Structurally-bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch

Rima Hazra; Sayan Layek; Somnath Banerjee; Soumen Chakrabarti; Animesh Mukherjee

cs.CLAug 25, 2026

Structurally-bounded Agentic Graph Exploration for Evidence-Grounded Scholarly DeepSearch

Authors: Rima Hazra, Sayan Layek, Somnath Banerjee, Soumen Chakrabarti, Animesh Mukherjee

Organizations: National University of Singapore · TCG CREST · Indian Institute of Technology Kharagpur · Singapore Institute of Technology · Indian Institute of Technology Bombay

Abstract

We present Crase, a bounded and inspectable alternative to deep research agents for scholarly search. Instead of an open-ended search loop, Crase queries a search engine once for seed papers, expands them along their 1.5-hop citation neighborhood, prunes citation edges whose claims lack entailment support, and ranks the remaining papers with a recency-aware random walk. This makes the candidate set, the reason each paper is kept, and the stopping condition explicit and fixed before inference. On LitSearch and one further benchmarks over a 500K-paper arXiv corpus, Crase outperforms deep research agents built on proprietary models by up to 3 $\times$ recall@50 at roughly a third of the cost.

Figures & tables

Figure 1: Unlike deep research agents that repeatedly search until the model decides to stop, Crase performs a seeded search, constructs a 1.5-hop citation graph, prunes unsupported edges, and ranks papers with a recency-aware walk.

Dataset	Method	Recall@K						MAP@K
Dataset	Method	@5	@10	@20	@30	@40	@50	@5	@10	@20	@30	@40	@50
ACL	DeepResearch-GPT	0.0335	0.0696	0.1289	0.1572	0.1881	0.2294	0.0107	0.0170	0.0234	0.0253	0.0267	0.0279
	DeepResearch-Claude	0.0309	0.0696	0.1160	0.1366	0.1804	0.2294	0.0085	0.0150	0.0196	0.0203	0.0223	0.0242
	Crase -Llama-PPR	0.0730	0.1236	0.1882	0.2584	0.3258	0.3848	0.0453	0.0530	0.0602	0.0656	0.0697	0.0730
	Spector2-Deepwalk	0.0284	0.0747	0.1263	0.1830	0.2242	0.2629	0.0170	0.0248	0.0303	0.0341	0.0366	0.0380
ICLR	DeepResearch-GPT	0.0183	0.0427	0.0610	0.0732	0.1098	0.1220	0.0082	0.0106	0.0119	0.0123	0.0150	0.0155

Table 1: Main results. Recall@K and MAP@K across the three datasets and retrieval setups. Best results bold-colored; second-best underlined. We observe that Crase -Llama-PPR consistently performs best on ACL and ICLR and remains competitive or best across most settings on arXiv .

Dataset	Variant	Recall@K						MAP@K
Dataset	Variant	@5	@10	@20	@30	@40	@50	@5	@10	@20	@30	@40	@50
ACL	Crase -Qwen-SALSA	0.0335	0.0747	0.1263	0.1675	0.2113	0.2552	0.0200	0.0258	0.0289	0.0317	0.0336	0.0359
	Recency	0.0335	0.0619	0.1521	0.2165	0.2655	0.3015	0.0160	0.0209	0.0281	0.0333	0.0360	0.0376
	Crase -Qwen-PPR	0.0309	0.0722	0.1572	0.2191	0.2706	0.3144	0.0148	0.0212	0.0284	0.0336	0.0362	0.0382
	Crase -Llama-SALSA	0.0393	0.0787	0.1433	0.1938	0.2584	0.3146	0.0177	0.0269	0.0332	0.0359	0.0398	0.0428
ICLR	Crase -Qwen-SALSA	0.0549	0.0793	0.1280	0.1768	0.2073	0.2195	0.0197	0.0234	0.0280	0.0314	0.0325	0.0329

Table 2: Ablations. Alternative claim extraction models ( Qwen / Llama ) and ranking strategies ( SALSA / PPR ) and the Recency baseline. Best per column is bold-colored; second-best is underlined. We observe that Crase -Llama-SALSA gives the strongest overall performance, while Crase -Qwen-PPR is competitive in Recall@K.

		Recall@ $\bm{K}$						MAP@ $\bm{K}$
Dataset	Seeds	5	10	20	30	40	50	5	10	20	30	40	50
	Query yr.	0.0295	0.0516	0.1204	0.1622	0.2187	0.2654	0.0117	0.0158	0.0223	0.0253	0.0283	0.0311
ACL	Current yr.	0.0309	0.0722	0.1572	0.2191	0.2706	0.3144	0.0148	0.0212	0.0284	0.0336	0.0362	0.0382
	$\Delta$	$+.0014$	$+.0206$	$+.0368$	$+.0569$	$+.0519$	$+.0490$	$+.0031$	$+.0054$	$+.0061$	$+.0083$	$+.0079$	$+.0071$
	Query yr.	0.0606	0.1138	0.2036	0.2216	0.2515	0.2994	0.0278	0.0338	0.0441	0.0454	0.0463	0.0483
ICLR	Current yr.	0.0488	0.1220	0.2195	0.2561	0.3171	0.3659	0.0245	0.0372	0.0497	0.0531	0.0561	0.0580

Table 3: Seed selection ablation. Query-year vs. current-year seeds (shaded; used in main experiments); bold marks the best result and

\Delta

the absolute change. Current-year seeds improve nearly all Recall@K and MAP@K scores, with the only drop at Recall@5 for ICLR.

