Source-linked AI summary
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
TL;DR
Scientific work needs a persistent substrate that carries interpretations across tasks while preserving links to source evidence. ASKS addresses this gap by compiling each source into validated Wiki and graph representations and evaluating the resulting trajectory across 56 papers. The demonstration yields a source-traceable research portrait with stable, low-churn Hub organization and predominantly additive canonical-node growth.
Problem
Scientific knowledge compilation addresses the need for later research tasks to inherit persistent, source-traceable interpretations rather than only transient retrieved context.
Method
ASKS uses source-local LLM encoding, validated document-local GraphDeltas, embedding-based organization, and deterministic graph rules to update persistent state.
Results
The 56-paper chronological demonstration produces a source-traceable author research portrait centered on tensor-network methods, with later application branches and stable higher-level organization.
Takeaways & Limitations
The compiled portrait supports navigation through graph-level organization while retaining links from compiled views and paths to preserved scientific records.
Takeaways & Limitations
The worked example evaluates formation and persistence only along one frozen chronological trajectory, with order robustness, causal decomposition, and domain interpretation evaluated separately.
Abstract
from arXiv · showhide
Sustained scientific work requires a knowledge substrate that carries interpretation across tasks and preserves paths to source evidence. We call this process \emph{scientific knowledge compilation} and implement it in ASKS, the \emph{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.
1 Introduction
Scientific knowledge compilation addresses the challenge of organizing heterogeneous scientific records so later research tasks can inherit persistent, source-traceable interpretations. ASKS formulates this process as repeated source-local compilation and graph-state transitions rather than retrieval alone or a complete global structure.
- Motivation: Sustained research requires organization spanning papers, terminology, methods, claims, open questions, and connections across time.
- Motivation: Repeated agent workflows require a persistent knowledge substrate that later agents can inherit and build upon.Retrieval supplies transient context for a present task, whereas compilation determines what later tasks can inherit.
- Approach: ASKS compiles each source into a human-readable Wiki view and machine-facing semantic representation before persistent graph updates.A validated GraphDelta records intended document-local changes, while embeddings and explicit rules organize new and existing knowledge.
- Approach: The system treats graph emergence as higher-level organization arising through repeated local interactions and graph-state transitions.This construction-level view connects source-local compilation with emergent organization rather than assuming a complete global graph from one source.
- Scope: The study evaluates formation and persistence by chronologically compiling 56 papers and measuring reuse, support, churn, Hub lifecycle, lineage, and portrait coverage.Order reconstruction, cross-domain transfer, and alternative-method comparisons are treated as separate evaluations.
2 ASKS: System Model
ASKS preserves scientific sources while compiling reusable human-readable and machine-facing views into an evolving, inspectable research map. Embeddings provide semantic organization, while explicit validation and graph rules constrain persistent state changes and maintain source-linked authority.
- System Model: ASKS preserves received research artifacts and produces compiled notes and relations that researchers and agents can revise and inherit.Navigation helps determine where to look, while important factual paths lead back to the received sources.
- System Model: The system separates source records, persistent compiled knowledge, and provisional research state because they serve different scientific roles.Source records retain original artifacts and metadata; compiled knowledge contains reusable Wiki pages, propositions, relations, and keywords.
- Ingest: Each ingest creates transient semantic slots, a validated document-local subgraph, and an in-memory GraphDelta before persistent graph updates.These intermediates make the source transformation inspectable, analogous to compiler intermediate representations.
- Compiled Surfaces: The Wiki view supports reading and navigation, while the separately stored graph view supports computation through reusable nodes and relations.Both sibling products are compiled from the same preserved source but can evolve under different validation contracts.
- Three Computational Roles: Embeddings organize identity, alignment, routing, membership, and lifecycle candidates, while lexical gates, graph affinity, thresholds, and lineage rules make updates admissible.Continuous embedding geometry is converted into discrete graph-state changes through explicit constraints.
- Ingest: Ingest is a controlled, logged state transition whose transactional fusion either commits the complete document update or restores the previous graph state.Programmatic checks enforce source addressing, validation, duplicate removal, savepoints, origin recording, and snapshots.
- Authority Boundary: Scientific authority remains with source evidence, while the graph and Wiki provide computational organization and readable compiled views.Broader scientific assessment combines multiple records, analysis, and expert review.
3 Knowledge Compilation: From Local Interpretation to Global Organization
ASKS compiles each source into validated local representations, then integrates them into persistent graph organization through embedding geometry, explicit rules, and provenance-preserving state transitions. This separates source interpretation from cross-document routing while retaining inspectable lineage and replayable graph evolution.
- 3 Knowledge Compilation: Knowledge compilation uses two phases: source-local encoding produces representations, while consolidation determines their placement relative to accumulated knowledge.The first phase is semantic and language-facing; the second is geometric, structural, and stateful.
- 3.1 Encoding: One Source, Two Complementary Outputs: Each source yields a persistent Wiki view and machine-facing semantic slots under separate contracts, with source paths retained as a factual traceback surface.Wiki content supports reading, while semantic slots are stored and validated separately for graph construction.
