ScholarStack: Layered Research Asset Orchestration and Cross-Task Reuse for Scientific Agents

ScholarSeed AI Team; Caoqinwei Gong; Xue Jiang; Wei Luo; Xiaoyu Qiu; Jiayi Sheng; Yi Wang; Zheng Yu

cs.AISep 20, 2026

ScholarStack: Layered Research Asset Orchestration and Cross-Task Reuse for Scientific Agents

Authors: ScholarSeed AI Team, Caoqinwei Gong, Xue Jiang, Wei Luo, Xiaoyu Qiu, Jiayi Sheng, Yi Wang, Zheng Yu, +12 more

Organizations: DAMO Academy, Alibaba Group

Abstract

Scientific agents support a range of literature-based research tasks, such as retrieval, question answering, evidence-grounded generation, and claim assessment. Most existing systems, however, are organized around individual tasks: the same papers are repeatedly retrieved, segmented, and interpreted, and the understanding built in one task is difficult to reuse in the next. We present ScholarStack, a layered research asset framework that compiles a paper collection into reusable, versioned, and provenance-preserving assets at three complementary levels: source-grounded paper-level statements, domain-level organization, and evidence-grounded cross-paper syntheses. A common access interface returns task-specific views at the evidence granularity each task requires, preserving study conditions, source traceability, and verification status. We instantiate the framework on four task families spanning ten task settings, comparing agents that use the compiled assets with task-specific baselines under matched base models. Quality gains concentrate on tasks that require cross-paper evidence, such as multi-paper question answering and literature review generation, and query-time token cost falls on every task where it is measured, with assets compiled once and reused across tasks. These results suggest that layered research assets can serve as shared infrastructure for scientific agents, shifting literature-based assistance from isolated document processing toward cumulative, evidence-grounded workflows.

Figures & tables

Research need	Reusable assets	Reuse benefits
Grounded reading and verification	L1: Source-grounded facts with study conditions, source locations, and checking states.	Reduce repeated source processing while preserving the context of reported findings.
Organization and comparison	L1 + L2: Canonical entities, categories, and relations linked to result-specific assumptions, measurements, and settings.	Reuse organizational judgments while avoiding comparisons across incompatible settings.
Cross-paper synthesis	L3 (grounded in L1 + L2): Scoped syntheses built over organized research objects and linked to source-grounded evidence.	Reuse scoped interpretations without extending conclusions beyond the examined evidence.
Inspection and revision	Across layers: Evidence dependencies, identifiers, versions, and checking states.	Preserve justification and identify what requires rechecking when evidence changes.

Table 1: Research requirements and corresponding reusable assets.

Figure 1: Overview of ScholarStack. Knowledge construction compiles documents into source-grounded facts (L1), domain organization (L2), and scoped cross-paper syntheses (L3), linked by shared identifiers and evidence references. During task execution, agents select and combine the required objects into a Knowledge View, retaining their conditions, evidence references, and recorded verification states. Different tasks use different subsets of the layers and operations. The feedback path denotes optional registration of task-derived candidates after the applicable checks; registration does not itself establish semantic verification.

Operation	Output and semantics
$\mathbf{Retrieve}(K,q)\rightarrow V$	Locate recorded facts, domain objects, or existing syntheses relevant to the request, retaining their sources and status.
$\mathbf{Select}(V,c)\rightarrow V^{\prime}$	Filter the view by explicit conditions, distinguishing confirmed matches from mismatches and missing or inapplicable information.
$\mathbf{Expand}(K,V,r)\rightarrow V^{\prime}$	Supplement the view with objects connected through the specified relations or evidence links, retaining the connections to the selected objects.
$\mathbf{Compare}(K,V,d)\rightarrow C$	Organize findings by the requested comparison dimensions, exposing shared and differing conditions and identifying unresolved comparability.
$\mathbf{Synthesize}(K,V,q)\rightarrow s$	Generate a scoped interpretation from the selected evidence, recording its supporting, opposing, and limiting basis as a candidate synthesis.