Method	Calls $\downarrow$	Tokens $\downarrow$	Time $\downarrow$	Cost $\downarrow$	R@50 $\uparrow$
		(K)	(s)	($)
DeepResearch-GPT	18	620	272	1.76	0.1220
DeepResearch-Claude	17	560	249	2.06	0.1220
Crase -Llama-PPR	5	235	104	0.37	0.3659

Table 4: Retrieval utility and interaction cost on ICLR. We report median per-query resource consumption together with recall@

50

. Crase achieves substantially higher retrieval recall while requiring fewer external interactions and lower inference cost.

\downarrow

indicates lower is better and

\uparrow

indicates higher is better.

Measure	Score
Crase –human agreement	84.0%
Retained-edge precision	88.0%
Pruned-edge correctness	80.0%
Unanimous human agreement	68.0%
Fleiss’ $\kappa$	0.71

Table 5: Human validation of evidence-graph decisions. Three domain experts independently assess 50 citation edges while being blinded to Crase ’s decision. Crase ’s retain/prune decisions are compared against the majority human judgment.

Model	MSciNLI	SciNLI
Qwen2.5-1.5B-Instruct	0.75	0.77
Qwen2.5-3B-Instruct	0.80	0.83

Table 6: F1 score on claim entailment. Best per dataset in bold . Qwen2.5-3B consistently outperforms Qwen2.5-1.5B on both datasets.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Category	#	Category	#
cs.LG	135,116	cs.NE	6,972
cs.CV	150,055	cs.MA	3,890
cs.AI	36,148	stat.ML	18,549
cs.CL	82,273
cs.IR	14,081

Appendix

Table 7: Per-category paper counts in

\mathcal{C}

. Papers with multiple tags are counted once per category.

Setting	Value
Base model	Qwen2.5-3B-Instruct
Training stage	Supervised fine-tuning (SFT)
Fine-tuning type	Full parameters
Optimization backend	DeepSpeed ZeRO-3
Prompt template	qwen
Max sequence length	2048

Appendix

Table 8: Fine-tuning configuration for the Crase entailment model.

Graph	R@50	$\mathbf{\|V\|}$	Entailment calls
1-hop	0.331	642	1,286
1.5-hop	0.372	642	3,947
2-hop	0.386	4,781	28,614

Appendix

Table 9: Effect of exploration radius on ICLR. The 1-hop and 1.5-hop constructions contain the same nodes; 1.5-hop additionally retains relations among retrieved neighbors.

$\bm{m}$	R@50	Median $\|V\|$	Relative cost
5	0.326	371	$0.61\times$
10	0.372	655	$1.00\times$
15	0.381	911	$1.42\times$
20	0.386	1,174	$1.83\times$

Appendix

Table 10: Effect of the seed budget on ICLR.

ID	Type	Extracted claim	Source phrase
Paper 1: A Review of Cooperation in Multi-agent Learning
C1	CONTEXTUAL	Cooperation in multi-agent learning intersects with game theory, economics, social sciences, and evolutionary biology.	“Cooperation in multi-agent learning (MAL) is a topic at the intersection of numerous disciplines”
C2	CONTEXTUAL	Research in multi-agent learning aims to understand how agents can coordinate effectively when goals are aligned.	“Research in this area aims to understand both how agents can coordinate effectively when goals are aligned”
C3	CONTEXTUAL	Research in multi-agent learning also aims to understand how agents may cooperate in settings with potential for conflict.	“how they may cooperate in settings where gains from working together are possible but possibilities for conflict abound”
C4	METHODOLOGICAL	The paper provides an overview of fundamental concepts, problem settings, and algorithms in multi-agent learning.	“we provide an overview of the fundamental concepts, problem settings and algorithms of multi-agent learning”
C5	CONTEXTUAL	The paper discusses open challenges in multi-agent learning to inspire new research avenues.	“Finally we discuss open challenges in the field with the aim of inspiring new avenues for research”

Appendix

Table 11: Outputs of the claim-extraction stage. Each paper is represented by a small set of atomic, typed claims together with the source phrase from which the claim was extracted. These source anchors make the intermediate representation directly inspectable before claim-level entailment is computed.

Figure 2: Recall@

K

(left) and MAP@

K

(right) under seed corruption. Lexically similar distractors degrade retrieval gracefully— Crase retains

\sim\!56\%

of clean Recall@50 and

\sim\!81\%

of clean MAP@50, because claim-level pruning suppresses the unsupported edges the distractor introduces. Random irrelevant seeds collapse to zero: an off-manifold seed induces a neighborhood disconnected from the ground truth, which Crase does not attempt to recover through further search.

