Large Knowledge Model: From Papers to a Scientific Reasoning Landscape

# Large Knowledge Model: From Papers to a Scientific Reasoning Landscape
Source: [https://arxiv.org/html/2609.27297](https://arxiv.org/html/2609.27297)
\\addtolist

\[1\]DP Technology, Beijing, China\\addtolist\[2\]Institute of Theoretical Physics, Chinese Academy of Sciences, Beijing, China\\addtolist\[3\]International Center for Quantum Materials, School of Physics, Peking University, Beijing, China\\addtolist\[4\]Lanzhou Center for Theoretical Physics, Lanzhou University, Lanzhou, China\\addtolist\[5\]AI for Science Institute, Beijing, China\\addtolist\[6\]School of Mathematical Sciences, Peking University, Beijing, China\\addtolist\[7\]Center for Machine Learning Research, Peking University, Beijing, China\\addtolist\[\*\]Corresponding authors\\addtolist\[\]Yuan Huang:[huangyuan@dp\.tech](mailto:[email protected]); Kun Chen:[chenkun@itp\.ac\.cn](mailto:[email protected])\\addtolist\[\]Ruoxue Liao:[liaorx@dp\.tech](mailto:[email protected]); Xinyu Li:[lixy@dp\.tech](mailto:[email protected])\\addtolist\[\]Linfeng Zhang:[zhanglf@dp\.tech](mailto:[email protected]); Weinan E:[weinan@math\.pku\.edu\.cn](mailto:[email protected])\\checkdata\[LKM Website\][https://lkm\.bohrium\.com/web/en](https://lkm.bohrium.com/web/en)

Sihan HuHongyu GuChao MaJiaxing ZhangZhiyong ZouCaiyu FanYan XiaoMingjun XuChenyu XieMingzhen JuZhehao MaQi ZhangBaozong WangYu LiZhiyuan YaoRuoxue LiaoXinyu LiLinfeng ZhangKun ChenWeinan E

###### Abstract

Accumulated scientific knowledge advances inquiry when prior findings help researchers choose new questions, design investigations, and interpret results\. Realizing this value at scale requires access to the reasoning that connects research problems, scientific procedures, conclusions, and evidence\. We introduce theLarge Knowledge Model \(LKM\), a scientific knowledge infrastructure that transforms the literature into a shared, computationally accessible reasoning resource\. LKM represents papers as source\-grounded reasoning graphs, couples structural traversal with semantic retrieval over the same objects, and aligns related questions, claims, and reasoning chains across papers\. This representation forms aScientific Reasoning Landscapewith three connected views: a Question Landscape that organizes research problems and open directions, a Workflow Landscape that exposes reusable scientific procedures, and an Evidence Landscape that connects conclusions to their support, disagreement, and conditions\. The unified substrate supports reasoning\-aware scientific search, evidence\-grounded question answering, comparative evidence analysis, and research planning\. Researchers and agents can retrieve relevant work through its scientific intent, synthesize answers with inspectable supporting arguments, and develop research plans informed by established workflows and unresolved evidence\. We describe a corpus\-scale system and evaluate scientific retrieval and knowledge\-intensive question answering\. With the answering model fixed, LKM retrieval improves accuracy by 9\.30%, 4\.20%, and 14\.69% on ChemBench, PubMedQA, and SciBench, respectively\. By connecting knowledge access to scientific reasoning and action, LKM provides a common foundation for discovering relevant research, reusing scientific knowledge, and coordinating cumulative inquiry across researchers, agents, and research cycles\.

††date:September 2026## 1Introduction

Scientific progress depends on how effectively the knowledge produced by one investigation informs the next\. Prior findings shape the questions worth asking, established procedures provide starting points for new studies, and accumulated evidence determines how conclusions should be interpreted\. The literature records these contributions across disciplines and generations\. Its practical value lies in making them available for further reasoning: discovering connections, selecting methods, resolving disagreement, and designing investigations that advance the scientific record\. Literature\-based discovery demonstrates how connecting previously separate findings can generate new scientific insight[Swanson \(1986\)](https://arxiv.org/html/2609.27297#bib.bib21)\.

Scientific search and question answering \(QA\) are central to this process\. Search determines which prior work enters an investigation; question answering turns that work into an organized account of what is known and why\. Structured corpora, scientific embeddings, and retrieval\-augmented language models have substantially improved these capabilities[Lo et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib11);[Cohan et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib5);[Asai et al\. \(2026\)](https://arxiv.org/html/2609.27297#bib.bib7)\. Language agents further connect retrieval with multi\-paper synthesis and the interpretation of conflicting information[Skarlinski et al\. \(2024\)](https://arxiv.org/html/2609.27297#bib.bib8)\. As these systems take on richer scientific tasks, the representation through which they access knowledge becomes increasingly consequential: it shapes what they retrieve, which evidence they connect, and how they carry prior findings into a new investigation\.

We introduce the*Large Knowledge Model*\(LKM\), a scientific knowledge infrastructure that makes accumulated scientific reasoning accessible for both knowledge use and research guidance\. LKM transforms papers into source\-grounded reasoning graphs and organizes their objects across a shared Scientific Reasoning Landscape\. This design connects scientific search, question answering, evidence analysis, and research planning through a common representation\. Relevant work can be found through the questions and reasoning it contributes; answers can be constructed from claims with inspectable support; and research plans can draw on unresolved problems, reusable procedures, and the evidence associated with their outcomes\.

### 1\.1Scientific Knowledge Beyond Documents

A scientific paper contains a structured argument within its linear presentation\. Its problem motivates subproblems; premises and observations enter ordered reasoning chains; those chains establish conclusions under particular methodological and experimental conditions\. Highlights identify informative contributions, weak points expose where the argument needs further examination, and open questions connect the study to subsequent inquiry\. These relationships often span sections, figures, equations, and references\. Making them explicit gives computational systems direct access to the organization that makes a scientific result interpretable and reusable\.

Argument\-centered representations have established the value of connecting scientific claims to supporting and challenging material[Clark et al\. \(2014\)](https://arxiv.org/html/2609.27297#bib.bib14)\. LKM develops this perspective into a persistent, corpus\-scale reasoning layer\. A question, claim, or reasoning chain becomes an addressable object with a scientific role and a path to its source context\. Corresponding objects across papers can then be connected while retaining the individual arguments in which they occur\. The literature becomes navigable through the scientific relationships that researchers need to inspect and recombine\.

This organization has immediate value for scientific knowledge access\. Question\-centered search locates work by the problem it addresses, including papers whose terminology differs from the query\. Claim\-centered retrieval connects a requested finding to the premises and reasoning that establish it\. Procedure\-centered access exposes recurring ways to investigate a problem across applications and disciplines\. The same organization guides research decisions: unresolved questions identify directions for inquiry, workflow families provide methodological alternatives, and claim\-centered evidence reveals where further measurement, validation, or explanation would be informative\.

### 1\.2The Large Knowledge Model

LKM represents each paper through connected problems, subproblems, premises, reasoning chains, conclusions, highlights, weak points, and open questions\. Source locations and references remain attached to these objects, preserving the connection between structured knowledge and the paper that produced it\. Ordered reasoning steps expose how a conclusion is reached, including the reuse of an earlier conclusion as a premise\. An agent can therefore retrieve a scientific statement and inspect the argument in which that statement has meaning\.

Structural and semantic access operate over the same knowledge objects\. Semantic vectors locate related questions, claims, and reasoning chains across differences in wording; graph relations expose their roles and dependencies\. Local\-to\-global bindings connect recurring objects across papers while preserving their paper\-specific realizations\. This integration makes semantic discovery and precise contextual expansion part of one access path: a match can be followed to supporting reasoning, source\-paper members, and related scientific objects\. The paper remains the unit of attribution, while its reasoning units become the units of retrieval, comparison, and reuse\.

