streaming new documents into the corpus in 10% batches (10,020 docs/batch), incremental HNSW insertion and region voting take only 8.66–9.71 seconds per batch without rebuilding the hierarchy or retraining the router, matching the full-rebuild oracle in Recall@30 (0.3107).
Reviewer_K97oOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces AtlasNav, a framework for interactive corpus access that builds a persistent, multi-view navigation structure before query time. It organizes documents through four views—Topic, Identity, Episode, and Relation—and combines hierarchical navigation with query-adaptive routing. The paper also introduces Evidence Blindness to characterize failures at successive stages of evidence access. Experiments report improved answer accuracy and reduced query-time inference costs compared with DCI and DR-DCI.

Critical strengths
  • The paper addresses a practically relevant problem: evidence may be available in a corpus but remain undiscovered within a limited interaction budget.

  • The framework combines reusable corpus organization with adaptive exploration while preserving access to original documents.

  • The evaluation spans multiple backbones and datasets, with consistent improvements over the main baselines on BrowseComp-Plus. The analysis also examines evidence access throughout the interaction process rather than relying solely on final-answer accuracy.

Critical weaknesses
  • The incremental value of Evidence Blindness is unclear. Its metrics appear closely related to evidence missing rates measured at different stages. While the distinction between document exposure, document opening, and evidence localization is useful, the paper should clarify what additional insights this formulation provides beyond existing coverage and localization metrics.

  • The construction of the four views is insufficiently specified. The method constructs four textual signatures before encoding them and building the corresponding graphs. However, embeddings alone do not establish the intended semantic distinctions. The appendix describes passage selection, backfill sampling, length limits, and encoding configurations, but does not sufficiently explain the rules used to extract each view’s text. This makes a central component difficult to assess and reproduce. If the construction relies primarily on manually designed selection heuristics, their novelty, robustness, and contribution should be established through appropriate comparisons.

  • The distinction from existing structured approaches is not sufficiently established. Persistent structures, cross-document connections, and hierarchical organization are already present in methods such as GraphRAG. Beyond the four-view design, it remains unclear which aspects of persistent navigation constitute the distinctive contribution. If navigation is a separable mechanism, applying it to an existing graph index—or combining it with GraphRAG—would help clarify its contribution.

  • The adequacy of the fixed four-view organization remains uncertain. Project-oriented tasks require integrating evidence across sources, topics, and evolving project states. Codebase-related tasks may additionally depend on symbol resolution, call relationships, types, and data flow. Broadly assigning these structures to Identity or Relation does not establish that the resulting representations preserve them effectively. The paper would benefit from a clearer account of the framework’s applicability and limitations.

  • Key baselines and ablations are missing. The comparisons establish improvements over DCI and DR-DCI, but do not isolate the benefits of the proposed design relative to competitive persistent structured indexes. Task-appropriate GraphRAG baselines and controlled comparisons against simpler indexing alternatives would strengthen the evaluation. Differences in published experimental settings explain why existing scores cannot be compared directly, but do not adequately justify omitting evaluations under unified conditions. Further ablations are also needed to distinguish the benefits of the four semantic views from those of using multiple representations and retrieval channels.

  • The efficiency analysis does not fully account for offline preparation. The paper reports partial embedding preparation costs, but does not provide a complete account of index construction and router preparation costs or compare them with existing methods. Consequently, the reported query-time savings do not establish an overall cost advantage. A fuller analysis of construction time, resource usage, and amortization across query volumes would make the efficiency claims more persuasive.

Questions for authors

See weaknesses above.

7 comments
AuthorsOpenReview ↗

Value of Evidence Blindness Beyond Coverage and Localization

1. Value of Evidence Blindness Beyond Coverage and Localization (Weakness 1)

Coverage and Localization (Section 2) are endpoint metrics: they only measure whether supporting documents or passages were touched by the end of a trajectory. When an agent fails under a finite budget, endpoint metrics treat the trajectory as a black box and cannot explain where the failure occurred.

The core value of Evidence Blindness is its process-level interpretability. It decomposes an endpoint miss into four sequential stages across model-visible events: Construction (evidence exists in the corpus), Surface (supporting documents appear in agent-visible observations), Open (full document text is read into context), and Locate (the decisive factual span enters context). This stage-wise interpretability is valuable in its own right and directly guides system design:

