LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge

Shi-Ju Ran; Kun Zhang; Xi Wu; Liu-Si Yang; Wen-Jun Li

cs.AIAug 30, 2026

LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge

Authors: Shi-Ju Ran, Kun Zhang, Xi Wu, Liu-Si Yang, Wen-Jun Li

Organizations: Center for Quantum Physics and Intelligent Sciences, Department of Physics, Capital Normal University, Beijing 100048, China · College of Artificial Intelligence, Putian University, Putian, Fujian 351100, China

Abstract

Sustained scientific work requires a knowledge substrate that carries interpretation across tasks and preserves paths to source evidence. We call this process scientific knowledge compilation and implement it in ASKS, the Agent-Driven Scientific Knowledge System. For each source, an LLM produces a readable Wiki view and machine-facing semantics. Deterministic checks convert the latter into a document-local GraphDelta, and embedding geometry together with explicit graph rules integrates the proposed changes into persistent state. Each ingest is an inspectable state transition over accumulated knowledge, with compiled Wiki and graph views linked to the preserved source record. We examine this process by chronologically compiling 56 published papers from one research program. Branch survival, cross-paper support, lineage, coverage, and churn yield a source-traceable author research portrait centered on tensor-network methods, with branches into quantum many-body research, tensor-network machine learning, and quantum-AI-oriented directions. In this run, higher-level Hub organization remains stable and low-churn. Canonical-node growth is predominantly additive. Graph-level measurements and navigation paths retain links to the source records from which they were compiled.

Figures & tables

Figure 1: Scientific knowledge compilation in ASKS. Each received scientific source $D_{t}$ remains preserved while an LLM produces two complementary source-local representations: a readable Wiki view $W_{t}$ and machine-facing semantic slots $S_{t}$ . Validated semantics together with deterministic source relations form an inspectable document-local GraphDelta $\Delta_{t}$ . Embedding geometry and explicit graph rules integrate the delta with accumulated state $G_{t}$ to produce $G_{t+1}$ . Repeating this transition turns local interpretation into an evolving research map. The persistent Wiki and graph surfaces guide navigation, while scientific claims resolve to source-addressed evidence.

Layer	Principal role	Produced representation or signal	Scientific authority
Source record	Preserve the received scientific record and stable addressing.	Original files, metadata, and machine-readable derivatives.	Primary evidence for what the received record states.
LLM	Interpret one source and express local meaning.	Readable Wiki content and validated semantic slots.	Source-local interpretation under validation and review.
Embeddings	Place local and accumulated knowledge in a shared semantic geometry.	Similarity signals used by identity, routing, membership, and lifecycle gates.	Semantic proximity evaluated within coded gates.
Graph	Retain source-linked relations, Hub organization, and construction history.	A persistent, revisable navigation structure.	Source-linked navigation whose evidence resolves to source records.

Table 1: Division of labor and authority in ASKS. The four layers cooperate in one compiler and carry distinct scientific responsibilities.

Figure 2: Chronological compilation trajectory in the 56-paper demonstration. (a) Canonical reuse

R(t)

and multi-source consolidation

M(t)

across chronological ingests. (b) Old-node Hub-membership churn

C(t)

shown on its own scale because its amplitude is much smaller. Light marker-lines are raw per-paper values, whereas dark curves are five-paper trailing means used as visual guides. The diagnostics show low churn throughout and slightly declining descriptive trends in reuse and consolidation.

Figure 3: Hub births, growth, and lineage during chronological compilation. Ribbon width encodes active Hub membership over graph state

G_{t}

. Filled and open circles mark multi-work and single-work births, dotted links mark committed parent–child lineage, and right-side labels report final member count at

G_{56}

. The lifecycle consists of 18 births. Split, merge, and retirement counts are each zero.

Figure 4: A source-traceable author research portrait compiled from 56 publications. (a) Topology-focused view of the final research tree. Vertical guides indicate Hub birth year, with small horizontal offsets used for readability. Colored anchors and light dotted connectors serve as display-level macro-domain aids. Solid black arrows show committed parent–child lineage, and node size encodes the final number of distinct supporting papers. The label msQFM denotes multi-spin quantum feature mapping. (b) Hub births by year and macro-domain together with the cumulative number of active Hubs. The strongest branching phase occurs in 2019–2021, while later births extend the TN core in 2023, TN-ML in 2024, and quantum-AI in 2026. The deterministic paper-to-Hub navigation mapping covers 50 of 56 papers and records six explicit abstentions. Hub birth, membership, lineage, and supporting-paper counts come from the frozen graph record. Macro-domain labels are post-hoc display interpretations, not generated Hub nodes, and do not enter the quantitative analysis.

Descriptor	Value	Reading
Survival $P_{\mathrm{survive}}$	$1.000$	All born branches persist to $G_{56}$ .
Multi-work birth $B_{\mathrm{multi}}$	$0.889$	Most births are already cross-paper.
Lineage branching $L_{\mathrm{branch}}$	$0.222$	Some births refine earlier Hubs.
Lifecycle asymmetry $A_{\mathrm{life}}$	$1.000$	Observed growth is purely additive.
Mapping coverage $Q_{\mathrm{map}}$	$0.893$	50 of 56 papers map into the final tree.
Mean old-node churn $\bar{C}$	$0.00467$	Membership is low-churn under the frozen configuration.