At corpus scale, LKM organizes this shared representation into three aligned views\. The*Question Landscape*connects investigated problems, subproblems, and open directions\. The*Workflow Landscape*organizes recurring research procedures through workflow families, exposing common inputs and outputs, reasoning skeletons, and technique choices\. The*Evidence Landscape*connects conclusions to supporting, opposing, and conditional material, together with disagreement and the conditions that shape interpretation\. Because the views derive from the same paper\-level objects, an inquiry can be followed from its question to relevant procedures and then to the evidence associated with their conclusions\.

Figure[1](https://arxiv.org/html/2609.27297#S1.F1)illustrates this organization through the Hubble\-constant tension\. The shared question connects studies of the same quantity; the workflow view distinguishes estimation procedures; and the evidence view makes their conclusions comparable in light of their assumptions and conditions\. This alignment gives researchers a structured basis for understanding disagreement and identifying what a subsequent investigation should clarify\.

![Refer to caption](https://arxiv.org/html/2609.27297v1/from-papers-to-global-reasoning-landscape-v3.png)Figure 1:From papers to a Scientific Reasoning Landscape, illustrated with the Hubble\-constant tension, including the cosmological analysis of Planck[Planck Collaboration \(2020\)](https://arxiv.org/html/2609.27297#bib.bib27)and the distance\-ladder perspective of Freedman[Freedman \(2021\)](https://arxiv.org/html/2609.27297#bib.bib28)\. Paper\-level reasoning objects align across Question, Workflow, and Evidence views, connecting scientific intent, research procedures, and comparable conclusions\. The same connections support search, evidence\-grounded synthesis, and research planning\. Persistent paper identifiers retain the path from landscape objects to their source reasoning, while newly ingested papers extend the shared scientific record\.
### 1\.3Scientific Search, Question Answering, and Research Planning

LKM makes scientific search an access point to research reasoning\. A query can retrieve relevant questions and claims, expand the chains that establish them, and locate the papers contributing that knowledge\. Hybrid search combines lexical and semantic signals, while graph\-aware expansion supplies the surrounding reasoning context\. Researchers and agents gain both relevant sources and the scientific structure needed to judge their usefulness for the task at hand\.

For question answering, this context provides material for explanation and evidence\-grounded synthesis\. Claims remain connected to premises, reasoning steps, and source references, allowing answers to draw on the arguments behind retrieved findings\. Across papers, the Evidence Landscape helps researchers examine where conclusions agree, where they conflict, and under which assumptions or experimental conditions they hold\. Search and QA thus serve complementary scientific functions: finding the work that matters and constructing an attributable account of what that work establishes\.

Research planning builds on these same connections\. An open question can be examined alongside the claims that partially address it, the weak points that motivate further study, and the workflow families available for investigation\. Researchers and agents can compare methodological choices and formulate a plan that links each proposed step to prior reasoning and evidence\. Shared objects also provide the representational basis for returning new observations and conclusions to the landscape\. LKM gives researchers and agents a shared memory for connecting existing knowledge, proposed investigations, and emerging evidence across successive research cycles\.

### 1\.4Contributions

This report makes four contributions:

1. 1\.A source\-grounded paper\-level reasoning representation that makes scientific questions, claims, and multi\-step arguments directly addressable, together with their supporting context and open directions\.
2. 2\.A unified graph and vector access layer that connects semantic discovery, structural traversal, and cross\-paper identities over the same scientific knowledge objects\.
3. 3\.A Scientific Reasoning Landscape with aligned Question, Workflow, and Evidence views, connecting problem navigation, methodological reuse, evidence comparison, and research planning\.
4. 4\.A corpus\-scale system for reasoning\-aware scientific search and evidence\-grounded QA, with evaluation of retrieval quality, scientific answer accuracy, and citation attribution\.

The current system operates over tens of millions of papers\. On matched PaSaMaster queries, LKM Hybrid achieves the highest normalized discounted cumulative gain at retrieval depths of 5, 10, and 20 among the evaluated APIs\. With GPT\-5\.4 fixed as the answering model, LKM retrieval improves accuracy over the bare model by 9\.30, 4\.20, and 14\.69 percentage points on ChemBench, PubMedQA, and SciBench, respectively\. The ScholarQA comparison examines how graph\-structured context contributes to answer construction and evidence attribution\. Together, these tasks characterize the practical value of LKM for finding, understanding, and using scientific knowledge\.

LKM connects the accumulated literature to the scientific activity it informs\. Its shared reasoning objects support knowledge access, make research decisions inspectable, and provide continuity across successive investigations\.[Section2](https://arxiv.org/html/2609.27297#S2)situates the work in prior research;[Sections3](https://arxiv.org/html/2609.27297#S3)and[4](https://arxiv.org/html/2609.27297#S4)present the representation, landscape views, and system realization;[Section5](https://arxiv.org/html/2609.27297#S5)evaluates scientific search and QA; and[Section6](https://arxiv.org/html/2609.27297#S6)develops the path toward a cumulative scientific ecosystem\.

## 2Related Work

### 2\.1Scientific Search and Question Answering

Dense retrieval enables semantic matching across large corpora[Karpukhin et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib3), and retrieval\-augmented generation connects retrieved knowledge to generated answers[Lewis et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib4)\. Scientific search and QA develop this capability into literature\-centered investigation\. OpenScholar combines scientific retrieval with citation\-backed synthesis[Asai et al\. \(2026\)](https://arxiv.org/html/2609.27297#bib.bib7); PaperQA2 uses language agents to search the literature, synthesize findings, and resolve conflicting information[Skarlinski et al\. \(2024\)](https://arxiv.org/html/2609.27297#bib.bib8)\. PaSa and PaSaMaster extend paper discovery through iterative agentic search and relevance assessment[He et al\. \(2025\)](https://arxiv.org/html/2609.27297#bib.bib9);[Du et al\. \(2026\)](https://arxiv.org/html/2609.27297#bib.bib10)\. These systems establish retrieval and synthesis as productive interfaces to scientific knowledge\.

LKM advances this trajectory by making the retrieved scientific content persistently structured and reusable\. Questions, claims, and reasoning chains are access points into a paper’s argument, and graph expansion connects them to premises, procedures, and source evidence\. This organization benefits both search and QA: a search result can identify a specific contribution and expose its justification, while an answer can draw on the relations connecting findings rather than treating each passage as an independent fragment\. The same objects remain available for subsequent comparison and investigation, turning individual retrieval interactions into access to a shared scientific reasoning resource\.

### 2\.2Scientific Knowledge Representation

Scientific corpora, extraction systems, and knowledge graphs organize complementary aspects of the literature\. S2ORC connects full text with bibliographic metadata and references[Lo et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib11); the Open Research Knowledge Graph represents scholarly contributions as structured, comparable knowledge[Jaradeh et al\. \(2019\)](https://arxiv.org/html/2609.27297#bib.bib12)\. SciERC extracts scientific entities and relations[Luan et al\. \(2018\)](https://arxiv.org/html/2609.27297#bib.bib15), while SciREX captures document\-level combinations of tasks, methods, datasets, and metrics[Jain et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib16)\. SPECTER and SciRepEval develop citation\-informed and task\-sensitive scientific document representations[Cohan et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib5);[Singh et al\. \(2023\)](https://arxiv.org/html/2609.27297#bib.bib6)\. Together, these approaches make scientific content easier to retrieve, connect, and compare\.

Argument\-centered representations address the structure through which scientific knowledge is established\. Argumentative zoning identifies rhetorical roles[Teufel and Moens \(2002\)](https://arxiv.org/html/2609.27297#bib.bib19); nanopublications associate assertions with provenance[Groth et al\. \(2010\)](https://arxiv.org/html/2609.27297#bib.bib13); and micropublications connect claims, evidence, methods, and supporting or challenging arguments[Clark et al\. \(2014\)](https://arxiv.org/html/2609.27297#bib.bib14)\. Particularly close to LKM,*Linking Scholarly Contents*constructs an argumentation graph with fine\-grained scholarly objects and intra\- and inter\-paper relations[Song et al\. \(2022\)](https://arxiv.org/html/2609.27297#bib.bib29)\. It demonstrates how argument\-level organization exposes connections that are difficult to access through bibliographic links alone\.