- 3.2 GraphDelta: The Document-Local Intermediate Representation: GraphDelta is an inspectable document-local intermediate representation containing proposed graph changes before attachment to persistent state.It includes source-linked objects, candidate edges, identity-resolution boundaries, canonical endpoints, and structural errors.
- 3.2 GraphDelta: The Document-Local Intermediate Representation: Identity resolution prioritizes canonical identifiers and aliases, then name matches and embedding-assisted gates; ambiguous candidates abstain rather than forcing reuse.Persistent reuse requires lexical identity evidence, label and embedding tests, and a sufficient winner margin.
- 3.5 Global Organization: Topical routing keeps source-local predicates distinct from global paper placement, which compares research-positioning text with active Hub scopes.The LLM interprets the source locally, while the research map accumulates through cross-source Hub routing.
- 3.3 Transactional Fusion and Provenance-Preserving Reuse: Transactional fusion either commits the complete document update or rolls back to the prior graph state when hard checks fail.Fusion reuses or creates nodes, writes distinct edges once, and verifies that the page, raw package, and source edge remain present.
- 3.3 Transactional Fusion and Provenance-Preserving Reuse: Identical cross-source edges retain one reusable graph relation while accumulating independent source contributions in explicit lineage.This preserves multi-source support without duplicating the canonical topology.
- 3.4–3.5 Organization Dynamics: Embedding-based organization applies identity, proposition, membership, birth, split, and merge rules to turn semantic similarity into graph structure.Claim-bearing propositions may be reused, related while remaining distinct, or left independent; membership hysteresis and lineage provide temporal and structural memory.
4 Using the Compiled Structure: Navigation Before Answering
ASKS uses compiled semantic and graph structure to guide navigation before source-addressed reading and answer construction. Query outputs retain support traces for claims, composite judgments, and unresolved research items.
- 4 Using the Compiled Structure: Querying tests whether the compiled state helps scientists locate useful context, while source-addressed reading establishes what the record says.Semantic and graph representations guide search; retrieval and RAG optimize access, whereas claim verification identifies support, contradiction, and open questions.
- Navigation Before Answering: Embedding retrieval proposes related keywords, propositions, pages, or source regions even when wording differs, while graph structure contributes relations, Hubs, and neighborhoods.Similarity and graph position allocate attention, but source-addressed evidence remains the factual authority.
- Navigation Before Answering: Compiled pages function as reading indexes and source locators, letting agents inspect summaries, choose sections, and follow stable references.Staged reading concentrates relevant context and makes the scope of reading explicit.
- Support Traces: A support trace records the source region behind each supported claim, dependencies among composite judgments, and open questions requiring further retrieval.Conflicts and explored paths are recorded when available.
- Provisional and Persistent State: Composite judgments remain provisional research state until an explicit construction or review operation promotes a useful hypothesis into persistent representation.Broader scientific status draws on external review rather than substrate-level search alone.
5 Graphs Emerge: From 56 Papers to an Author Research Portrait
Repeated local interpretations accumulate into a persistent, source-traceable research organization. In the 56-paper trajectory, the graph remains low-churn while canonical vocabulary grows mainly through creation and branches around a tensor-network core.
- 204 canonical-node reuses and 878 creations across 56 papers yield Rcum(56) = 0.1885, with reuse slightly declining over time.The descriptive ordinary-least-squares slope is −5.41×10−4 per paper.
- Multi-source consolidation finishes at 0.06764 as the eligible-node denominator grows faster than the multi-source numerator.The numerator increases from 24 multi-source nodes at the peak to 60 at G56, while the denominator grows from 169 to 887.
- Mean old-node membership churn is 0.00467, with 26 of 55 defined steps showing zero change, indicating stable higher-level organization.The maximum churn is 0.03139.
- The trajectory commits 18 Hub births, all remaining active at G56, while split, merge, and retirement counts are each zero.This run therefore exercises birth and membership dynamics but not the other lifecycle operations.
- The resulting portrait preserves methodological continuity around a tensor-network core while branching into quantum many-body, tensor-network machine learning, and quantum-AI directions.The strongest branching phase occurs in 2019–2021, when 10 Hub births occur across 18 papers.
- The portrait has Psurvive = 1, Alife = 1, Bmulti = 0.889, Lbranch = 0.222, Qmap = 0.893, and ¯C = 0.00467.These descriptors jointly summarize persistent birth-driven structure, cross-paper support, selective refinement, coverage, and low churn.
6 Reliability and Bounded Authority
ASKS treats reliability as part of the scientific architecture by preserving source records, validating outputs, and making graph changes rollback-safe. Its bounded-authority design automates mechanically checkable operations while requiring explicit gates for scientific meaning and identity decisions.
- Source records remain preserved through compilation, while Wiki and semantic outputs follow separate validation contracts.
- Document-local GraphDeltas isolate graph changes, and hard fusion violations trigger rollback at the document boundary.
- Replay binds the source manifest, local model outputs, embedding identifier, configuration, operation versions, and ordered construction calls.
- The demonstration uses frozen local bundles, an isolated clean-state graph, hash-bound Hub decisions, step reports, and snapshot receipts.