  • Separating navigation failure from selection failure (Construction →\rightarrow Surface →\rightarrow Open): A drop from Construction to Surface means the index failed to expose the right documents, requiring better cross-document organization. A drop from Surface to Open means the right documents appeared in the observations, but the agent failed to prioritize and open them before running out of turns. For example, on PhantomWiki (Appendix B.2, Table 8), scaling from 100K to 1M files drops the Surface-to-Open conversion rate from 81.82% to 57.43%. While endpoint Coverage only shows an overall drop, Evidence Blindness reveals that at million-scale, the bottleneck shifts from finding candidates to deciding which surfaced files to open. This directly motivated our representative region anchors (Appendix D.3) to give the agent clear previews before opening files.
  • Separating reading failure from multi-hop bottlenecks (Open →\rightarrow Locate and EBAllEB_{\text{All}}): A drop from Open to Locate shows that a document was opened but its decisive span was missed. Furthermore, unlike standard Coverage (which averages across documents, similar to 1−EBMean1 - EB_{\text{Mean}}), tracking EBAnyEB_{\text{Any}} (total failure), EBMeanEB_{\text{Mean}} (average loss), and EBAllEB_{\text{All}} (incomplete evidence set) exposes why multi-hop search is fragile. In Figure 1(b), realizing all required slots (EBAll=0EB_{\text{All}} = 0) yields 94.2%–95.4% accuracy, whereas missing even one slot (EBAll>0EB_{\text{All}} > 0) drops accuracy to 26.3%–51.2%. This directly motivated our router's **worst-positive bottleneck
AuthorsOpenReview ↗

Specification and Validation of the Four-View Construction

2. Specification and Validation of the Four-View Construction (Weakness 2)

2.1 Deterministic Text Extraction Rules We do not encode the same document text four times. Before running the embedding model, we extract four distinct text summaries ("signatures") from each document using deterministic, corpus-only rules (included in our initial Supplementary Material, and now written out step-by-step in Appendix D.3):

  1. Sentence segmentation and feature counting: Each cleaned document is split into deduplicated sentences (18–720 characters). For each sentence, we count seven lexical/regex indicators—named entities (entity), four-digit years (year), numbers (number), relational connectives (relation), temporal markers (time), event/place nouns (event), and identity/category nouns (identity)—plus token length (info, capped at 45).
  2. View-specific linear scoring: Each sentence is scored by four linear rules targeting distinct cross-document connections:
    • Topic: 0.25⋅info+1.0⋅entity+1.0⋅event0.25 \cdot \text{info} + 1.0 \cdot \text{entity} + 1.0 \cdot \text{event} (overall subject and content areas).
    • Identity: 4.0⋅entity+1.5⋅identity+0.1⋅info4.0 \cdot \text{entity} + 1.5 \cdot \text{identity} + 0.1 \cdot \text{info} (people, organizations, works, and titles).
    • Episode: 4.0⋅year+2.0⋅time+1.5⋅event+0.35⋅number4.0 \cdot \text{year} + 2.0 \cdot \text{time} + 1.5 \cdot \text{event} + 0.35 \cdot \text{number} (time periods, places, events, and stages).
    • Relation: 3.0⋅relation+1.25⋅entity+0.75⋅number3.0 \cdot \text{relation} + 1.25 \cdot \text{entity} + 0.75 \cdot \text{number} (roles, causes, comparisons, and quantities).
  3. Diversity filtering and uniform backfill: For each view, we greedily pick top-scoring sentences while skipping any sentence with ≥0.68\ge 0.68 token Jaccard overlap with already selected ones. To ensure keyword lists never create a hard cutoff, we append evenly spaced backfill sentences sampled uniformly across the document:
    • Topic (≤6,144\le 6,144 chars): top 8 diverse sentences + 8 uniform backfill sentences.
    • Identity, Episode, Relation (≤4,096\le 4,096 chars each): top 12 diverse sentences + 3 uniform backfill sentences.
  4. View-conditioned encoding and indexing: We prepend a brief view instruction and document header to each signature (and pair queries with matching view instructions), encode them with Qwen3.7-Text-Embedding (2,560-D), project each view via an independent 192-D PCA matrix, and build four separate HNSW graphs (Appendix D.3, Table 12).

2.2 Validating the Four Semantic Views (Not Just "Four Channels") Three results confirm that the gains come from our semantic views rather than simply using four channels (end-to-end QA ablations are reported in Section 3 below):

  1. Distinct graph topologies (Section 4.2): The four HNSW graphs share only 26%–39% of their edges, and 79% of documents have a different nearest neighbor in at least one view.
  2. Role-swap ablation under fixed four-channel capacity (Appendix E.1.2, Table 15): On the 166-question BrowseComp-Plus subset, we kept all four views, the exact same four-channel capacity, and the same router, and only swapped which views form parent regions vs. leaf regions. Using Topic + Identity for parents and Episode + Relation for leaves achieves 95.78% accuracy (27.43 turns), whereas reversing them drops accuracy by 7.23 points to 88.55% (p=0.00754p = 0.00754, 36.02 turns). If the gain came merely from channel count, swapping view roles would not degrade performance.
  3. New offline retrieval ablation across all 830 BrowseComp-Plus questions: Evaluating initial retrieval before agent interaction across all 830 questions shows that four-view fusion consistently outperforms every single view and a single full-text dense baseline:
Offline Retrieval Variant (BrowseComp-Plus, All 830 Questions)Doc Recall@10 ↑Doc Recall@30 ↑Surface@10 ↑Surface@30 ↑
Single Full-Text Dense Embedding0.17700.27300.31690.4349
Topic View Only0.18430.29770.32290.4771
Identity View Only0.18610.30260.31200.4723
Episode View Only0.18860.29080.31330.4578
Relation View Only0.17740.29360.29280.4675
Full Atlas 4-View Fused0.19030.31070.33130.4892