LKM builds on this insight by organizing paper\-level questions and multi\-step reasoning together with claims, contextual annotations, and source links\. Paper\-local records preserve each contribution’s argumentative setting; canonical bindings support exact object reuse; and object\-level vectors support semantic discovery across formulations\. This combination makes argument structure operational for corpus\-scale access: users can retrieve a proposition, inspect its derivation, and locate related scientific reasoning through the same underlying object identities\.

### 2\.3Literature Synthesis and Workflow Mining

Literature\-based discovery shows that connections across separate bodies of work can generate new scientific insight[Swanson \(1986\)](https://arxiv.org/html/2609.27297#bib.bib21)\. Workflow\-centric research objects preserve the components and context of computational investigations[Belhajjame et al\. \(2015\)](https://arxiv.org/html/2609.27297#bib.bib20), and annotated laboratory protocols expose experimental actions and their relations for machine reading[Kulkarni et al\. \(2018\)](https://arxiv.org/html/2609.27297#bib.bib18)\. Scientific claim verification connects claims to supporting or refuting evidence and rationales[Wadden et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib17)\. These perspectives reveal three complementary dimensions of scientific activity: the questions being investigated, the procedures used to investigate them, and the evidence obtained\.

*The Discovery Engine*develops a closely related vision of navigable scientific knowledge landscapes, combining structured knowledge artifacts, graph and vector views, source grounding, and agent\-oriented synthesis[Baulin et al\. \(2025\)](https://arxiv.org/html/2609.27297#bib.bib30)\. This framework highlights the value of organizing knowledge for conceptual navigation and discovery across fields\. LKM gives such navigation a shared paper\-level substrate and aligns it through Question, Workflow, and Evidence landscapes\. Their connection enables a research question to lead to alternative investigative approaches and then to the claims and conditions associated with their outcomes\. The resulting perspective supports scientific planning by bringing the problem, the available approaches, and their evidential basis into one navigable structure\.

### 2\.4Scientific Reasoning as Shared Infrastructure

Graph\-based retrieval illustrates how structural context improves access to interconnected knowledge\. GraphRAG uses entity graphs and community summaries for query\-focused synthesis[Edge et al\. \(2024\)](https://arxiv.org/html/2609.27297#bib.bib22), and LightRAG combines graph structure with vector\-based retrieval[Guo et al\. \(2025\)](https://arxiv.org/html/2609.27297#bib.bib23)\. ALCE makes citation quality an explicit dimension of answer evaluation[Gao et al\. \(2023\)](https://arxiv.org/html/2609.27297#bib.bib24)\. These advances motivate retrieval architectures in which semantic access, relational context, and evidence attribution work together\.

LKM centers this integration on scientific reasoning\. Its contribution connects three levels: paper\-level graphs expose how findings are established; graph and vector access make their components retrievable and reusable; and aligned landscapes organize those components into cross\-paper views of questions, approaches, and evidence\. Scientific search and QA provide immediate applications and empirical tests of this infrastructure, while landscape navigation extends its utility to synthesis, methodological comparison, and research planning\. The organizing insight is that scientific knowledge becomes more actionable when a finding remains connected to the problem it addresses and the reasoning that supports it\.

## 3Representing Scientific Papers as Reasoning Graphs

Scientific knowledge becomes more useful when readers and computational agents can access not only a result, but also the question it answers, the reasoning that establishes it, and the conditions under which it applies\. LKM makes these components individually addressable and explicitly connected\. We represent a paper as a typed directed graph with inference factors,𝒢=\(𝒱,ℱ,ℰ\)\\mathcal\{G\}=\(\\mathcal\{V\},\\mathcal\{F\},\\mathcal\{E\}\), where𝒱\\mathcal\{V\}is the set of scientific knowledge objects,ℱ\\mathcal\{F\}is the set of inference factors connecting premises to conclusions, andℰ\\mathcal\{E\}is the set of typed relations\. This representation turns the paper into a reusable reasoning resource: users can enter through a question, claim, or method and follow the scientific dependencies relevant to their task\.

### 3\.1Reasoning Primitives and Relations

###### Definition 1\(Claim\)\.

A*claim*is a proposition node that appears as a premise or conclusion in one or more reasoning chains or inference factors\.

###### Definition 2\(Question\)\.

A*question*is a node encoding a research problem, subproblem, or open question; it frames the claims and reasoning chains that address it\.

###### Definition 3\(Setting\)\.

A*setting*records experimental or domain context attached to the claims, chains, and evidence to which it applies\.

###### Definition 4\(Reasoning chain\)\.

A*reasoning chain*is an ordered sequence of stepsσ=\(s1,…,sT\)\\sigma=\(s\_\{1\},\\ldots,s\_\{T\}\), whereTTis the number of steps andsts\_\{t\}is thett\-th step, connecting premises, operations, and a conclusion; it is attached to the factor whose inference it explains\.

###### Definition 5\(Evidence\)\.

*Evidence*for claimccis the set of observations, premises, reasoning steps, and source anchors attached to the chains that conclude incc\.

###### Definition 6\(Method\)\.

A*method*is the inference rule or operator subtype of a factor, together with the procedure, tools, data, and model choices recorded in its reasoning chain\.

###### Definition 7\(Relations\)\.

An*inference factor*maps an ordered premise set to a conclusion and records its method, parameters, and attached reasoning chain\.*Relations*includeaddresses\(problem to claim\),premise\_of\(premise to reasoning chain or conclusion\),concludes\(reasoning chain to conclusion\),highlight\_of, andweakpoint\_of, together with links between questions, chains, claims, settings, and sources\.

### 3\.2Paper\-level Reasoning Graphs

A paper’s exposition is linear, but its scientific contribution is assembled through dependencies across observations, intermediate results, methodological choices, and conclusions\. Argumentative zoning and claim–evidence representations make important parts of this structure explicit[Teufel and Moens \(2002\)](https://arxiv.org/html/2609.27297#bib.bib19);[Clark et al\. \(2014\)](https://arxiv.org/html/2609.27297#bib.bib14);[Song et al\. \(2022\)](https://arxiv.org/html/2609.27297#bib.bib29)\. LKM organizes these dependencies around research questions and multi\-step reasoning, enabling a reader to reconstruct how a finding was established and how it contributes to the larger investigation\.

For a paperpp, the paper\-local graph is

𝒢ploc=\(𝒱ploc,ℱploc,ℰploc\),\\mathcal\{G\}^\{\\mathrm\{loc\}\}\_\{p\}=\(\\mathcal\{V\}^\{\\mathrm\{loc\}\}\_\{p\},\\mathcal\{F\}^\{\\mathrm\{loc\}\}\_\{p\},\\mathcal\{E\}^\{\\mathrm\{loc\}\}\_\{p\}\),where𝒱ploc\\mathcal\{V\}^\{\\mathrm\{loc\}\}\_\{p\}contains paper\-scoped typed objects,ℱploc\\mathcal\{F\}^\{\\mathrm\{loc\}\}\_\{p\}contains inference factors, andℰploc\\mathcal\{E\}^\{\\mathrm\{loc\}\}\_\{p\}contains typed relations\. Source references, locations where supplied by extraction, and paper\-specific roles accompany the objects they describe\. These attributes connect graph\-level exploration to the underlying scientific account\.

Construction fuses four complementary structured views\. The conclusion view identifies the paper\-level problem, conclusions, subproblems, open questions, and structural links\. The reasoning\-chain view records the ordered steps supporting each conclusion\. The review view contributes premise assessments, highlights, weak points, and available inference parameters\. The refined view supplies decontextualized proposition text, titles, and references\. Their integration combines statements that can be understood outside their original paragraph with the paper\-specific reasoning and source context that establish their meaning\. A conclusion becomes a claim reached through an inference factor and its attached chain; premises, reused conclusions, and methodological operations retain their distinct roles\.