- Mechanically checkable operations can be automated, but assigning scientific meaning, merging identities, or approving new Hub scope requires an explicit gate.
7 Conclusion
ASKS implements persistent scientific knowledge construction as source-to-graph compilation with validated local changes, embedding-based organization, and deterministic graph rules. In a 56-paper chronological demonstration, it produces a low-churn, source-traceable research portrait while retaining construction history and source links.
- ASKS compiles source-local LLM encodings into validated GraphDeltas, embedding-based organization, and persistent graph state with replayable source links.
- The 56-paper demonstration produces a low-churn Hub organization and a source-traceable author research portrait with explicit mapping coverage.
- The portrait centers on tensor-network methods and branches into quantum many-body research, tensor-network machine learning, and quantum-AI-oriented directions.
- Graph emergence is treated as source-linked organization, while scientific novelty remains a separate review judgment.
Code and Data Availability
The public ASKS release distributes compiled and sanitized research artifacts associated with the manuscript. It includes the source system, paper Wiki pages, Hub pages, graph data, metrics, validation records, and verification materials.
- The Ran-ASKS v0.2.0 release includes frozen paper artifact 1.0.0 and 56 sanitized paper Wiki pages.
- The release contains 18 Hub pages, the portable final graph, the reviewed corpus manifest, reported metrics, and author-portrait tables.
- Validation records, code provenance, and verification checksums are included, while primary PDFs, parsed Raw text, credentials, and private production knowledge are excluded.
Supplementary Material
The paper is titled “LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge.”
- The title presents LLM interpretation, embedding-based organization, and graph emergence as the paper’s central progression.
Supplementary Contents
The supplementary material records the frozen demonstration boundary, organization rules, ingest transaction boundary, recompilation contracts, replay and impact-analysis procedures, and publication manifest.
- The supplement includes the frozen demonstration boundary and reproducibility record.
- It documents organization rules, the complete ingest state machine, and transaction boundaries.
- It also covers re-ingest, recompilation, replay, impact analysis, construction memory, and the frozen publication manifest.
- Supplementary references are provided separately.
A Frozen Demonstration Boundary and Reproducibility Record
The demonstration is a frozen, isolated compilation run over a bounded corpus, with reproducibility locks, transactional graph updates, deterministic checks, and explicit identity and lifecycle constraints.
- Demonstration boundary: 56 independently published works were selected from 65 candidate records under fixed inclusion rules, with verified canonical PDFs and formal publication-year ordering.The corpus also contains three excluded candidates and six related-only records.
- Reproducibility record: Construction used two locked phases: fresh source compilation and auditing in Phase L, followed by ordered graph fusion from empty state G0 in Phase G.The locks bind artifacts, models, embeddings, thresholds, schemas, code hashes, runtime, and isolated write paths.
- Reproducibility record: All 56 fusion steps and 57 snapshots passed integrity, artifact-receipt, and graph-validation checks within the frozen author-corpus scope.The checks establish identity and internal consistency of the reported construction trajectory.
- Complete ingest: The compiler proceeds from source preprocessing and LLM encoding through semantic validation, GraphDelta construction, attachment planning, transactional fusion, organization, and final validation.The complete ingest is an ordered compiler pass whose fusion either succeeds completely or restores the previous state.
- Identity and validation: Boundary mentions use canonical identifiers, aliases, decomposed names, and an embedding-assisted identity gate, with ambiguity producing explicit abstention.Empty triples, self-loops, and exact duplicates are handled as hard constraints before fusion.
- Lifecycle boundaries: Re-ingest, recompilation, split, and merge contracts preserve source records and construction history while allowing active navigation to change.The production path also uses lexical fallback identity resolution during embedding-service outages.
E Replay, Impact Analysis, and Construction Memory
The paper treats replay, dependency-aware impact analysis, typed orchestration, and decision memory as engineering supports for maintaining a source-linked scientific knowledge system.
- Replay: Frozen artifacts and isolated snapshots define a complete replay boundary, while selected production sessions can serve as golden replays before system changes.Replay bindings cover prompts, models, schemas, graph code, and related execution artifacts.
- Impact analysis: The engineering dependency graph links capabilities and contracts to affected code, validation rules, and graph surfaces for impact analysis and targeted regression selection.It is physically separated from the scientific graph, which organizes source-linked knowledge, while session logs record execution diagnostics.
- Typed orchestration: Typed work states, capabilities, and tools distinguish persistent context, composable guidance, and bounded execution on the source-linked knowledge substrate.Registered tools use typed schemas, structured results, guards, and content-addressed checkpoints whose bindings determine reuse or retry.
- Construction memory: Decision memory records why predicates, Hub scopes, parsers, and validation rules changed, complementing source provenance with engineering rationale.High-consequence writes remain bounded by explicit gates even when an agent proposes the action.
- Corpus record: The frozen corpus is listed in chronological manifest order, with run labels, formal-publication dates, and bibliographically verified DOI corrections.The continued table includes later works spanning tensor-network, quantum-machine-learning, and quantum-many-body topics.