3. Distinction from GraphRAG, New Baselines, and End-to-End Ablations (Weaknesses 3 & 5)

3.1 Conceptual Distinction from GraphRAG and HippoRAG While GraphRAG and HippoRAG also build persistent structures (Section 2, Appendix F.1), AtlasNav differs in two key ways:

  • Static retrieval vs. interactive navigation: GraphRAG and HippoRAG use graphs for one-shot retrieval (running community summarization or Personalized PageRank once to fill the prompt). AtlasNav exposes its hierarchy as an interactive navigation interface (Section 4.2, Appendix D.5) with stable parent-leaf addresses, region anchor previews, and cross-region bridges that the agent actively explores across turns (overview -> expand -> follow bridges -> open files).
  • Lightweight signature graphs vs. LLM OpenIE extraction: GraphRAG and HippoRAG run an LLM offline over every document to extract entities, triples, or summaries, whereas AtlasNav builds graphs purely from deterministic signatures and embeddings with zero offline LLM calls.

3.2 Experimental Setup for New Controlled End-to-End Ablations During the rebuttal period, we ran 10 controlled end-to-end evaluations on the 166-question BrowseComp-Plus subset (the same stratified subset used in Appendix E.1.2, Table 15) using two backbones—Qwen (Qwen-3.5-flash) and DeepSeek (deepseek-v4.1-flash-1017). All variants share the exact same DCI agent harness, tool-call protocol, and turn budget, differing only in the components being isolated:

  • Full AtlasNav: All 4 semantic views + BM25 + learned multi-view router + hierarchical parent-leaf navigation.
  • 4-View + BM25 + Uniform Router + Hierarchy (Isolating the Learned Router): Keeps all 4 semantic views, BM25, and hierarchical navigation, but replaces the learned router with fixed equal weights across the 4 views and a 1:1 semantic/BM25 fusion.
  • 4-View + BM25 + Learned Router + Flat (Isolating Hierarchical Navigation): Keeps all 4 semantic views, BM25, and the learned router's RRF document ranking (Eq. 6), but disables parent-leaf region overview, region expansion, and cross-region bridges, exposing the ranked documents as a flat list.
  • Topic / Identity / Episode / Relation + BM25 (Applying Navigation to a Single View Graph): Each variant retains BM25 (1:1 fixed fusion) and retains hierarchical Atlas navigation, but builds the Atlas from only one semantic view (removing the other three views and the multi-view router). Note that Identity + BM25 and Relation + BM25 directly test applying AtlasNav's navigation mechanism to a single entity graph or relation graph.
  • 4-View Atlas (No BM25) & BM25 Only: Isolate the four semantic views without lexical matching, and pure lexical matching without the Atlas.
  • RARG (Relevance-Guided Interaction Baseline): Evaluates the competitive relevance-guided interaction baseline (Table 20) under the exact same agent harness and budget.
Variant (BrowseComp-Plus, 166 Questions)Qwen Acc (%) ↑Qwen Norm. Cost ↓Qwen Turn-15 Blindness (Surface / Open / Locate) ↓Qwen Turn-60 Blindness (Surface / Open / Locate) ↓DeepSeek Acc (%) ↑DeepSeek Norm. Cost ↓DeepSeek Turn-60 Blindness (Surface / Open / Locate) ↓
BM25 Only23.491.50340.717 / 0.825 / 0.8430.620 / 0.723 / 0.74795.181.03940.024 / 0.151 / 0.265
Topic + BM25 (Single-view Atlas)33.131.33650.566 / 0.735 / 0.7710.518 / 0.675 / 0.70596.391.05050.018 / 0.127 / 0.253
Identity + BM25 (Single-view Atlas)30.721.38440.620 / 0.759 / 0.7890.536 / 0.675 / 0.72396.391.23550.042 / 0.127 / 0.253
Episode + BM25 (Single-view Atlas)30.721.47450.614 / 0.765 / 0.8070.566 / 0.687 / 0.72994.581.09910.024 / 0.133 / 0.247
Relation + BM25 (Single-view Atlas)37.351.41100.572 / 0.741 / 0.7590.494 / 0.633 / 0.66396.991.06370.024 / 0.090 / 0.211
4-View Atlas (No BM25)41.571.22890.572 / 0.687 / 0.7290.476 / 0.590 / 0.63995.181.14710.042 / 0.163 / 0.283
RARG (Relevance-Guided Baseline)45.181.13910.530 / 0.639 / 0.6930.451 / 0.572 / 0.60893.981.11900.078 / 0.169 / 0.229
4-View + BM25 + Learned Router + Flat55.421.10410.398 / 0.494 / 0.5480.325 / 0.434 / 0.482———
4-View + BM25 + Uniform Router + Hierarchy60.841.11070.470 / 0.554 / 0.6140.289 / 0.380 / 0.440———
Full AtlasNav64.461.00000.392 / 0.500 / 0.5420.253 / 0.343 / 0.39896.991.00000.024 / 0.127 / 0.223