Figure 2:A real reasoning neighborhood from*Attention Is All You Need*[Vaswani et al\. \(2017\)](https://arxiv.org/html/2609.27297#bib.bib31)\.\(a\) Stored dependencies connect architectural components to the Transformer and its empirical applications\. \(b\) The architecture inference expands into five ordered steps\. Arrows reproduce LKM relations; labels condense source records\.Figure[2](https://arxiv.org/html/2609.27297#S3.F2)shows the graph extracted from*Attention Is All You Need*[Vaswani et al\. \(2017\)](https://arxiv.org/html/2609.27297#bib.bib31)\. The scaled dot\-product attention conclusion is reused in the multi\-head attention argument\. Multi\-head attention, the position\-wise feed\-forward network, sinusoidal positional encoding, and the complexity analysis then enter the inference establishing the Transformer architecture\. This architecture is subsequently reused in arguments about translation performance and constituency parsing\. Expanding the architecture factor reveals five ordered steps explaining how its components are assembled and how they function\. The graph therefore exposes a progression from reusable primitives to an integrated method and its empirical applications\. Readers can inspect an individual contribution, recover its scientific dependencies, or follow it into the investigations that use it\.

The paper\-level graph is consequently both a reading interface and a unit of knowledge reuse\. Questions identify what an investigation seeks to resolve; chains expose how it proceeds; claims capture its outcomes; and attached evidence and settings provide the basis for interpretation\. This organization makes a paper’s contribution available at the granularity needed for scientific search, question answering, evidence comparison, and research planning\.

### 3\.3Canonical Storage, Vectorization, and Access

LKM preserves paper\-local realizations while giving recurring objects shared corpus\-level identities\. Local records retain typed variables, inference factors, ordered steps, source packages, references, and structural metadata\. Canonical bindings connect these records to global representatives and retain their source\-paper membership\. A shared object can thus bring together its occurrences across the corpus, while each occurrence remains accessible in the argument that gives it scientific context\.

The implementation establishes public\-variable identity through a SHA\-256 hash of the variable type, stored proposition content, and sorted parameter\-name/type declarations\. Exact matches share a global representative; private variables receive separate identities\. Factor integration matches the mapped premise IDs, conclusion ID, factor type, and subtype\. These mechanisms provide stable reference points for repeated propositions and inference structures\. Semantic relatedness is represented through vector neighborhoods and clustering, giving the system complementary ways to recognize an identical object and discover related formulations\. Together, these operations support cross\-paper organization at both exact and semantic levels\.

ByteHouse persists local and global records, canonical bindings, parameter records, embeddings, and clustering results; optional topology services support traversal and visualization\. Prior records attach assessments to claims, and factor\-parameter records attach conditional assessments to inference relations, retaining their parameterization sources\. Keeping these assessments separate from topology allows the scientific structure and its associated assessments to evolve through distinct update paths\. The same claim can remain a stable access point as additional source occurrences and assessment records enter the system\.

Graph identities also anchor semantic access\. For a public global variablevv, letxvx\_\{v\}be its decontextualized text and letEmbed\\operatorname\{Embed\}denote the configured text encoder\. The default configuration produces

𝐳v=Embed⁡\(xv\)∈ℝ512,\\mathbf\{z\}\_\{v\}=\\operatorname\{Embed\}\(x\_\{v\}\)\\in\\mathbb\{R\}^\{512\},where𝐳v\\mathbf\{z\}\_\{v\}is the variable’s dense semantic vector\. Embedding records include node type, graph role, and source identifiers\. Reasoning chains receive separate vectors by encoding their ordered, concatenated step text, with the associated conclusion retained as metadata\. Users can therefore retrieve both*what was established*and*how it was established*: claim vectors support proposition\-centered discovery, while chain vectors support discovery of reasoning procedures and methodological approaches\.

The access layer connects semantic retrieval to graph expansion and source inspection\. A query retrieves relevant questions, claims, or chains; their identifiers resolve to source\-paper objects and associated inference structure\. A retrieved conclusion can be expanded into its premises, ordered reasoning, supporting observations, and contextual annotations\. This continuity gives search results an actionable scientific structure, supplies QA with connected evidence, and exposes reusable components for the Question, Workflow, and Evidence landscapes developed in Section[4](https://arxiv.org/html/2609.27297#S4)\.

### 3\.4Current Scale and Scientific Search

The current deployment contains reasoning graphs for more than 40 million papers and abstract representations for more than 70 million papers, spanning the major scientific disciplines\. These two coverage layers combine broad discovery with detailed access to scientific arguments\. Abstract representations support corpus\-wide recall, while reasoning graphs support inspection of the findings, procedures, and evidence within retrieved papers\.

Users and agents can search by natural\-language question, concept, method, dataset, model, tool, or claim, then inspect the corresponding paper and reasoning objects\. Shared identifiers connect vector hits, graph relations, and source context, enabling a continuous transition from finding relevant literature to understanding and comparing its contributions\. Citation\-aware retrieval and paper\-graph inspection provide additionnal paths through the corpus\. Search and QA are therefore direct applications of the reasoning substrate: search locates useful scientific objects, and QA assembles their connected knowledge into evidence\-grounded responses\. As new papers are structured and integrated, they also enrich the questions, workflows, and evidence available for cross\-paper analysis and scientific planning\.

## 4The Scientific Reasoning Landscape

A paper\-level graph exposes the reasoning within an investigation\. The Scientific Reasoning Landscape connects that reasoning across the literature, making accumulated findings useful for understanding a field and deciding how to investigate it next\. Its organizing principle is to keep scientific questions, investigative procedures, and evidence aligned: a researcher can examine what needs to be resolved, how related work has approached it, and what the resulting evidence establishes\.

### 4\.1Questions, Workflows, and Evidence as Aligned Views

Let𝒫\\mathcal\{P\}denote the corpus of ingested papers, with each paperp∈𝒫p\\in\\mathcal\{P\}contributing the paper\-local graph𝒢ploc\\mathcal\{G\}^\{\\mathrm\{loc\}\}\_\{p\}defined in Section[3](https://arxiv.org/html/2609.27297#S3)\. We define the*Scientific Reasoning Landscape*as the cross\-paper organization

ℒ=\(𝒬,𝒲,ℬ,ℛ\),\\mathcal\{L\}=\(\\mathcal\{Q\},\\mathcal\{W\},\\mathcal\{B\},\\mathcal\{R\}\),where𝒬\\mathcal\{Q\}is the set of research questions,𝒲\\mathcal\{W\}is the set of workflow families,ℬ\\mathcal\{B\}is the set of claim\-centered evidence analyses, andℛ\\mathcal\{R\}is the set of relations connecting these objects\. The three views derive from the same paper\-level substrate\. Question objects identify scientific intent; reasoning chains supply investigative structure; and claims, premises, source references, and contextual annotations supply the material for evidence analysis\.

This alignment makes the views complementary paths through scientific knowledge\. Starting from a question, a researcher can discover relevant contributions, examine the procedures behind them, and compare their evidential basis\. Starting from a finding, the researcher can recover the question it addresses and the method that produced it\. Figure[3](https://arxiv.org/html/2609.27297#S4.F3)illustrates how these paths remain connected through their underlying paper objects\.

Figure 3:Three aligned views of scientific reasoning\.Shared paper objects connect question scoping, workflow adaptation, and evidence interpretation through exact bindings, semantic neighborhoods, typed relations, and source links\. Q denotes a question, R reasoning, and C a claim\.
### 4\.2Question Landscape

The Question Landscape organizes scientific knowledge around the problems that motivate research\. Paper\-level problems, conclusion\-specific subproblems, and open questions preserve the relationship between an investigation’s objective, its individual contributions, and its subsequent directions\. In the paper graph,addresseslinks a problem to its conclusions, whilesubproblem\_ofassociates a subproblem with its corresponding conclusion\. Shared identities and semantic neighborhoods connect these question objects across papers, enabling related investigations to be explored together\.