Table 2: Core operations over the knowledge layer.

K=(F,O,S)

comprises source-grounded facts, domain organization, and scoped syntheses;

q

is a retrieval or synthesis request;

V

is a selected view of typed knowledge objects, and

V^{\prime}

is the view returned by selection or expansion;

c

specifies selection constraints,

r

the relations or evidence links to follow, and

d

comparison dimensions;

C

is a Comparison Result, and

s

a candidate Scoped Synthesis. The signatures specify semantic inputs and outputs rather than a fixed execution policy.

Figure 2: Knowledge-guided agentic literature retrieval. A compiled Knowledge View informs search planning and evidence selection. Candidate papers are checked against the request using retrieved source evidence, which also supports the final ranking. Fixed-corpus and open-world settings differ in the permitted search scope.

Method	Corpus	Asset Size	R@1	R@5	R@10	MRR@40	Tokens
BM25	64,183	–	0.220	0.470	0.537	0.331	–
Dense	64,183	–	0.257	0.503	0.533	0.379	–
Hybrid	64,183	–	0.347	0.537	0.583	0.452	–
Full-text	64,183	900	0.650	0.767	0.787	0.707	49.3k
ScholarStack	64,183	900	0.670	0.773	0.773	0.717	47.9k

Table 3: Fixed-corpus literature search results on LitSearch ( Ajith et al., 2024 ) . Tokens are per-query means. Bold marks the best retrieval value among methods evaluated on LitSearch. Corpus and asset size refer to the number of papers.

	In-source				Out-of-source
Method	R@1	R@5	R@10	MRR@40	R@1	R@5	R@10	MRR@40
Full-text	0.711	0.807	0.807	0.781	0.613	0.742	0.774	0.662
ScholarStack	0.763	0.825	0.825	0.818	0.613	0.742	0.742	0.654

Table 4: Fixed-corpus literature search results by knowledge-source coverage: 19 queries with at least one gold paper among the 900 asset sources (in-source), and 31 with none (out-of-source). All gold papers are evaluated against the same retrieval corpus. Bold marks the best value within each model and coverage group, including ties.

Method	Recall	Precision	F1-Score	Macro W-Recall	Micro W-Recall
Full-text	0.124	0.138	0.128	0.117	0.113
ScholarStack	0.134	0.158	0.141	0.124	0.126

Table 5: Pooled open-world retrieval results on 89 SAGE queries. All metrics use the explicit Tier 1 set, which contains approximately five papers. Recall, Precision, F1, and Macro Weighted Recall are macro-averaged over queries; Micro Weighted Recall pools weighted labels.

Method	Weighted F1	Recall	Precision	Tokens
Full-text	0.645	0.761	0.560	35.6k
ScholarStack	0.662	0.779	0.576	21.4k

Table 6: Novelty-assessment results for

n{=}40

overlap-present targets across 12 fields and

135

gold overlaps. Weighted identification F1 uses weights of

1.0

for core axes and

0.3

for peripheral axes; token counts are per-query means.

Method	Overall	Extr.	Abst.	Bool.	Unans.	Evidence-F1	Tokens
	Answer-F1
Full-text	0.6146	0.6596	0.2697	0.6950	0.9198	0.6015	5.4k
Human	0.6090	–	–	–	–	0.7160	–
RAG	0.5613	0.5792	0.2527	0.6571	0.9091	0.5610	2.1k
ScholarStack	0.5708	0.5988	0.2631	0.6495	0.9114	0.5887	1.9k

Table 7: Single-paper QA on QASPER (416 papers / 1,428 questions). Overall denotes Answer-F1 across all questions; Extr., Abst., Bool., and Unans. denote extractive, abstractive, boolean, and unanswerable questions, respectively. Answer-F1 measures token-level answer overlap; Evidence-F1 measures paragraph-level evidence overlap. Human is the reported human-answer reference from QASPER ( Dasigi et al., 2021 ) . Tokens denotes mean input tokens per question; bold marks improvements over RAG.

