Human reviewers stopped rewarding lexical complexity as LLMs made it cheap
A frozen-rater design separates preference drift from composition change across 124K reviews
cs.DL
Digital Libraries
Covers all aspects of the digital library design and document and text creation. Note that there will be some overlap with Information Retrieval (which is a separate subject area). Roughly includes material in ACM Subject Classes H.3.5, H.3.6, H.3.7, I.7.
sort pith recommended most recent
Radical cross-field recombination is shrinking in modern science
Citations spanning fully separate subfields decline while overlapping-field recombination grows
· “The conservative turn in science: The changing character of knowledge recombination”
Economic funnel and data-level novelty power AI culture harvesting
Formal five-stage pipeline uses cheap filters before expensive AI, turning deduplication into data deepening
TEXTWAL+LaTeX recovers 88% of URLs from arXiv papers
Systematic study across six formats gives practical advice for link rot analysis, knowledge graph building, and reproducibility
· “URL Extraction from Scholarly Documents: A Cross-Format Comparative Analysis”
Co-evolution loop prevents drift between ontology and data
X-DigCheck provides a domain-independent environment for continuous profile–graph alignment across regeneration cycles.
One field's solver finds another's without sharing words
A domain-stripped computational skeleton surfaces same-problem papers in unrelated fields, lifting retrieval precision from 0.22 to 0.56.
33.2 hours of correction beat 97% more data on medieval Latin parser
An expert correcting machine pre-annotations for 33 hours produced a parser that outperforms all existing Latin treebank models.
· “Correction as Annotation: Bootstrapping a Dependency Parser for Documentary Medieval Latin”
1,152 Vietnamese-diaspora scholars
Students, collaborators, and event organizers get a searchable, evidence-checked way to find diaspora academics.
· “VietProfs: A Public Directory of the Vietnamese Academic and Research Diaspora”
Human outlines and reviews lift automated survey quality above baselines
Multi‑agent system retrieves from multiple databases, re‑ranks by impact, and uses real peer‑review comments to revise drafts, achieving…
Math preprints up 33.5% in 2026
A synthetic control of 18 other arXiv archives isolates a math-specific surge tied to AI capability thresholds.
· “The Generative AI Gold Rush in Theoretical and Computational Research”
Three kingdoms dominate philanthropic and nonprofit studies
Citation networks reveal weakly connected power centers of nonprofits, social movements, and voluntary action
· “Constructing the Field of Philanthropic and Nonprofit Studies: Evidence from Citation Networks”
Thin metadata layer chains provenance and DOIs to experiment data
Machine-readable records captured at collection time let AI agents find and judge data too large to move.
MultiGhostBench: no single method wins multilingual LLM attribution
New 928-book benchmark across six languages tests domain, author, and language transfer; Transformer detectors cross languages but with…
Score anchors steer LLM reviews
LLMs cover human strength points more reliably than weakness points when guided by an external score.
· “Guiding LLM Peer Reviewers: The Impact of Score Anchors on Review Evidence and Accuracy”
LLM reviews amplify argument
Five theory-guided lenses on 11,560 ICLR reviews show AI and human critique redistribute evaluative work.
· “Beyond Human-Likeness: Mapping the Scientific Critique Profiles of LLMs and Human Reviewers”
LLMs warm scientific citations
Six models consistently replace critical citations with support, cite older popular work, and ignore authors' collaborators.
· “Citing Less Critically: LLMs Reshape the Rhetoric and Reach of Scientific Citation”
250 years of math reviews become 168 million queryable links
A new open graph joins expert-curated math semantics to centuries of publications, beyond citation networks.
· “The zbMATH Open Knowledge Graph: Tracing Centuries of Mathematical Research”
LLM quality scores tilt toward viruses
In 73,489 health-science articles, two models consistently favoured molecular research over surveys and clinical studies.
Paper-mill research reaches policy
Study of 1,975 Pharmakon Network papers shows cited outputs persist and authors keep funding after exposure.