Question\-centered access reveals how a field develops through successive contributions\. A broad problem can be examined through the subproblems that make it tractable, the claims that address different aspects of it, and the open directions identified by contributing papers\. Researchers can locate work through its scientific intent even when terminology differs, and inspect the reasoning behind these contributions\. An open question becomes an entry point into relevant claims, reasoning chains, and source papers\.

This perspective supports research scoping and planning\. It brings a proposed question into contact with prior findings, methodological alternatives, and evidence gaps, helping researchers identify what a new investigation could clarify\. The landscape thereby connects literature discovery to the formulation of useful research questions\.

### 4\.3Workflow Landscape

The Workflow Landscape organizes recurring ways of producing and validating scientific knowledge\. Workflow\-centric research objects and protocol extraction establish the value of making investigative structure accessible[Belhajjame et al\. \(2015\)](https://arxiv.org/html/2609.27297#bib.bib20);[Kulkarni et al\. \(2018\)](https://arxiv.org/html/2609.27297#bib.bib18)\. In LKM, a*workflow family*groups reasoning chains around a recurring scientific procedure\. A family description records its definition and corpus participation and can specify common inputs and outputs, a reasoning skeleton, problem variants, task templates, and technique slots\. Technique slots record alternative techniques and their applicability, exposing choices within a shared procedure\.

A family thus connects an abstract approach to the concrete investigations that instantiate it\. Its skeleton identifies the purposes of successive steps; task templates and problem variants describe uses of that structure; and technique alternatives expose how it can be adapted\. Laboratory type distinguishes dry, wet, and mixed investigations, while verification mode describes numerical, LLM\-based, multimodal, or experimental assessment\. These attributes help researchers compare approaches in light of their available resources and validation requirements\.

The resulting view makes methodological knowledge discoverable across individual papers and applications\. Researchers can inspect how a procedure has been used, compare alternative implementations, and identify a starting point for a new investigation\. For agents, the same structure provides reusable planning material: a sequence of scientific purposes together with the choices and source examples needed to instantiate it\.

### 4\.4Evidence Landscape

The Evidence Landscape organizes what the literature establishes around individual claims and the conditions of their interpretation\. Scientific claim verification makes support and refutation explicit[Wadden et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib17); LKM connects these assessments to the reasoning and source context behind the findings\. An evidence item links a claim to a source paper, a reasoning\-chain preview, and available chain analysis, including steps, references, strengths, weaknesses, failure modes, and probability parameters\.

For a target claim, the Evidence Engine constructs an analysis that groups retrieved evidence as supporting, opposing, conditional, insufficient, or methodologically divergent\. These categories preserve distinctions that matter scientifically: a result obtained under different conditions may qualify a claim or expose a methodological difference rather than directly oppose it\. Contradiction groups organize disagreements by their members, cause, explanation, and severity\. Quality assessments and evidence balance accompany the underlying items, allowing researchers to inspect both an overall account and the reasoning behind its constituent findings\.

This organization turns evidence synthesis into a basis for further investigation\. Researchers can examine why results differ, which conditions affect their interpretation, and where additional work would be informative\. Research\-gap records identify methodological, measurement, validation, transfer, mechanism, and scope questions, with source evidence and available next\-step suggestions\. The view therefore supports attributable answers and comparative analysis while connecting unresolved evidence to concrete directions for research\.

Figure[4](https://arxiv.org/html/2609.27297#S4.F4)illustrates the aligned views through three source\-grounded studies of efficient attention: FlashAttention, FlashAttention\-2, and Linformer[Dao et al\. \(2022\)](https://arxiv.org/html/2609.27297#bib.bib32);[Dao \(2023\)](https://arxiv.org/html/2609.27297#bib.bib33);[Wang et al\. \(2020\)](https://arxiv.org/html/2609.27297#bib.bib34)\. Their paper graphs expose complementary routes through the same research problem: reducing memory traffic, improving GPU work partitioning, and introducing low\-rank approximation\. Examining the methods together with their evidence helps researchers select an approach for the required accuracy, hardware, and workload\.

Figure 4:Three source\-grounded routes to efficient attention\.Linformer, FlashAttention, and FlashAttention\-2 connect distinct subproblems to methods, inference factors, and contextualized findings\. Solid arrows reproduce LKM relations; dashed links organize the sources under a shared inquiry\. C and R identifiers are paper\-local\.
### 4\.5Landscape Alignment and Evolution

The value of the landscape comes from continuity between its views\. Canonical bindings connect exact recurring public objects and matching factor structures to global representatives, retaining their paper\-local realizations\. Semantic representations support discovery of related formulations, and clustering organizes recurring question or reasoning patterns\. Workflow\-family membership and claim\-centered evidence records provide additional cross\-paper organization\. These connections let researchers move between a problem, the approaches used to investigate it, and the evidence associated with their outcomes\.

The same structure provides continuity as the literature grows\. Ingestion creates and commits local graph records; integration establishes bindings and matches public variables by content hash and factors by structure\. Embedding workers extend semantic access to variables and reasoning chains\. Clustering and family construction organize the resulting cross\-paper patterns, while source\-linked parameter records retain their associated assessments\. Stable object identities allow new paper occurrences to enter an existing scientific context, and semantic organization can be refreshed as the corpus develops\.

Scientific access interfaces make this organization usable\. Hybrid node search retrieves claims and questions; reasoning search retrieves whole chains; claim\-level reasoning views expose premises and ordered steps; and paper\-graph access returns the scientific structure of an individual study\. Workflow interfaces expose families and their methodological alternatives, while the Evidence Engine develops claim\-centered analyses from retrieved sources and reasoning\. Together, these paths support finding relevant research, synthesizing evidence\-grounded answers, comparing investigative approaches, and planning further study\. LKM makes the accumulated literature a connected resource for scientific decisions, with each landscape view retaining a route to the contributions on which it is built\.

## 5Evaluation

### 5\.1Experimental Setup

Our evaluation examines whether LKM’s structured representation yields measurable gains at the interfaces through which researchers and agents encounter scientific knowledge: search, question answering, and evidence use\. We therefore evaluate scientific search and knowledge\-intensive question answering under matched, operational settings\.

For scientific search, we use the 244 matched queries from PaSaMaster[Du et al\. \(2026\)](https://arxiv.org/html/2609.27297#bib.bib10)\. We compare the Science Navigator API, the LKM Semantic API, and the LKM Hybrid API\. Ranking quality is measured byNDCG​@​k\\mathrm\{NDCG\}@k, wherek∈\{5,10,20\}k\\in\\\{5,10,20\\\}denotes the retrieval depth\. We additionally report candidate\-pool recall, ranking quality, and latency for the agent\-integrated setting\.

For question answering, GPT\-5\.4 is fixed as the answering model in all configurations\. We evaluate ChemBench[Mirza et al\. \(2025\)](https://arxiv.org/html/2609.27297#bib.bib25)\(n=2,785n=2\{,\}785\), PubMedQA[Jin et al\. \(2019\)](https://arxiv.org/html/2609.27297#bib.bib2)\(n=500n=500\), SciBench[Wang et al\. \(2024\)](https://arxiv.org/html/2609.27297#bib.bib26)\(n=580n=580\), and ScholarQA[Asai et al\. \(2024\)](https://arxiv.org/html/2609.27297#bib.bib1)\. The baselines include a bare model, Web Search, arXiv retrieval, Google Scholar, and Science Navigator\. For ChemBench, PubMedQA, and SciBench, we report overall answer accuracy\. For ScholarQA, we assess answer quality together with citation F1; citation precision and recall characterize the faithfulness with which answers attribute evidence\.

### 5\.2Scientific Search

Figure[5](https://arxiv.org/html/2609.27297#S5.F5)presents the search results on the matched PaSaMaster queries\.