Method	Ans-F1	Unans-F1	Acc.	W-F1	Macro-F1	FF	AE	Tokens
	Answerability					Generation (Rouge-L)
Full-text	0.8056	0.5071	0.7212	0.7381	0.6564	0.1548	0.2048	12.6k
Human	0.7318	–	–	–	–	0.1501	0.1605	0.3k
RAG	0.7356	0.4759	0.6485	0.6768	0.6057	0.1416	0.1564	1.9k
ScholarStack	0.7485	0.4783	0.6606	0.6874	0.6134	0.1447	0.1558	1.4k

Table 8: Single-paper QA on PeerQA (208 papers / 579 questions). Answerability is evaluated on 495 questions: Ans-F1 and Unans-F1 denote F1 for answerable and unanswerable questions; Acc., W-F1, and Macro-F1 denote accuracy, class-frequency-weighted F1, and unweighted mean class F1. Generation ROUGE-L is evaluated on 245 questions against free-form reference answers (FF) and annotated evidence (AE). Human denotes model answers using human-annotated evidence, with answerability scored only on answerable questions ( Baumgärtner et al., 2025 ) . Tokens denotes mean input tokens per question; bold marks improvements over RAG.

Method	Corr.	Comp.	Synth.	Cal.	Dir.	Total	Tokens	Sec.
Full-text	1.11	2.01	1.41	0.92	3.01	8.46	20.8k	13.13
ScholarStack	3.89	3.35	3.48	3.90	3.80	18.41	5.5k	7.54

Table 9: Multi-paper QA on 797 MDAQA questions. Corr., Comp., Synth., Cal., and Dir. denote correctness, completeness, cross-paper synthesis, calibration, and directness. Scores are from 0 to 4, and Total sums the five scores (0 to 20). Tokens and Sec. are mean tokens and generation seconds per question over available usage records, excluding knowledge construction and judging.

Method	Rel.	Cov.	Spec.	Supp.	Overall
Full-text	4.688	4.219	4.359	4.234	4.156
ScholarStack	4.578	4.250	4.453	4.438	4.203

Table 10: Quality-dimension performance on the OARelatedWork benchmark (each dimension scored out of 5). Rel., Cov., Spec., and Supp. denote Relevance, Coverage, Specificity, and Support.

Method	True $\uparrow$	Wrong Cit. $\downarrow$	Unverif. $\downarrow$	False $\downarrow$	Citation F1 $\uparrow$
Full-text	91.82	2.14	4.23	1.81	0.95
ScholarStack	92.43	1.10	5.31	1.15	0.97

Table 11: Factuality and citation performance on the OARelatedWork benchmark. Claim-label values are sample-mean percentages; citation F1 uses a 0–1 scale.

System	F1	Prec.	Rec.	Hits
GPT-Researcher	0.039	0.290	0.022	72
STORM	0.015	0.250	0.008	21
LangChain ODR	0.007	0.055	0.004	13
Lacuna Deep Research	0.052	0.339	0.028	99
Claude Code DR	0.090	0.553	0.049	171
ScholarStack	0.214	0.260	0.182	637

Table 12: ReportBench-ML citation overlap against expert-survey references.

System	Overall	Comp.	Insight	Instr.	Read.
GPT-Researcher	5.24	5.31	4.87	4.41	7.62
STORM	2.90	3.45	2.67	1.66	4.20
LangChain ODR	7.42	7.48	7.23	7.22	8.16
Lacuna Deep Research	7.82	8.01	7.61	7.57	8.34
Claude Code DR	7.54	7.04	7.43	8.28	7.88
ScholarStack	9.18	9.23	9.30	9.16	8.81

Table 13: RACE report-quality scores on ReportBench-ML. Scores are out of 10.