· “Tracing high-profile attention to questionable research as a case for funder due diligence”
VLM pipeline extracts structured auction data at 87% accuracy
A new 1,378-lot benchmark compares commercial, institutional, and local VLM deployments for historical lot extraction.
· “Lot Machine: Multimodal Lot Extraction from Auction Catalogs”
Citation impact becomes pitch: tool sonifies research performance
SoniMet maps field-weighted citation counts to auditory parameters, adding an audio channel to bibliometric analysis for individual…
· “SoniMet - A tool for sonifying and visualizing the performance of single researchers”
56 papers compiled into a stable knowledge map by LLM
ASKS system yields a source-traceable research portrait with zero splits and low membership churn in a longitudinal demonstration.
Filling conceptual holes in science draws outsized attention
Persistent homology of concept embeddings reveals two asymmetries: gaps pay off in empirical fields, yet their number explodes while…
· “Filling holes in science draws collective attention, but most higher-order holes remain unexplored”
Only 6.1% of open AI models are curatable today
Across 191,375 Hugging Face repos and 2,214 papers, visibility far outpaces the evidence libraries need.
China's home citation bias shrinks over two decades
A global citation network analysis reveals modest domestic preference and convergence with US in disruptive research impact.
Nordic health catalogues fail every required EU metadata property
Measurement of 2,811 health dataset descriptions finds none meets all eight mandatory fields of HealthDCAT-AP Release 7
· “Measuring the Installed Base: Nordic Health Dataset Catalogues Against HealthDCAT-AP Release 7”
NPMI drop metric recovers patent gaps with high specificity
BLANC compares cluster associations before and after keyword filtering, flagging 'established globally, unexplored locally' at ~30%…
AI writing boosts disruption but narrows knowledge sources after 2023
Two million papers show positive link to disruptive citations yet weaker cross-field recombination and narrower inputs.
· “AI-assisted writing and the reorganization of scientific knowledge”
Framework certifies AI research by separating validity from authorship
Two-layer system assesses knowledge claims and human input levels to work within existing journal systems.
· “Rethinking Publication: A Certification Framework for AI-Enabled Research”
Full OA aligns closer to patents than hybrid OA
Patents cite hybrid models more often, but gold and diamond OA show stronger semantic links, especially inside patent bodies.
· “Discoverability matters: Open access models and the translation of science into patents”
Open scholarly infrastructures get a four-layer computational blueprint
Maps open-infrastructure principles onto a concrete hardware-to-application stack.
· “Towards a Definition of the Computational Architecture of Open Scholarly Infrastructures”
LLM pipeline turns flat heritage records into FAIR Digital Object graphs
On 637 records it links 86% of metadata slots; the missing piece is human-verified correctness.
· “Automated Construction of FAIR Digital Object Knowledge Graphs from Flat Cultural Heritage Records”
118,000 microscope events reveal how experiments really run
Beyond data and metadata: how an instrument was used is itself preserve-worthy science.
· “Beyond FAIR Data: Instrument Traces for Active and Autonomous Scientific Experimentation”
One matrix makes taxonomy-aware topic distances exact
A rooted taxonomy's height function fully determines a metric on topic profiles, computable in linear time.
· “Taxonomy-aware distances between scholarly topic profiles via an exact simplex embedding”
40 million affiliations now map to 638,843 sub-units
The new open dataset adds department-level detail and a validated parent-child hierarchy for universities.
MeGraS turns media files into queryable graph nodes
Images, video, and audio become first-class nodes, so SPARQL can search and segment them directly.
· “MediaGraph: A Content-Aware Data Model and Query Framework for Multimodal Knowledge Graphs”
Leibniz chokepoint splits 900 years of math lineage
Tracing 64 Fields Medalists back 57 generations, the paper finds 84.4% of lineages converge on five medieval scholars.
5 of 6 research projects contradict their own citation records
An audit of 117 open-source projects finds their own metadata surfaces often contradict each other.
· “A Multi-Surface Consistency Audit of Software Citation Metadata”
A field's birth is visible as a drop in its knowledge-map modularity
Across three test areas, the knowledge base's modularity falls as the field forms and can later rebound.