Figure 5:Scientific search quality on PaSaMaster\.The LKM Hybrid API achieves the highestNDCG\\mathrm\{NDCG\}at @5, @10, and @20, reaching 22\.19, 21\.00, and 20\.93, respectively\.The LKM Hybrid API obtainsNDCG​@​5=22\.19\\mathrm\{NDCG\}@5=22\.19,NDCG​@​10=21\.00\\mathrm\{NDCG\}@10=21\.00, andNDCG​@​20=20\.93\\mathrm\{NDCG\}@20=20\.93\. The corresponding values for the LKM Semantic API are 21\.17, 20\.43, and 20\.54, while the Science Navigator API obtains 21\.44, 19\.77, and 19\.28\. Thus, the hybrid configuration improves over Science Navigator by 0\.75, 1\.23, and 1\.65 percentage points at the three retrieval depths\. Relative to the strongest alternative at each depth, the corresponding margins are 0\.75, 0\.57, and 0\.39 points\.

The matched\-query analysis shows that the improvement is consistent under the same query set rather than being caused by different query coverage\. In the agent\-integrated setting, reasoning\-guided candidate generation further increases candidate\-pool recall by approximately 2\.93–3\.45 percentage points\. These results show that reasoning structures function as effective search units and expand the evidence candidate space available to downstream reasoning\.

### 5\.3Knowledge\-intensive Question Answering

We first evaluate retrieval\-assisted scientific question answering on ChemBench, PubMedQA, and SciBench\. The results are shown in Figure[6](https://arxiv.org/html/2609.27297#S5.F6)\.

Figure 6:Scientific QA accuracy with fixed GPT\-5\.4\.LKM retrieval improves over the bare model by 9\.30, 4\.20, and 14\.69 percentage points on ChemBench, PubMedQA, and SciBench, respectively\.LKM retrieval achieves 74\.08% on ChemBench, 81\.40% on PubMedQA, and 90\.90% on SciBench\. Compared with the bare model, these results yield gains of 9\.30, 4\.20, and 14\.69 percentage points\. LKM also outperforms Science Navigator by 1\.26, 0\.20, and 2\.28 percentage points, respectively\. Because the answering model is fixed, the gains directly demonstrate the value of structured scientific retrieval for knowledge\-intensive reasoning\.

We further evaluate whether graph\-structured context improves evidence use on ScholarQA\. Figure[7](https://arxiv.org/html/2609.27297#S5.F7)compares a unified API\-baseline workflow with a reranked evidence\-synthesis workflow\. The unified workflow includes the base model, arXiv, Google Scholar, Science Navigator, and LKM Search\. The reranked workflow compares Science Navigator, LKM Search, and LKM Graph\. LKM Graph provides graph\-addressable claims, reasoning chains, and evidence relations in addition to retrieved scientific content\.

Figure 7:Answer and citation quality on ScholarQA\.LKM Graph substantially improves citation F1 over search\-only context on both ScholarQA\-CS and ScholarQA\-Multi while maintaining comparable answer quality\.LKM Graph reaches citation F1 scores of 54\.06% on ScholarQA\-CS and 57\.79% on ScholarQA\-Multi, improving over LKM Search by 5\.71 and 6\.67 percentage points, respectively\. The corresponding answer\-quality scores remain comparable across search and graph contexts\. Paired analysis further shows that graph context improves both citation precision and citation recall, with confidence intervals for the graph\-over\-search differences excluding zero\.

Taken together, the results trace a coherent path from better retrieval to stronger answers and more faithful evidence attribution\. LKM turns scientific reasoning from latent document content into an operational substrate for constructing and citing scientific answers\.

## 6Discussion and Future Directions

LKM is designed as infrastructure for a scientific system that learns across research cycles\. Its purpose is not simply to place more literature in a structured store, but to retain the state of inquiry: what was asked, what was attempted, what evidence was produced, and what remains open\. The retrieval and attribution interfaces evaluated here make that substrate operational\.

### 6\.1From Open Questions to Shared Scientific Workflows

Open questions provide a natural entry point into the system\. By aligning addressed and unresolved questions into canonical and common\-question families, LKM makes gaps visible across papers and domains\. A newly posed question can then be connected to the reasoning chains, evidence, and methodological resources that bear on it, rather than treated as an isolated prompt\.

These links expose reusable workflow structures: recurring inputs and outputs, methodological skeletons, tools, software, datasets, and models\. Researchers and agents can inspect how a workflow has been used, under which conditions, and with what outcome, then adapt it to a new question\. A workflow thus becomes a shared scientific object that can be reused, compared, and improved\.

### 6\.2Cross\-domain Synthesis and Self\-iteration

Literature\-based discovery has long shown the scientific value of connecting separate bodies of knowledge[Swanson \(1986\)](https://arxiv.org/html/2609.27297#bib.bib21)\. Because questions, workflows, and evidence are aligned through common paper\-level objects, LKM can connect ideas expressed in different vocabularies or developed in separate fields\. A technique, dataset, or reasoning pattern established for one problem can be surfaced as a candidate approach to another\. The agent can combine these aligned objects to propose a hypothesis, workflow, or experimental design while retaining the source conditions that justify each transfer\.

The resulting plan can be represented as an executable research task\. Its intermediate steps, observations, failures, and results become new reasoning objects linked to the claims and workflows that motivated them\. Positive results extend a line of reasoning; negative results sharpen its scope or create a new question\. In both cases, the outcome returns to LKM and enriches the landscape available to the next research cycle\.

### 6\.3What the Current Experiments Establish

The current experiments establish the first stages of this cycle\. Hybrid reasoning\-aware retrieval improves candidate ranking and pool recall; structured context improves knowledge\-intensive QA; and graph\-addressable context improves citation precision and recall while preserving comparable answer quality\. Together, these findings show that the shared objects support discovery of relevant work, reasoning over it, and accountable reuse of evidence\.

This capability creates a clear path to ecosystem\-level scientific functions: identifying consequential open questions, composing reproducible studies from shared workflows, executing those workflows with agents, and integrating the resulting evidence into subsequent inquiry\. LKM supplies the memory and coordination layer that makes those functions cumulative rather than episodic\.

### 6\.4Building a Scientifically Accountable Ecosystem

A reasoning\-native ecosystem gains value when every new object remains inspectable in the context that produced it\. LKM therefore represents provenance, confidence, correction paths, failures, and negative outcomes alongside positive evidence\. Human researchers can inspect and revise the graph, while agents can expose the questions, workflows, and evidence behind a proposed action\. These properties make iterative scientific use transparent, measurable, and extensible\.

## 7Conclusion

Scientific AI requires a persistent, shared account of how knowledge is produced\. LKM provides that account by transforming papers into source\-grounded reasoning graphs and aligning questions, workflows, claims, evidence, and resources in one addressable space\. The same objects support semantic retrieval, structural traversal, comparison, and citation, allowing an agent to move from a question to evidence without breaking the provenance chain\.

The current results anchor this foundation empirically: reasoning\-aware retrieval improves access to relevant work, structured context improves scientific QA, and graph\-addressable evidence strengthens citation attribution\. As researchers and agents use LKM to propose, execute, critique, and reproduce scientific work, their outputs can return to the shared landscape as new evidence, revised claims, and new questions\.

LKM is therefore more than a representation of the literature\. It is a memory and coordination layer for a scientific ecosystem that preserves what was tried, why it worked, where it failed, and what should be investigated next\. By keeping scientific reasoning computationally usable and source accountable, it creates the substrate on which cumulative scientific agents can be built\.