Stage	Model contribution	Structural control	What can be inspected?
Planning	Generates focused sub-queries	Exactly $L$ external search calls	Research plan and retrieved seeds
Graph construction	Extracts atomic claims	Candidate graph fixed after seed construction	Nodes, citation edges, extracted claims
Evidence scoring	Predicts local claim support	Each citation edge scored independently	Entailment matrix and $\Delta(u,v)$
Pruning	Provides support scores	Fixed threshold determines edge fate	Why each edge is retained or removed
Ranking	No new evidence is generated	Deterministic PPR/SALSA over the weighted graph	Edge weights and final ranking scores
Stopping	None	Graph scoring completes and ranking converges	Explicit stopping condition

Appendix

Table 12: Where does model autonomy remain in Crase ? Model judgments are restricted to local, inspectable operations, while exploration, pruning, ranking, and termination remain structurally constrained after seed construction.

Seed paper	Fate
$\rho_{1}$ “faithful explicit helpfulness measurement”
A resource theory of quantum memories and their faithful verification with minimal assumptions	✗
Pretty good measurement for bosonic Gaussian ensembles	✗
Topology-faithful nonparametric estimation and tracking of bulk interface networks	✗
$\rho_{2}$ “human explanations language models”
Sparse Autoencoders Find Highly Interpretable Features in Language Models	✓

Appendix

Table 13: The seed set is noisy by design,and it does not matter. The complete seed set

S

(

|S|=29

) for the query of Trace 1, grouped by contributing sub-query

\rho_{\ell}

. ✓ = seed survives Stage 2 claim-level pruning and enters the ranking walk; ✗ = every incident edge receives coverage affinity

\Delta=0

and the seed is removed from the evidence graph. Keyword search over a 500K-paper corpus inevitably retrieves lexical-match artifacts (quantum measurement, galaxy catalogs, retail image classification); in a search-and-synthesis loop these would enter the context of every subsequent step, whereas in Crase the evidence graph , not the search engine , enforces relevance.

Explore similar work

Argus: Evidence Assembly for Scalable Deep Research Agents
May 15, 2026Zhen Zhang, Liangcai Su, Zhuo Chen +7Multi-Agent LLM Systems Tool-Augmented Language Model Agents
MiroMind AI
Deep Research Pretraining via Predictive Navigation
Aug 1, 2026Jiang Zhou, Zhiyuan Fan, Xing Wu +3Deep Research Agents LLM Agent Training
Hunyuan Team, Tencent
ScholarQuest: A Taxonomy-Guided Benchmark for Agentic Academic Paper Search in Open Literature Environments
Jun 18, 2026Tingyue Pan, Mingyue Cheng, Daoyu Wang +4Agentic Search Information Retrieval
State Key Lab of Cognitive Intelligence, University of Science and Technology of China

May 15, 2026J/K move · Enter open · S save

Argus: Evidence Assembly for Scalable Deep Research Agents

Zhen Zhang, Liangcai Su, Zhuo Chen, Xiang Lin+6

MiroMind AI

Multi-Agent LLM Systems Tool-Augmented Language Model Agents Deep Research Agents Agentic Search Agentic Reasoning

Open paper AskPDF arXiv

Deep research agents have achieved remarkable progress on complex information seeking tasks. Even long ReAct style rollouts explore only a single trajectory, while recent state of the art systems scale inference time compute via parallel search and aggregation. Yet deep research answers are composed of complementary pieces of evidence, which parallel rollouts often duplicate rather than complete, yielding diminishing returns while pushing the aggregation context toward the model's limit. We propose Argus, an agentic system in which a Searcher and a Navigator cooperate to treat deep research as assembling a jigsaw from complementary evidence pieces, rather than brute forcing the whole answer in parallel. The Searcher collects evidence traces for a given sub-query through ReAct-style interaction. The Navigator maintains a shared evidence graph, verifying which pieces are still missing, dispatching Searchers to gather them, and reasoning over the completed graph to produce a source-traced final answer. We train the Navigator with reinforcement learning to verify, dispatch, and synthesize, while independently training the Searcher to remain a standard ReAct agent. The resulting Navigator supports rollouts with a single Searcher or many in parallel without retraining. With both Searcher and Navigator built on a 35B-A3B MoE backbone, Argus gains 5.5 points with a single Searcher and 12.7 points with 8 parallel Searchers, averaged over eight benchmarks. With 64 Searchers it reaches 86.2 on BrowseComp, surpassing every proprietary agent we benchmark, while the Navigator's reasoning context stays under 21.5K tokens.

Abstract

Explore similar work

Argus: Evidence Assembly for Scalable Deep Research Agents

Deep Research Pretraining via Predictive Navigation

ScholarQuest: A Taxonomy-Guided Benchmark for Agentic Academic Paper Search in Open Literature Environments