· “Declining Modularity of Intellectual Bases During the Emergence of Research Areas”
Annual APC data for 12,540 journals from 14 publishers now open
Year-by-year list prices from 2019–2025 make open-access fee trends and publisher comparisons easy to compute.
· “A dataset of article processing charges from 14 scholarly publishers, 2019-2025”
A 1,951-figure benchmark maps how AI should read scientific images
Four expert-scored tasks move from spotting axes to judging evidence, with Bloom-informed questions scoring each level of understanding.
· “A Pathway to General-Purpose Scientific AI: Multimodal Comprehension of Scientific Images”
New corpus maps 500 MPB songs across melody
Ten iconic composers, 17,053 chords, 23,447 notes—enough to test what defines MPB's shared style.
Co-led teams drive novelty at large-scale science facilities
Across 273,109 papers from 76 facilities, novelty rises when in-house scientists co-lead with external users.
· “Co-leading Teams Drive Scientific Novelty in Large-scale Research Infrastructures”
Explainable transfer learning targets earlier Alzheimer's diagnosis
It consolidates models, accuracy ranges, and open challenges so new AI researchers can enter AD diagnosis.
Frontier LLMs miss up to 63% of wrong-page legal cites
Even the best model misses 40% of pinpoint-page errors in opinions, so AI citechecking cannot verify support.
No backend: house history runs on Wikidata and OpenHistoricalMap
A map tool lets any genealogist add sourced building records straight into the global open knowledge graph.
· “Domus: An Open-Data Web Platform for House History Research”
Knowledge-base grading inflates translation scores by up to 2.7x
Across four models, the gain appears only where the grading key shares the injected resource.
Pure science loses NSERC ground as medical
A 24-year look at Canadian grant data finds medical sciences climbing while chemistry and space science fall.
· “Patterns of Research Funding Across Research Subjects: The Case of NSERC”
Peace research's own words vary by discipline and country
Keyword analysis of 113,690 SDG-16 papers finds welfare language in high-peace countries, security language in low-peace ones.
Pre-registered No-Go halted the authors' own confirmatory claim
21-card protocol frozen before production made a No-Go boolean halt the work; outsiders can recompute it all.
Matched test: broad scope alone flags delisted venues
A PageRank-style journal metric tracks impact while withstanding fake-citation attacks about ten times better than raw counts.
Network scans could expose predatory journals
A review argues coordinated publishing fraud shows up in shared authors and citation rings that single-journal checks miss.
89% of biomedical papers carry LLM writing markers by 2025
A word-frequency analysis of 1.2 million papers finds LLM editing in nearly all, led by Discussion sections.
· “Most biomedical publications show signs of LLM-assisted writing”
Webcam eye-tracking lifts keyphrase extraction from Chinese abstracts
Fixation number plus total fixation duration gives the best F1 gains on Chinese academic abstracts.
· “Leveraging Human Reading Behavior for Keyphrase Extraction: A Webcam-based Eye-tracking Corpus”
Scholar assessment splits into six auditable evidence trails
HexEval scores intrinsic quality and external behavior separately, so each rating can be traced and checked.
· “HexEval: An Evidence-Driven Hexagonal Framework for Multidimensional Scholar Assessment”
Male first authors edge ahead in ChatGPT quality scores
Author names are withheld, yet a small male tilt remains across most UK research fields, an 89,744-article study finds.
· “Does ChatGPT score research quality differently by gender?”
189,101 statistics papers, four citation networks, one public dataset
Built from 62 journals, StatCite pairs paper metadata with co-citation, coupling, and journal-level graphs for public analysis.
· “StatCite: A Large-scale Citation Network Dataset for Statistics and Data Science”
LLM and graph pipeline adds 113,892 concepts to OpenAlex
Keyword clusters become named concepts covering 98.8% of OpenAlex topics, rated ~89% relevant by experts.