Table 2: Graph descriptors of the chronological research portrait. Definitions are given in Eqs. (6)–(7).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S1 : Direct-contribution sensitivity for the 2024 density-outlier paper (D047). Solid curves show the observed frozen trajectory. Dashed curves remove D047’s direct attachment, support, or created-old-node contribution under the fixed formal outcomes. The upper panel reports cumulative reuse

R_{\mathrm{cum}}(t)

. The comparison is deterministic and preserves all downstream fusion decisions.

Process	Frozen signal and gate	Compilation effect
Node identity	lexical identity evidence, label similarity $\geq 0.92$ , semantic similarity $\geq 0.90$ , combined gate $\geq 0.91$ , winner margin $\geq 0.04$ , candidate floor $0.72$	reuse, abstain, or create a local node
Proposition alignment	cosine $>0.98$ for reuse, relation band $0.90$ – $0.98$	merge, relate, or remain distinct
Paper routing	positioning text versus active Hub scope, routing floor $0.50$ and margin $0.04$	paper-to-Hub navigation relation
Hub membership	enter $0.66$ , retain $0.58$ , at most three Hubs, child preference $+0.05$	rebuild membership with hysteresis and specificity bias
Hub overload	active membership count $>20$	redistribute toward children or trigger split analysis
Hub birth	pairwise similarity $\geq 0.82$ , at least four members, mean cohesion $\geq 0.60$	create a hash-bound candidate for explicit review

Appendix

Table 3: Frozen organization parameters used in the chronological demonstration.

Mechanism	Implemented examples	Registration and discovery	Invocation and operational role
Work state	research : project-scoped research context with status and research memory	Task route and the research engineering capability package	--task research restores project context and carries evidence, decisions, experiments, and manuscript state across the sustained activity
Work state	Staged query loop and transactional ingest state machine	Task route, stage cards, transaction checkpoints, and resumable receipts	--query-stage selects start , evidence , continue , or answer . Ingest stages advance after validation and checkpoint completion
Capability	write/academic : shared writing conventions plus the physics-manuscript overlay	write capability profile with a discipline-specific overlay	--capability write --capability-profile academic loads claim-strength, scope, notation, and manuscript-QA conventions while preserving the active research state
Tool	lookup , relations , read-raw , recall , and remember	DSH ToolRegistry and the unified wg.py command surface	Structured arguments and results execute graph navigation, source-addressed reading, and project-memory operations under evidence guards

Appendix

Table 4: Representative implemented work states, composable capabilities, and executable tools in the ASKS scientific knowledge system. The examples cover a selected set of major implemented mechanisms built on the source-linked knowledge substrate.

Run ID	Year	Title	DOI
D001	2010	Phase transitions and thermodynamics of the two-dimensional Ising model on a distorted kagome lattice	doi:10.1103/PhysRevB.82.134434
D002	2011	Linearized Tensor Renormalization Group Algorithm for the Calculation of Thermodynamic Properties of Quantum Lattice Models	doi:10.1103/PhysRevLett.106.127202
D003	2012	Phase diagrams, distinct conformal anomalies, and thermodynamics of spin-1 bond-alternating Heisenberg antiferromagnetic chain in magnetic fields	doi:10.1103/PhysRevB.85.134425
D004	2012	Optimized decimation of tensor networks with super-orthogonalization for two-dimensional quantum lattice models	doi:10.1103/PhysRevB.86.134429
D005	2013	Honeycomb Heisenberg spin ladder: Unusual ground state and thermodynamic properties	doi:10.1209/0295-5075/104/57009
D006	2013	Theory of network contractor dynamics for exploring thermodynamic properties of two-dimensional quantum lattice models	doi:10.1103/PhysRevB.88.064407

Appendix

Table 5: Frozen 56-work corpus in the chronological order defined by the publication manifest.

KG Construction

Knowledge graph construction turns text, tables, images, and databases into structured entities and relationships that machines can query and reason over.

See all 195 papers

Agents-K1: Towards Agent-native Knowledge Orchestration
Jun 11, 2026Zongsheng Cao, Bihao Zhan, Jinxin Shi +23KG Construction AI for Science
Shanghai Artificial Intelligence Laboratory · East China Normal University · Fudan University
Large Knowledge Model: A Knowledge Foundation for Agentic Science at Scale
Sep 23, 2026Yuan Huang, Sihan Hu, Hongyu Gu +18KG Construction AI Agents for Scientific Discovery
DP Technology, Beijing, China · Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing, China · International Center for Quantum Materials, School of Physics, Peking University, Beijing, China +4
ScholarStack: Layered Research Asset Orchestration and Cross-Task Reuse for Scientific Agents
Sep 20, 2026ScholarSeed AI Team, Caoqinwei Gong, Xue Jiang +17AI Agents for Scientific Discovery Evidence-Grounded Generation
DAMO Academy, Alibaba Group

Jun 11, 2026J/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

KG Construction AI for Science Document Information Extraction Knowledge Representation LLM Agents

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 cites edges, omitting key entities, claims, evidence, mechanisms, and method lineages essential for scientific reasoning. To this end, we introduce 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 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

KG Construction

Explore similar work

Agents-K1: Towards Agent-native Knowledge Orchestration

Large Knowledge Model: A Knowledge Foundation for Agentic Science at Scale

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