(Note: Normalized Cost is query-time agent inference cost divided by Full AtlasNav's cost within each backbone; Blindness reports all-slot EBAllEB_{\text{All}} at Surface / Open / Locate.)

(Continued below...)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Manuscript Revisions and Table Numbering Mapping

We thank the Area Chair and all three reviewers (Reviewers K97o, HZUp, and GcGi) for their constructive feedback. We have uploaded a revised manuscript PDF incorporating the new controlled experiments, algorithmic specifications, lifecycle cost accounting, and writing edits discussed in our individual responses:

  1. Writing Clarity & Direct Phrasing (Main Text, Pages 1–9): Replaced abstract compound terms (e.g., "evidence realization", "evidence directions", "complementary cross-document structure") with plain, direct technical descriptions across the Abstract, Introduction, and Sections 3–5 (Reviewer HZUp).
  2. Applicability & Scope Boundaries (Section 5.4, Page 9 & Appendix A, Page 16): Clarified empirical performance on codebase tasks (GitHub: 82.80% Overall, 93.33% Correctness in Table 10), exhaustive aggregation boundaries on project/completeness queries, and how domain-specific structural graphs can be added as extra views (Reviewer K97o).
  3. Explicit Four-View Signature Rules (Appendix D.3, Page 21): Added the exact deterministic sentence-level feature definitions, linear scoring equations, Jaccard diversity ceiling (0.680.68), and uniform positional backfill budgets for Topic, Identity, Episode, and Relation (Reviewer K97o).
  4. New Offline & End-to-End Component Ablations and RARG Comparison (Appendix E.1.2, Page 25, New Table 17): Added full 830-question offline retrieval ablations and 10 controlled end-to-end evaluations on the 166-question subset across Qwen-3.5-flash and DeepSeek-v4.1-flash-1017, isolating the 4 semantic views, BM25 fusion, hierarchical navigation, and learned routing, alongside a direct matched-budget comparison with RARG (Reviewers K97o, HZUp, GcGi).
  5. Complete Offline Cost Accounting, Storage Footprint & Streaming Incremental Updates (Appendix E.1.3, Page 26, Table 18b): Added the complete wall-clock time (31.4 min non-embedding local build), disk storage (865.9 MiB active runtime index), query-volume amortization analysis, and streaming incremental update results (8.66–9.71 s per 10% batch across 70% →\rightarrow 80% →\rightarrow 90% →\rightarrow 100% splits, preserving 98.87%–99.06% HNSW neighbor Recall@48 and matching the full-rebuild oracle at 0.3107 Recall@30) (Reviewers K97o, HZUp, GcGi).

Note on Table Numbering in the Revised PDF

Because we inserted the new end-to-end ablation and RARG comparison table as Table 17 in Appendix E.1.2 (Page 25), subsequent tables in Appendices E–G have shifted by +1+1 in the revised PDF compared to the initial submission referenced in our individual responses:

  • New Table 17 (Page 25): End-to-end component ablations and matched-budget RARG comparison (newly added).
  • Table 17(a/b) in initial submission →\rightarrow now Table 18(a/b) (Page 26) (Cost distributions & offline amortization).
  • Table 18 in initial submission →\rightarrow now Table 19 (Page 27) (Representative BrowseComp-Plus trajectories).
  • Table 19 in initial submission →\rightarrow now Table 20 (Page 27) (EnterpriseRAG-Bench category & source breakdown).
  • Table 20 in initial submission →\rightarrow now Table 21 (Page 28) (Conceptual corpus-access design space).
  • Table 21 in initial submission →\rightarrow now Table 22 (Page 29) (2Wiki-Global results).
  • Table 22 in initial submission →\rightarrow now Table 23 (Page 29) (FanOutQA results).
  • Table 23 in initial submission →\rightarrow now Table 24 (Page 31) (Notation summary).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.