Search agents can use cross-document graphs to discover sources through multi-hop connections and choose subsequent actions based on candidate rankings and relationships within a limited observation window. However, documents may rank highly because their numerous connections make them easier to reach, without necessarily providing more valuable evidence. Graph diffusion methods such as Personalized PageRank (PPR) aggregate propagation contributions from multiple sources, allowing highly connected distractor nodes to crowd the observation window. Flat node scores also obscure cross-document evidence trails, limiting continued exploration. We propose Contrastive Evidence Paths (CEP), which scores candidates by their strongest individual paths on a query-conditioned document graph and calibrates these scores against the same graph's background visitation distribution to mitigate structural bias. CEP ranks candidates by the calibrated scores and retains the corresponding source paths. Building on CEP, we develop Structured Corpus Interaction (SCI), which organizes candidates and their connecting paths into a structured Map, enabling agents to verify source documents and pursue missing information through expansion from anchor documents. Experiments on BrowseComp-Plus and TREC-COVID show that SCI achieves state-of-the-art answer accuracy and nDCG@10 with the lowest generation cost. Compared with the strongest baselines by task performance, maximum reductions in tool calls and total generation costs exceed 50% and 40%, respectively. Further analysis confirms that SCI with CEP suppresses ranking inflation induced by structural connections and improves gold recall and exploration efficiency.
Figures & tables
Figure 1. Illustrative comparison of graph diffusion retrieval and contrastive evidence paths retrieval. Conceptual comparison on the same document graph. Graph diffusion prioritizes a highly connected distractor and exposes a flat node ranking. CEP combines strongest path support with a background visitation distribution to calibrate candidate priorities and retains source paths in the agent observation window.
Figure 2. Overview of SCI. (a) Shared entities connect source documents into a document graph with static transition probabilities. (b) CEP adjusts transition probabilities using the original question, computes strongest path support Mi(v) from query-relevant starting points, and calibrates candidate scores using the background visitation distribution bQ(v) . It collects documents and builds a map. (c) The agent uses the map and previews to choose actions: verify collected source text with read and bash , expand from an anchor to address an information gap with graphexpand(a,q) , or initiate global retrieval with graphsearch . Shared entities connect documents into a graph. The original question adjusts transition probabilities, while the original question and current query jointly select starting points. CEP calibrates strongest path support using background visits and collects documents into a workspace with a navigation map. The agent explores through reading, command-line checks, global search, and anchor expansion.
GPT-6-Luna
DeepSeek-V4.1-Flash
Qwen3.8-Flash
Method
Acc. ↑
Turns ↓
Tools ↓
Total Cost ↓
Acc. ↑
Turns ↓
Tools ↓
Total Cost ↓
Acc. ↑
Turns ↓
Tools ↓
Total Cost ↓
Conventional retrieval agents
Retrieval-Agent
80.00
16.43
35.97
$33.15
87.95
20.47
36.57
$26.22
82.41
24.54
52.64
$133.87
NeMo Agent
72.05
8.09
27.64
$23.06
80.12
10.47
30.69
$25.72
76.87
15.77
31.86
$56.86
Corpus interaction agents
DCI
78.07
19.94
29.85
$52.83
83.86
23.43
34.79
$40.77
84.58
33.51
37.27
$61.46
Table 1. Main results on BrowseComp-Plus with a 100K-document corpus. Turns and Tools are averaged per question. Cost is the total generation cost in USD for all 830 queries. Bold indicates the best values.
Method
nDCG@10 ↑
Turns ↓
Tools ↓
Total Cost ↓
DCI
68.85
10.18
56.00
$1.11
DR-DCI
67.17
6.20
23.20
$0.16
RISE-BM25
58.42
9.14
10.72
$0.19
RARG++
75.97
11.98
20.66
$0.29
AtlasNav
57.79
7.00
25.06
$0.26
SCI
81.77
5.24
13.72
$0.15
Table 2. Main results on TREC-COVID with GPT-6-Luna. Turns and Tools are averaged per query. Cost is the total generation cost for 50 queries. Best values are bold.
Figure 3. Corpus scaling results. (a) Accuracy of SCI and RARG++. (b) Mean tool time of graphsearch and graphexpand . Accuracy of SCI and RARG++ and mean execution times of graphsearch and graphexpand as the corpus grows from 100K to 1M documents. SCI maintains higher accuracy across all tested sizes. At 1M documents, graphsearch and graphexpand take 0.877 and 0.264 seconds on average, respectively.