Method	Coverage F1	Recall	Precision	Tokens
Base LLM	0.339	0.665	0.255	18.7k
Full-text	0.398	0.665	0.329	20.7k
ScholarStack	0.404	0.642	0.331	12.4k

Table 14: Experimental-design generation results. Coverage F1 against the paper’s actual experiments (weighted by experiment kind). Tokens indicate per query.

Comparison	Wins	Losses
Full-text vs. Base LLM	32	6
ScholarStack vs. Base LLM	34	4
ScholarStack vs. Full-text	16	22

Table 15: Blind pairwise preference (GPT-6 judge; anonymized, randomized order, anti-verbosity instruction). Wins/losses/ties over

38

targets.

Method	Novelty	Significance	Clarity	Feasibility	Effectiveness	Mean
Full-text (w/o Skill)	0.486	0.457	0.343	0.257	0.443	0.397
Full-text (w/ Skill)	0.500	0.514	0.571	0.543	0.500	0.526
ScholarStack	0.514	0.529	0.586	0.700	0.557	0.577

Table 16: Comparison of idea-generation quality across five dimensions on the benchmark.

Method	Correct	Accuracy (%)	Tokens/query
Full-text	37/80	46.25	12.0k
ScholarStack	41/80	51.25	2.8k

Table 17: Scientific claim verification on

80

NLPCC instances (

20

per label). Tokens/query includes input, output, and retries, excluding knowledge construction. Bold marks the better observed value in each metric.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Object	Core fields	Example types or values
Scientific Fact	statement, context, evidence	Method description; theoretical result; limitation
Domain Taxonomy	dimensions, categories, membership criteria	Method family; assumption regime; guarantee type
Canonical Entity	name, type, aliases, source anchors	Method: SAGA; dataset: MNIST
Category Assignment	subject, dimension, category, basis, inference type, certainty	SAGA $\in$ variance reduction; system judgment
Typed Relation	source, relation, target, evidence, payload	Paper evaluated on dataset; reported metric
Scoped Synthesis	account, scope, supporting / opposing / limiting basis	Method-family overview; selected studies

Appendix

Table 18: Object fields used in the example. These are presentation fields for the representation in Section 3.2 , not a complete storage schema. All objects have an identity; verification and lifecycle states are annotations. The final column gives illustrative types and values, not exhaustive enumerations.

Dimension	Assigned category	Basis
Method family	Variance reduction	saga.method : historical gradient-table correction
Assumption regime	Convex	saga.convex-rate : component assumptions at SAGA:9
Guarantee type	Sublinear rate	saga.convex-rate : convex result at SAGA:29

Appendix

Table 19: Distinct organizational views of the same method. The assumption and guarantee rows share the convex result as their displayed basis. These labels describe that result, not every guarantee available for SAGA. The family assignment records three agreeing votes; all three displayed assignments have stored confirmed certainty, current validity, and passed evidence verification.

Role	Source fact	Contribution and source locator
Supporting	Prox-SVRG method [ Xiao and Zhang, 2014 ]	Periodic full-gradient computation; Prox-SVRG:83 states when the full gradient is computed.
Supporting	saga.method [ Defazio et al., 2014 ]	Historical gradient storage; SAGA:20 . Same fact as the L2 family-assignment basis.
Supporting	Saddle-point method [ Palaniappan and Bach, 2016 ]	Extension to saddle-point problems: Saddle:19 ; convex-concave setting: Saddle:14 .
Limiting	saga.convex-rate [ Defazio et al., 2014 ]	Specifies the guarantee in the convex setting ( SAGA:9,29 ); this upper bound does not establish that faster convergence is impossible.
Limiting	Saddle-point comparison [ Palaniappan and Bach, 2016 ]	Uniform-sampling SAGA/SVRG may be inferior to an accelerated batch method; Saddle:118 .

Appendix

Table 20: Evidence roles in the displayed synthesis. A supporting entry contributes to the account; a limiting entry qualifies its application. Locators identify sentences in each paper’s saved source map. The convex-rate entry uses the corrected basis at SAGA:29 .