· “SCALE: Scientific Concept Aggregation via LLMs and Embeddings for Fine-Grained Taxonomy Extension”
Congress's press releases doubled their em-dashes in 2025
A preregistered scan of 146,239 releases finds the punctuation shift came late and broad, with caveats.
One small set of climate-health pairings dominates 22,695 studies
Extreme heat and flood-hurricane mental health pairs recur far beyond chance, even after adjusting for term popularity.
LODE 2.0 rebuilds ontology docs for modular knowledge graphs
Keeps the classic single-page style, adds per-entity RDF provenance, Markdown rendering, and static export.
New repository standardizes 83 higher-order network datasets
Each dataset links to its source, citation guidance, conversion code, and fixed version snapshots for reproducible reuse.
· “A repository for discovery and reuse of higher-order network datasets”
AI upscaling adds no accuracy to thermal 3D scans
Six resolution tiers on Florence's Loggia: native wins; AI detail fails multi-view checks.
· “Loggia dei Lanzi: AI Thermography Enhancement Comparisons through 3D Photogrammetry”
BIP! Ranker computes eight citation metrics on billion-edge graphs
One distributed library spans influence, popularity, momentum, and field-normalised scores, running in minutes on 2.1 billion citations.
· “BIP! Ranker: A Software Library for Citation-Based Impact Indicators on Large-Scale Graphs”
LLM productivity boost survives artifact check
Rate-matched random placebo still leaves a positive gap, and four designs that avoid the timing artifact agree.
16 schemas turn lab methods into comparable machine data
Expert-finished field lists for biology, materials, imaging, physics, and psychology—ready for extraction and knowledge graphs.
18 centuries of citation style held steady in Chinese histories
A span-grounded AI extractor found 5,766 Analects reuses across 24 histories—wording drifted while practice held.
LLM literature screening is maturing faster than LLM retrieval
A 34-paper survey finds rapid growth, a heavy biomedical focus, and little direct evaluation of retrieval tools.
· “Scientific Knowledge Discovery in the Age of Large Language Models”
Peer review favors top institutions—until outcomes are checked
At equal reviewer scores, borderline papers without a top-25 author lose acceptance points, yet show no better downstream results.
· “Bias at the Borderline: Who Gets the Benefit of the Doubt in Peer Review?”
ChatGPT ranks articles as reliably as individual reviewers
In a 200-article study, ChatGPT's averaged rankings matched or beat single human experts—but detailed PDF critiques did not improve scores.
Same field lands in different buckets on every platform
Four bibliometric platforms classify by different mixes of experts, AI, citations, and granularity—know the scheme before you search.
Nearly 1 in 5 journal articles now cite preprints
Even preprints never peer-reviewed are cited as often as published ones — and the trend is exponential since 2014.
· “Preprints Without Curation Are Increasingly Cited by Journals”
Tiny unsupervised training beats labeled baselines on ancient text reuse
Four to eight thousand raw Latin or Greek sentences turn token models into strong reuse detectors and retrievers.
Plain language unlocks curated math-model graphs
Schema retrieval and Steiner join plans keep language-model answers tied to expert knowledge.
Word spotting lets graph agents beat RAG on noisy archives
Executable Cypher plus OCR-variant expansion recovers counts and multi-hop facts standard retrieval misses.
Plain text that is the jeongganbo score
A spatial text format and automatic traditional layout finally make Korean mensural notation editable data, not hand-built images.
Pipeline mines 13,740 X-ray absorption spectra from battery papers
Automated figure-and-text reading makes literature-embedded spectra machine-readable for training and cross-lab use.
· “Harnessing X-ray Absorption Spectroscopy Data through Multimodal Mining of Battery Literature”
MLIP papers become queryable data via one ontology
Methods, DFT training data, and benchmark results reduce to a single SPARQL query instead of a manual literature read.
New z-index turns retractions into scientific debt
Adding impact factors of retracted papers gives a net-value score, h-z, that can go deeply negative.
Only 4 of 984 papers report on government software engineering
Mainstream venues published almost nothing on how public agencies build software — a gap for evidence-based practice.