## References

- Asaiet al\.\(2026\)A\. Asai, J\. He, R\. Shao, W\. Shi,et al\.Synthesizing scientific literature with retrieval\-augmented language models\.Nature650,pp\. 857–863\.External Links:[Document](https://dx.doi.org/10.1038/s41586-025-10072-4)Cited by:[§1](https://arxiv.org/html/2609.27297#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1)\.
- Asaiet al\.\(2024\)A\. Asai, J\. He, R\. Shao, W\. Shi, A\. Singh, J\. C\. Chang, K\. Lo, L\. Soldaini, S\. Feldman, M\. D’Arcy, D\. Wadden, M\. Latzke, M\. Tian, P\. Ji, S\. Liu, H\. Tong, B\. Wu, Y\. Xiong, L\. Zettlemoyer, G\. Neubig, D\. Weld, D\. Downey, W\. Yih, P\. W\. Koh, and H\. HajishirziOpenScholar: synthesizing scientific literature with retrieval\-augmented lms\.arXiv preprint arXiv:2411\.14199\.Cited by:[§5\.1](https://arxiv.org/html/2609.27297#S5.SS1.p3.1)\.
- Baulinet al\.\(2025\)V\. Baulin, A\. Cook, D\. Friedman, J\. Lumiruusu, A\. Pashea, S\. Rahman, and B\. WaldeckThe discovery engine: a framework for AI\-driven synthesis and navigation of scientific knowledge landscapes\.arXiv preprint arXiv:2505\.17500\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2505.17500)Cited by:[§2\.3](https://arxiv.org/html/2609.27297#S2.SS3.p2.1)\.
- Belhajjameet al\.\(2015\)K\. Belhajjame, J\. Zhao, D\. Garijo, M\. Gamble,et al\.Using a suite of ontologies for preserving workflow\-centric research objects\.Journal of Web Semantics32,pp\. 16–42\.External Links:[Document](https://dx.doi.org/10.1016/j.websem.2015.01.003)Cited by:[§2\.3](https://arxiv.org/html/2609.27297#S2.SS3.p1.1),[§4\.3](https://arxiv.org/html/2609.27297#S4.SS3.p1.1)\.
- Clarket al\.\(2014\)T\. Clark, P\. N\. Ciccarese, and C\. A\. GobleMicropublications: a semantic model for claims, evidence, arguments and annotations in biomedical communications\.Journal of Biomedical Semantics5,pp\. 28\.External Links:[Document](https://dx.doi.org/10.1186/2041-1480-5-28)Cited by:[§1\.1](https://arxiv.org/html/2609.27297#S1.SS1.p2.1),[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p2.1),[§3\.2](https://arxiv.org/html/2609.27297#S3.SS2.p1.1)\.
- Cohanet al\.\(2020\)A\. Cohan, S\. Feldman, I\. Beltagy, D\. Downey, and D\. WeldSPECTER: document\-level representation learning using citation\-informed transformers\.InProceedings of ACL,pp\. 2270–2282\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.207)Cited by:[§1](https://arxiv.org/html/2609.27297#S1.p2.1),[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Daoet al\.\(2022\)T\. Dao, D\. Y\. Fu, S\. Ermon, A\. Rudra, and C\. RéFlashAttention: fast and memory\-efficient exact attention with io\-awareness\.arXiv preprint arXiv:2205\.14135\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2205.14135)Cited by:[§4\.4](https://arxiv.org/html/2609.27297#S4.SS4.p4.1)\.
- Dao \(2023\)T\. DaoFlashAttention\-2: faster attention with better parallelism and work partitioning\.arXiv preprint arXiv:2307\.08691\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2307.08691)Cited by:[§4\.4](https://arxiv.org/html/2609.27297#S4.SS4.p4.1)\.
- Duet al\.\(2026\)Y\. Du, T\. Jin, J\. Kang, X\. Pang, J\. Chai, T\. Miao, F\. Liu, W\. Wang, S\. Yao, Y\. Zhang,et al\.Towards self\-evolving agentic literature retrieval\.arXiv preprint arXiv:2605\.14306\.External Links:[Link](https://arxiv.org/abs/2605.14306)Cited by:[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2609.27297#S5.SS1.p2.1)\.
- Edgeet al\.\(2024\)D\. Edge, H\. Trinh, N\. Cheng, J\. Bradley, A\. Chao, A\. Mody, S\. Truitt, D\. Metropolitansky, R\. O\. Ness, and J\. LarsonFrom local to global: a graph RAG approach to query\-focused summarization\.arXiv preprint arXiv:2404\.16130\.External Links:[Link](https://arxiv.org/abs/2404.16130)Cited by:[§2\.4](https://arxiv.org/html/2609.27297#S2.SS4.p1.1)\.
- Freedman \(2021\)W\. L\. FreedmanMeasurements of the hubble constant: tensions in perspective\.arXiv preprint arXiv:2106\.15656\.External Links:[Link](https://arxiv.org/abs/2106.15656)Cited by:[Figure 1](https://arxiv.org/html/2609.27297#S1.F1)\.
- Gaoet al\.\(2023\)T\. Gao, H\. Yen, J\. Yu, and D\. ChenEnabling large language models to generate text with citations\.InProceedings of EMNLP,External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.398)Cited by:[§2\.4](https://arxiv.org/html/2609.27297#S2.SS4.p1.1)\.
- Grothet al\.\(2010\)P\. Groth, A\. Gibson, and J\. VelteropThe anatomy of a nanopublication\.Information Services & Use30,pp\. 51–56\.External Links:[Document](https://dx.doi.org/10.3233/ISU-2010-0613)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p2.1)\.
- Guoet al\.\(2025\)Z\. Guo, L\. Xia, Y\. Yu, T\. Ao, and C\. HuangLightRAG: simple and fast retrieval\-augmented generation\.InFindings of EMNLP,pp\. 10746–10761\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.568)Cited by:[§2\.4](https://arxiv.org/html/2609.27297#S2.SS4.p1.1)\.
- Heet al\.\(2025\)Y\. He, G\. Huang, P\. Feng, Y\. Lin, Y\. Zhang, H\. Li, and W\. EPaSa: an LLM agent for comprehensive academic paper search\.InProceedings of ACL,pp\. 11663–11679\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.572)Cited by:[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1)\.
- Jainet al\.\(2020\)S\. Jain, M\. van Zuylen, H\. Hajishirzi, and I\. BeltagySciREX: a challenge dataset for document\-level information extraction\.InProceedings of ACL,pp\. 7506–7516\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.670)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Jaradehet al\.\(2019\)M\. Y\. Jaradeh, A\. Oelen, K\. E\. Farfar, M\. Prinz, J\. D’Souza, G\. Kismihok, M\. Stocker, and S\. AuerOpen research knowledge graph: next generation infrastructure for semantic scholarly knowledge\.arXiv preprint arXiv:1901\.10816\.External Links:[Link](https://arxiv.org/abs/1901.10816)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Jinet al\.\(2019\)Q\. Jin, B\. Dhingra, Z\. Liu, W\. W\. Cohen, and X\. LuPubMedQA: a dataset for biomedical research question answering\.InProceedings of EMNLP\-IJCNLP,pp\. 2567–2577\.External Links:[Document](https://dx.doi.org/10.18653/v1/D19-1259)Cited by:[§5\.1](https://arxiv.org/html/2609.27297#S5.SS1.p3.1)\.
- Karpukhinet al\.\(2020\)V\. Karpukhin, B\. Oguz, S\. Min, P\. Lewis, L\. Wu, S\. Edunov, D\. Chen, and W\. YihDense passage retrieval for open\-domain question answering\.InProceedings of EMNLP,pp\. 6769–6781\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.550)Cited by:[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1)\.
- Kulkarniet al\.\(2018\)C\. Kulkarni, W\. Xu, A\. Ritter, and R\. MachirajuAn annotated corpus for machine reading of instructions in wet lab protocols\.InProceedings of NAACL\-HLT,pp\. 97–106\.External Links:[Document](https://dx.doi.org/10.18653/v1/N18-2016)Cited by:[§2\.3](https://arxiv.org/html/2609.27297#S2.SS3.p1.1),[§4\.3](https://arxiv.org/html/2609.27297#S4.SS3.p1.1)\.
- Lewiset al\.\(2020\)P\. Lewis, E\. Perez, A\. Piktus, F\. Petroni,et al\.Retrieval\-augmented generation for knowledge\-intensive nlp tasks\.InAdvances in Neural Information Processing Systems,Vol\.33\.External Links:[Link](https://arxiv.org/abs/2005.11401)Cited by:[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1)\.
- Loet al\.\(2020\)K\. Lo, L\. L\. Wang, M\. Neumann, R\. Kinney, and D\. WeldS2ORC: the semantic scholar open research corpus\.InProceedings of ACL,pp\. 4969–4983\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.acl-main.447)Cited by:[§1](https://arxiv.org/html/2609.27297#S1.p2.1),[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Luanet al\.\(2018\)Y\. Luan, L\. He, M\. Ostendorf, and H\. HajishirziMulti\-task identification of entities, relations, and coreference for scientific knowledge graph construction\.InProceedings of EMNLP,pp\. 3219–3232\.External Links:[Document](https://dx.doi.org/10.18653/v1/D18-1360)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Mirzaet al\.\(2025\)A\. Mirza, N\. Alampara, S\. Kunchapu, M\. Rios\-Garcia,et al\.A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists\.Nature Chemistry17,pp\. 1027–1034\.External Links:[Document](https://dx.doi.org/10.1038/s41557-025-01815-x)Cited by:[§5\.1](https://arxiv.org/html/2609.27297#S5.SS1.p3.1)\.
- Planck Collaboration \(2020\)Planck CollaborationPlanck 2018 results\. VI\. cosmological parameters\.Astronomy & Astrophysics641,pp\. A6\.External Links:[Document](https://dx.doi.org/10.1051/0004-6361/201833910)Cited by:[Figure 1](https://arxiv.org/html/2609.27297#S1.F1)\.
- Singhet al\.\(2023\)A\. Singh, M\. D’Arcy, A\. Cohan, D\. Downey, and S\. FeldmanSciRepEval: a multi\-format benchmark for scientific document representations\.InProceedings of EMNLP,pp\. 5548–5566\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.338)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p1.1)\.
- Skarlinskiet al\.\(2024\)M\. D\. Skarlinski, S\. Cox, J\. M\. Laurent, J\. D\. Braza, M\. Hinks, M\. J\. Hammerling, M\. Ponnapati, S\. G\. Rodriques, and A\. D\. WhiteLanguage agents achieve superhuman synthesis of scientific knowledge\.arXiv preprint arXiv:2409\.13740\.External Links:[Link](https://arxiv.org/abs/2409.13740)Cited by:[§1](https://arxiv.org/html/2609.27297#S1.p2.1),[§2\.1](https://arxiv.org/html/2609.27297#S2.SS1.p1.1)\.
- Songet al\.\(2022\)N\. Song, H\. Cheng, H\. Zhou, and X\. WangLinking scholarly contents: the design and construction of an argumentation graph\.Knowledge Organization49\(4\),pp\. 213–235\.External Links:[Document](https://dx.doi.org/10.5771/0943-7444-2022-4-213)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p2.1),[§3\.2](https://arxiv.org/html/2609.27297#S3.SS2.p1.1)\.
- Swanson \(1986\)D\. R\. SwansonUndiscovered public knowledge\.The Library Quarterly56\(2\),pp\. 103–118\.External Links:[Document](https://dx.doi.org/10.1086/601720)Cited by:[§1](https://arxiv.org/html/2609.27297#S1.p1.1),[§2\.3](https://arxiv.org/html/2609.27297#S2.SS3.p1.1),[§6\.2](https://arxiv.org/html/2609.27297#S6.SS2.p1.1)\.
- Teufel and Moens \(2002\)S\. Teufel and M\. MoensSummarizing scientific articles: experiments with relevance and rhetorical status\.Computational Linguistics28\(4\),pp\. 409–445\.External Links:[Document](https://dx.doi.org/10.1162/089120102762671936)Cited by:[§2\.2](https://arxiv.org/html/2609.27297#S2.SS2.p2.1),[§3\.2](https://arxiv.org/html/2609.27297#S3.SS2.p1.1)\.
- Vaswaniet al\.\(2017\)A\. Vaswani, N\. Shazeer, N\. Parmar, J\. Uszkoreit, L\. Jones, A\. N\. Gomez, L\. Kaiser, and I\. PolosukhinAttention is all you need\.arXiv preprint arXiv:1706\.03762\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.1706.03762)Cited by:[Figure 2](https://arxiv.org/html/2609.27297#S3.F2.2),[Figure 2](https://arxiv.org/html/2609.27297#S3.F2.3),[§3\.2](https://arxiv.org/html/2609.27297#S3.SS2.p4.1)\.
- Waddenet al\.\(2020\)D\. Wadden, S\. Lin, K\. Lo, L\. L\. Wang, M\. van Zuylen, A\. Cohan, and H\. HajishirziFact or fiction: verifying scientific claims\.InProceedings of EMNLP,pp\. 7534–7550\.External Links:[Document](https://dx.doi.org/10.18653/v1/2020.emnlp-main.609)Cited by:[§2\.3](https://arxiv.org/html/2609.27297#S2.SS3.p1.1),[§4\.4](https://arxiv.org/html/2609.27297#S4.SS4.p1.1)\.
- Wanget al\.\(2020\)S\. Wang, B\. Z\. Li, M\. Khabsa, H\. Fang, and H\. MaLinformer: self\-attention with linear complexity\.arXiv preprint arXiv:2006\.04768\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2006.04768)Cited by:[§4\.4](https://arxiv.org/html/2609.27297#S4.SS4.p4.1)\.
- Wanget al\.\(2024\)X\. Wang, Z\. Hu, P\. Lu, Y\. Zhu, J\. Zhang, S\. Subramaniam, A\. R\. Loomba, S\. Zhang, Y\. Sun, and W\. WangSciBench: evaluating college\-level scientific problem\-solving abilities of large language models\.InProceedings of ICML,Proceedings of Machine Learning Research, Vol\.235,pp\. 50622–50649\.External Links:[Link](https://proceedings.mlr.press/v235/wang24z.html)Cited by:[§5\.1](https://arxiv.org/html/2609.27297#S5.SS1.p3.1)\.