Stage	Corpus searches	Title lookups	Asset lookups	Abstract inspections	Full-text reads
0: Initial planning	3	2	1	0	0
1: Expansion and checks	3	2	1	6	4
2: Final verification	2	2	0	4	2
3: Ranking	0	0	0	0	0

Appendix

Table 21: Maximum numbers of tool calls per stage. Asset lookups are available to Full-text and ScholarStack but not to the no-asset control. All allocations are subject to the cumulative limits in Table 22 .

Resource	Limit
Model calls / candidate-context size	4 / 180 papers
Corpus searches / title lookups	8 (20 papers per search) / 6
Metadata resolutions / abstract inspections	10 / 10
Asset lookups and returned content	2 lookups; at most 3 records per lookup; 8,000 characters in total
Full-text reads	6 calls covering at most 6 distinct papers
Full-text response length	4,000 characters per call; 6,000 per paper; 24,000 in total

Appendix

Table 22: Execution and response limits per query. Byte limits are measured in UTF-8; character and token limits are enforced separately.

Method	Corpus	Asset Size	R1	R5	R10	MRR40	Tokens
Full-text	64,183	900	0.670	0.773	0.773	0.721	41.1k
ScholarStack	64,183	900	0.710	0.803	0.803	0.763	39.5k

Appendix

Table 23: Fixed-corpus retrieval results with GPT-5.6 Terra. Source papers denotes the number of papers used to construct the auxiliary assets. Runtime and token usage are averaged per query; GPT token counts are reported-usage lower bounds. Bold indicates the best value in each retrieval-metric column.

	In-source				Out-of-source
Method	R1	R5	R10	MRR40	R1	R5	R10	MRR40
Codex raw (w/o full-text)	0.605	0.719	0.719	0.684	0.677	0.774	0.774	0.720
Full-text	0.711	0.772	0.772	0.763	0.645	0.774	0.774	0.695
ScholarStack	0.711	0.798	0.798	0.781	0.710	0.807	0.807	0.753

Appendix

Table 24: Fixed-corpus retrieval results by knowledge-source coverage with GPT-5.6 Terra. Bold indicates the best value in each column, including ties.

Domain	SAGE queries	Source papers
Computer science	29	1,754
Natural science	28	1,749
Healthcare	32	1,748

Appendix

Table 25: Frozen source collections for open-world retrieval. "Source papers" denotes the number of full-text documents underlying both the Full-text and ScholarStack conditions.

Field	Count	Field	Count	Field	Count
Linguistics	5	Physics	4	Education	2
Medicine	5	Psychology	4	Art	1
Engineering	5	Business	4	Chemistry	1
Mathematics	4	Biology	4	Materials Sci.	1

Appendix

Table 26: Per-field breakdown of the

40

overlap-present novelty targets (12 fields).

	Answerability					Generation (Rouge-L)
Method	Ans-F1	Unans-F1	Acc.	W-F1	Macro-F1	FF	AE	Tokens
Full-text	0.806	0.507	0.721	0.738	0.656	0.155	0.205	12,577
Human (reference)	0.732	–	–	–	–	0.150	0.161	301
RAG10	0.736	0.476	0.649	0.677	0.606	0.142	0.156	1,864
RAG20	0.766	0.489	0.679	0.703	0.627	0.149	0.184	3,481
ScholarStack	0.749	0.478	0.661	0.687	0.613	0.145	0.156	1,350

Appendix

Table 27: All evaluated conditions on PeerQA (208 papers / 579 questions). RAG@

k

retrieves the top-

k

passages. Answerability is evaluated on 495 questions: Ans-F1 and Unans-F1 denote F1 for answerable and unanswerable questions; Acc., W-F1, and Macro-F1 denote accuracy, class-frequency-weighted F1, and unweighted mean class F1. Generation ROUGE-L is evaluated on 245 questions against free-form reference answers (FF) and annotated evidence (AE). Human denotes model answers using human-annotated evidence, with answerability scored only on answerable questions. Tokens denotes mean input tokens per question; bold marks ScholarStack’s improvements over RAG10.