Retrieval budget per query
Top10
Type
Method
100
500
1,000
preview
Conventional retrieval
BM25
34.74
50.80
58.31
19.85
Dense
45.14
67.76
77.16
20.83
Hybrid
41.92
62.02
71.28
20.33
Graph retrieval
PPR
45.83
65.05
74.12
21.88
PCST
40.71
60.35
68.42
–
Table 3. Retrieval algorithm ablation on BrowseComp-Plus. Gold recall (%) is measured after initial retrieval. PCST returns a selected subtree rather than a ranked list, so Top10 preview recall is not reported.
Figure 4. CEP component ablation. (a) Initial workspace gold recall across retrieval budgets. (b) Top10 preview gold recall. Workspace recall curves and preview recall bars for full CEP, without path support, without background calibration, and without both.
Configuration
Acc. ↑
Turns ↓
Tools ↓
graphsearch
graphexpand
read
bash
SCI
93
14.15
18.61
5.15
0.88
6.68
5.90
w/o map
83
12.52
19.35
5.47
0.70
7.57
5.61
w/o expansion
83
13.08
21.40
4.46
–
7.93
9.01
w/o both
80
12.66
21.95
4.74
–
8.39
8.82
Table 4. Structured map ablation on BrowseComp-Plus with GPT-6-Luna. Turns and tool calls are averaged per question. Bold indicates the best values.
Figure 5. Evolution of evidence coverage. Coverage after the initial search versus cumulative coverage over the full agent trajectory. First-search and cumulative gold recall for workspace documents, map sources, and preview sources.
Expansion outcome
Documents
Read
Read or Bash
Newly acquired
67
32
37
Newly shown in map
372
180
212
Table 5. Gold documents newly added to the workspace or first shown in the map from the existing workspace by local expansion. The last two columns count documents later accessed via read and via read or bash , respectively.
Mean degree
Highly connected (%)
Top- k
PPR
CEP
PPR
CEP
10
182.29
66.56
56.76
33.55
100
220.63
57.29
56.13
29.36
1,000
171.00
61.30
50.26
30.68
Table 6. Mean degree and proportion of highly connected nodes among documents retrieved by PPR and CEP.
Figure 6. Gold document rankings under PPR and CEP. (a) Rank pairs within the top 1,000 under both methods, with medians computed over all 2,407 rank pairs. (b) Proportions of rank wins and ties across all rank pairs.
Method
Mean distractor rank Before → After
Top10 entry (%) Before → After
PPR
50.67→24.96
0.00→15.22
CEP w/o calibration
42.21→42.21
0.84→0.84
Full CEP
45.79→75.87
0.00→0.00
Table 7. Mean rank and Top10 entry rate of distractor documents before and after controlled edge addition. Higher mean ranks and lower Top10 entry rates indicate stronger suppression of distractors.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Setting
Entity document frequency limit τ
4
Rank offset k
60
Retrieval starting points per query R
1000
Background restart probability α
0.5
Background estimation iterations Tb
24
Hop discount λ
0.5
Appendix
Table 8. SCI retrieval parameters.
Model
Input
Cache hit
Cache write
Output
GPT-6-Luna ( ≤ 272K)
0.10
0.01
0.125
0.50
GPT-6-Luna ( > 272K)
0.20
0.02
0.25
0.75
DeepSeek-V4.1-Flash
0.15
0.003
–
0.60
Qwen3.8-Flash
0.15
0.016
0.20
0.47
Appendix
Table 9. Generation pricing in USD per million tokens. GPT-6-Luna tiers refer to input tokens per request. DeepSeek-V4.1-Flash uses off-peak rates.
Method
Acc. ↑
Turns ↓
Tools ↓
Specially trained deep research agents
OpenResearcher-30B-A3B
46.00
67.64
66.44
QUEST-35B-RL
48.00
56.14
53.79
Corpus interaction agents (Qwen3.8-27B)
DCI
77.00
27.53
74.94
DR-DCI
71.00
27.60
70.26
Appendix
Table 10. Supplementary agent comparison on 100 BrowseComp-Plus queries.
Method
Corpus size
Acc. ↑
Turns ↓
Tools ↓
Total Cost ↓
Native graph RAG pipelines
HippoRAG2
1,545
42.00
—
—
—
PathRAG
1,545
28.00
—
—
—
LinearRAG
1,545
28.00
—
—
—
Graph retrieval tools within Pi Agent
HippoRAG2 + Pi Agent
1,545
89.00
12.49
26.74
$2.43
Appendix
Table 11. Reference comparison of graph-based RAG methods and SCI. Cost is the total generation cost in USD for 100 queries. For a fair cost comparison among agents, we only report costs for agent methods.