## Appendix ALKM Interfaces and Agent Integration

### A\.1LKM Website

The public LKM site exposes four complementary components\.[Knowledge Explorer](https://lkm.bohrium.com/web/en)searches claims and questions, opens reasoning chains, and shows the source\-paper graph\.[Evidence Engine](https://lkm.bohrium.com/web/en/evidence-engine)retrieves evidence that supports or challenges a research question and produces a cited synthesis\.[Science Landscape](https://lkm.bohrium.com/web/en/landscape)organizes common\-question clusters by discipline\.[Workflows](https://lkm.bohrium.com/web/en/workflows)presents reproducible workflow families derived from reasoning chains\. All four operate on source\-grounded objects and preserve the link to the originating paper\.

### A\.2Integration Modes

Agents can access LKM in three ways\. The[OpenAPI](https://s.apifox.cn/33d12311-ec59-4a5c-a849-391704fe7f84)provides direct HTTP access and is appropriate for production services and custom orchestration\. The[bohr CLI](https://docs.bohrium.com/docs/bohrctl/install/)offers shell\-level access for research scripts, batch jobs, and reproducible command\-line workflows\.Bohrium skillsare discovered withbohr skills listand expose the same operations as tool calls inside an agent runtime, handling request construction and response decoding\. These modes share the same identifiers and semantics; the choice is an integration concern rather than a different knowledge source\.

### A\.3LKM OpenAPI

For/search, valid scopes areabstract,claim,premise,conclusion,question,problem,open\_question,subproblem, andreasoning\_chain\.hybridcombines semantic and lexical retrieval;semanticfavors speed;lexicaltargets exact terms\. A successful response hascode=0; hits are indata\.variables, paper metadata indata\.papers, and pagination is indicated byhas\_more\. Scores are ranking signals, not scientific confidence\.