Machine learning (313 answerability / 158 generation)
Method	Ans-F1	Unans-F1	Acc.	W-F1	Macro-F1	RgL $*{\mathrm{FF}}$	RgL $*{\mathrm{AE}}$
Full-text	0.808	0.517	0.725	0.747	0.662	0.142	0.199
RAG	0.722	0.460	0.633	0.666	0.591	0.130	0.149
ScholarStack	0.729	0.459	0.639	0.672	0.594	0.128	0.141
NLP (133 / 69)
Full-text	0.822	0.593	0.752	0.761	0.707	0.194	0.235

Appendix

Table 28: Per-domain results on PeerQA. Parentheses give the numbers of answerability / generation questions. Ans-F1 and Unans-F1 denote F1 for answerable and unanswerable questions; Acc., W-F1, and Macro-F1 denote accuracy, class-frequency-weighted F1, and unweighted mean class F1. RgL

*{\mathrm{FF}}

and RgL

*{\mathrm{AE}}

denote generation ROUGE-L against free-form reference answers and annotated evidence, respectively. Bold marks the best Macro-F1 in each domain.

Question	Gold	Full-text	RAG	ScholarStack
QASPER
What inference models are used?	XGBoost, AdaBoost, Random Forest	XGBoost, AdaBoost and Random Forest	Gaussian Processes	XGBoost, AdaBoost, Random Forest
How many actions are present in the dataset?	14,769	14,769	No Answer	14,769
PeerQA
Does the optimization happen only over map locations observed so far?	No; it also covers unobserved (hallucinated) regions	No; any region observed or hallucinated	No Answer	No; goals from any observed or hallucinated region
What values are represented in Table 2?	Zero-shot classification performance across 5 datasets and 14 language models	Zero-shot classification performance for all models and datasets	No Answer	Performance statistics for all models and datasets

Appendix

Table 29: Selected qualitative examples on QASPER and PeerQA. Gold denotes the reference answer. Reference answers and model responses are abridged for presentation; answer completeness varies. Aggregate quality and input usage are reported in Tables 7 and 8 .

Method	Corr.	Comp.	Synth.	Cal.	Dir.	Total	Tokens	Sec.
Full-text + Bare	1.11	2.01	1.41	0.92	3.01	8.46	20.8k	13.13
ScholarStack (L1) + Bare	3.90	3.56	2.99	3.78	3.82	18.04	4.1k	9.83
ScholarStack (L1) + Skill	3.91	3.60	3.03	3.82	3.83	18.18	4.2k	10.45
ScholarStack (L1+L2) + Bare	3.82	3.25	2.80	3.26	3.88	17.01	4.8k	7.20
ScholarStack (L1+L2) + Skill	3.89	3.53	3.02	3.78	3.84	18.05	4.9k	10.67
ScholarStack (L1+L2+L3) + Bare	3.89	3.35	3.48	3.90	3.80	18.41	5.5k	7.54

Appendix

Table 30: Full ablation and generation-cost results on the 797-question MDAQA bank. Bare denotes the answering model using the ordinary prompt and supplied evidence, without the additional skill instructions. Skill denotes Cross-Paper Evidence Synthesis skill, which guides comparison of study conditions, synthesis of findings, and attribution of claims to evidence. Corr., Comp., Synth., Cal., and Dir. denote correctness, completeness, cross-paper synthesis, calibration, and directness. Scores are weighted means on a 0–4 scale, except Total (0–20). Generation-cost columns use the available per-question usage audit. Construction and judging are excluded. Bold marks the best value in each column, where higher is better for the six score columns and lower is better for the two cost columns.

System	Input(k)	Output(k)	Cache hit(%)	Cost(USD)
Claude Code DR	3,128	106	78.94	7.18
ScholarStack	10,533	117	96.02	3.49

Appendix

Table 31: Per-report resource use and standardized cost for the two locally evaluated review-generation workflows. k denotes one thousand tokens.