Figure 7. Two local expansion cases on BrowseComp-Plus, with full questions, abbreviated tool calls, and returned map excerpts. Quotes reproduce source text. R denotes the interaction turn. Bracketed numbers are corpus document IDs. Two recorded local expansion trajectories. Case A retrieves a previously uncollected paper through Yuval Peres, then reads and cites it to verify a coauthorship clue supporting the answer Russell David Lyons. Case B first shows an already collected Hirata Family archive in the map, then reads and cites it to corroborate the answer Jun, whose name was already present in the anchor.
Method
GPT-6- Luna
DeepSeek- V4.1-Flash
Qwen3.8- Flash
Retrieval-Agent
67.21 s
98.30 s
561.95 s
NeMo Agent
49.44 s
113.57 s
534.30 s
DCI
133.45 s
146.19 s
493.76 s
DR-DCI
62.04 s
164.89 s
652.18 s
RISE
59.57 s
111.83 s
419.95 s
RARG++
172.86 s
232.50 s
960.02 s
Appendix
Table 12. Mean end-to-end execution time (s) per query on BrowseComp-Plus. Bold indicates the lowest values.
Scientific agents can retrieve relevant passages yet still lose functional order, mix evidence across sources, or state conclusions that exceed the retrieved record. We introduce PathAnchor, a bounded scientific reasoning system built on path-structured evidence workspaces. Instead of treating passages or extracted concepts as independent units, the system retrieves source-linked Material-Sensor-Signal-System trajectories that preserve role, direction, and the evidence supporting each transition. A controller uses three read-only tools to search paper-specific trajectories, trace paths across candidate sources, and open exact evidence before producing a claim-cited answer and an explicit evidence boundary. On 120 single- and cross-paper flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems. Under a matched controller, corpus, and six-call budget, replacing unordered concept graphs with path-structured records raises source recall from 61.3% to 82.9%, increases answers whose claims all cite opened evidence from 69.2% to 90.0%, and reduces tool calls. These results show that evidence organization affects retrieval and citation completeness under fixed agent resources.
Qiuhui Chen, Jiafan Lu, Shuaimin Tang +5
School of Information Science and Engineering, East China University of Science and Technology · College of Intelligent Robotics and Advanced Manufacturing, Fudan University · Tencent
LLM search agents are often evaluated on final-answer accuracy, overlooking the process. Analyzing a search strategy requires understanding how credible evidence is retrieved to address question constraints. This valuable information is buried in raw search trajectories that are long and difficult to parse. We introduce SearchAtlas, a framework that converts search trajectories into structured graphs whose edges represent how evidence is propagated across the reasoning trace, from the query that retrieves it to the final answer. Our automated parsing pipeline achieves a mean edge F1 of 86.0% against human-annotated graphs and remains consistent across repeated runs. We analyze five search agents on three benchmarks, revealing systematic differences in search scale and evidence aggregation. SearchAtlas exposes fragmented answer support, question constraints that do not reach the answer, and unverified parametric knowledge entering the response. These process failures are strongly associated with incorrect answers, even more so than an LLM judge given either the raw trajectory or the ordered query list, suggesting that the constructed graphs provide useful interpretability. Moreover, an audit of cases in which process-diagnostic scores disagree with final-answer correctness shows that they capture information not reducible to answer accuracy.
Answering complex questions over large document collections requires assembling complementary evidence across sections and documents. GraphRAG offers structured retrieval but typically uses fixed traversal, while agentic RAG operates over weakly structured interfaces. Our key insight is that agents should navigate document structure within and across documents rather than repeatedly search from scratch. We introduce DocNavRAG, which organizes document hierarchies and cross-region relations into a navigable graph, exposes graph operations for locating, navigating, expanding, and fetching, and maintains an evolving evidence state to guide retrieval until sufficient evidence is collected. Across four long- and multi-document QA benchmarks, DocNavRAG improves answer quality and context sufficiency over the strongest baseline by 7.8% and 17.7% on average.
Dongyang Xie, Yao Tian, Hao Zhang +5
School of Computer Science, Wuhan University · The Hong Kong University of Science and Technology · The Chinese University of Hong Kong +1