- 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
-
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.
-
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
-
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.
-
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.
-
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
-
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.
-
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?
-
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
This progression cleanly separates each component's contribution:
- 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%) andBM25 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). - Hierarchical navigation (+9.04 points, -9.4% cost): Comparing
Full AtlasNav(64.46%) against4-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. - 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. - Comparison with
RARG: Under matched budgets, Full AtlasNav outperformsRARGby +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:
- 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
RARGexhibits 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. - 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.
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 Time | Compute & Memory Footprint | Disk Storage | Monetary / API Cost |
|---|---|---|---|---|
| 1. 4-View Signature Extraction | 119.30 s (~2.0 min) | CPU only (deterministic scoring) | 391.0 MiB (Parquet) | CNY 0.00 |
| 2. 4-View Corpus Embedding | 5,245.96 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 router | 146.7 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 Preparation | 7,127.91 s (1.98 hrs) | Local CPU/GPU + Embedding API | 865.9 MiB | CNY 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 GiBfor RARG /Qwen3-Embedding-8B) and 82% smaller than the BM25 SQLite index (4.7 GiB, on top of1.7 GiBraw corpus Parquet).
2.2 Amortized Total Cost Across Query Volumes () and Reuse/Update Conditions Following Eq. (18) in Appendix C.3, the amortized total cost per query across queries is . On DeepSeek ( for AtlasNav vs. for DR-DCI and for Raw DCI; Table 17b), the amortized total cost per query across query volumes is:
| Query Volume () | Raw DCI Total/Query | DR-DCI Total/Query | AtlasNav Total/Query (Full Build) | AtlasNav Savings vs. DR-DCI |
|---|---|---|---|---|
| (Break-even vs. Raw DCI) | CNY 1.4798 | CNY 0.9034 | CNY 1.4763 | -CNY 0.5729 |
| (Break-even vs. DR-DCI) | CNY 1.4798 | CNY 0.9034 | CNY 0.9029 | +CNY 0.0005 |
| (1× Benchmark Evaluation) | CNY 1.4798 | CNY 0.9034 | CNY 0.7845 | +CNY 0.1189 (-13.2%) |
| CNY 1.4798 | CNY 0.9034 | CNY 0.6562 | +CNY 0.2472 (-27.4%) | |
| CNY 1.4798 | CNY 0.9034 | CNY 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 queries relative to DR-DCI (or 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).
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) 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 Stage | Active Docs | Graph & Region Update Time | HNSW Neighbor Recall@48 (4-View Mean) | Adaptive Leaf Recall (830 Qs) |
|---|---|---|---|---|
| 70% Static Init | 70,136 | 17.32 s (initial 4-view HNSW build) | 100.00% (initial static graph) | 0.8036 |
| +10% Stream (80%) | 80,156 | 8.661 s (0.49% of full Atlas build) | 99.064% (vs. 80% active rebuild) | 0.8067 |
| +10% Stream (90%) | 90,176 | 9.147 s (0.52% of full Atlas build) | 98.949% (vs. 90% active rebuild) | 0.8040 |
| +10% Stream (100% Incremental) | 100,195 | 9.708 s (0.55% of full Atlas build) | 98.873% (vs. 100% active rebuild) | 0.8077 |
| 100% Full-Rebuild Oracle | 100,195 | 1,762.650 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 with document count while HNSW insertion scales logarithmically , 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).
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)
BM25 OnlyTopic + BM25(Single-view Atlas)Identity + BM25(Single-view Atlas)Episode + BM25(Single-view Atlas)Relation + BM25(Single-view Atlas)4-View Atlas (No BM25)RARG(Relevance-Guided Baseline)Full AtlasNavTable B: Step-by-Step Isolation of Multi-View Fusion, Hierarchical Navigation, and Learned Routing on Qwen (166 Questions)
Relation + BM25)4-View + BM25 + Learned Router + Flat(w/o Hierarchy)4-View + BM25 + Uniform Router + Hierarchy(w/o Learned Router)Full AtlasNav (4-View + BM25 + Learned Router + Hierarchy)(Note: Normalized Cost is query-time agent inference cost relative to Full AtlasNav =1.0000 within each backbone; Blindness reports all-slot EBAll at Surface / Open / Locate.)
(Continued below...)