Field	Count	Field	Count	Field	Count
Medicine	5	Physics	4	Education	2
Engineering	5	Business	4	Art	1
Linguistics	4	Biology	4	Chemistry	1
Mathematics	4	Psychology	3	Materials Sci.	1

Appendix

Table 32: Per-field breakdown of the

38

experimental-design targets (12 fields).

Domain	Count
AI and engineering	14
Chemistry and materials	12
Life sciences and medicine	4
Public health and social sciences	3
Environmental science	2
Total	35

Appendix

Table 33: Domain composition of the idea-generation benchmark.

Dimension	Definition
Novelty	Are the problems or approaches new? Is this a novel combination of familiar techniques? Is it clear how this work differs from previous contributions? Is related work adequately referenced?
Significance	Is the idea important? Are other people (practitioners or researchers) likely to use these ideas or build on them? Does the idea address a difficult problem in a better way than previous research? Does it provide a unique theoretical or pragmatic approach?
Clarity	Is the paper clearly written? Is it well-organized? Does it adequately inform the reader?
Feasibility	Can the idea be realized with existing technology or methods? Are there any technical difficulties or bottlenecks? Is the idea clear and logical? Are there any obvious errors or unreasonable parts in the idea, and can the experiments be designed normally according to this idea.
Effectiveness	How likely the proposed idea is going to work well (e.g., better than existing baselines).

Appendix

Table 34: Definitions of evaluation dimensions from Li et al. [2024] .

Explore similar work

Agents-K1: Towards Agent-native Knowledge Orchestration
Jun 11, 2026Zongsheng Cao, Bihao Zhan, Jinxin Shi +23Knowledge Graphs Scientific Discovery
Shanghai Artificial Intelligence Laboratory · East China Normal University · Fudan University
The Last Human-Written Paper: Agent-Native Research Artifacts
Apr 27, 2026Jiachen Liu, Jiaxin Pei, Jintao Huang +34Research Automation Narratives
University of Michigan · Stanford University · Ohio State University +22
AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery
Apr 28, 2026Lei Xiong, Kun Luo, Ziyi Xia +15Research Automation Deep Research
AutoResearchBench Team*

Jun 11, 2026, cs.AIJ/K move · Enter open · S save

Agents-K1: Towards Agent-native Knowledge Orchestration

Zongsheng Cao, Bihao Zhan, Jinxin Shi, Jiong Wang+22

Shanghai Artificial Intelligence Laboratory · East China Normal University · Fudan University

Knowledge Graphs Scientific Discovery Agent2Agent

Open paper AskPDF arXiv

Current LLM-based research agents have advanced through agent orchestration, yet largely overlook scientific knowledge orchestration. Existing works often reduce papers to abstracts, surface mentions, and flat \texttt{cites} edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning. To this end, we introduce \textbf{Agents-K1}, an end-to-end knowledge orchestration pipeline that converts raw documents into agent-native scientific knowledge graphs. Agents-K1 integrates three components under a unifying theoretical foundation: a multimodal parser whose five-module schema captures entities, multimodal evidence, citations, and typed inter-entity relations across the full paper rather than abstracts alone; a 4B information-extraction backbone trained with GRPO under a rule-based reward; and a graphanything CLI, a tri-source agent interface that unifies web search, multimodal graph retrieval, and cross-document traversal. On top of this, we process 2.46 million scientific papers across six subjects to produce \textbf{Scholar-KG}, of which we release a one-million-paper subset, and the full Scholar-KG is accessible via the SCP link below. The same pipeline can be extended to general-domain corpora and to schema-conformant data synthesis. Extensive experiments demonstrate that Agents-K1 achieves superior performance in scientific information extraction, knowledge graph construction, and multi-hop scientific reasoning.

Abstract

Explore similar work

Agents-K1: Towards Agent-native Knowledge Orchestration

The Last Human-Written Paper: Agent-Native Research Artifacts

AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery