JCDL '26: Proceedings of the 26th ACM/IEEE Joint Conference on Digital Libraries

Scoring open-ended creativity is expensive. The gold standard for evaluating open-ended creativity responses, including divergent thinking tasks such as the Alternate Uses Task (AUT), is the Consensual Assessment Technique (CAT), where panels of human experts rate creative products. CAT is resource-intensive, limiting the extent to which institutions can measure creativity. Advances in large language models offer an alternative: a CAT-inspired persona panel using large language models (LLMs) as judges to score open-ended responses automatically. Our selected large language models, gpt-5.5, gpt-5-mini, and Claude Haiku 4.5, were assigned expert personas, and we aggregated each model's judgments and compared them against two controls: a self-consistency control using repeated judgments from a single generic judge, and a basic prompted baseline. On gpt-5-mini, the CAT-inspired persona panel improves over a single prompted judge ( r=0.48 to 0.55). It does not outperform the self-consistency control, suggesting that the benefit of the CAT-inspired persona panel is aggregation rather than persona diversity. Accuracy improves substantially with many-shot grounding: supplying many labeled in-context exemplars nearly closes the gap to Ocsai, the fine-tuned reference, while persona diversity adds nothing measurable. For digital libraries and learning repositories that collect open-ended submissions at scale, scalable scoring methods are especially consequential. 

Research conceptualization involves the process of synthesizing existing knowledge, identifying research gaps, generating novel connections, and formulating testable questions, which is the most cognitively demanding stage of research training. Students struggle in this process, facing difficulties in selecting research topics and generating ideas from multiple perspectives. According to social learning theory, such skills are normally acquired by observing instructors, mentors, or peers work through problems and internalize their research strategies. However, human experts are constrained by individual cognitive boundaries, and one-on-one mentoring is difficult to scale, leaving students with scarce opportunities to observe diverse exemplary thinking. Recent advances in LLMs with reasoning capabilities offer a way to close this gap. AI can serve as a thinking partner whose reasoning processes can inspire human cognitive development. We developed InspAIre, a conversational system that makes AI reasoning visible and interactive during research conceptualization, and evaluated it in two experiments. In Study 1, a within-subject experiment with 72 university students showed significant improvements in research conceptualization performance after observing and interacting with AI reasoning. In Study 2, a between-subject experiment with 76 university students showed that students who observed AI reasoning significantly outperformed those receiving direct AI answers, and that inspiration mediated this effect. These findings suggest that observable AI reasoning and interactive design can effectively enhance students’ research capabilities, transforming AI from answer providers into collaborative cognitive partners in educational contexts.

Named-entity recognition (NER) on historical and OCR-degraded text remains a difficult operating point for both supervised taggers and large language models (LLMs). Specialist encoders are fast and accurate in-domain but require task- and domain-specific training; frontier LLMs are strong general extractors, but their cost, deployment constraints, and run-to-run behaviour may be unsuitable for sovereign or high-throughput knowledge-graph construction pipelines. We ask whether structured orchestration can make an open-weight LLM operationally competitive for NER while preserving the cost, latency, stability, and sovereignty advantages that motivate open deployment. We study SOLE, a structured zero-shot architecture (specified extraction prompt, deterministic post-processing, entity-centric resolution) over an interchangeable substrate. Raw open-weight extraction collapses on historical OCR (0.216 partial-mention F1, zero predictions on 71.6% of documents); the same substrate under SOLE reaches about 0.78 entity-level F1. On HIPE-2020 French, SOLE on an open-weight substrate reaches per-mention parity with a frontier reference (0.818 vs 0.813, within run-to-run noise) at roughly 6.5 × lower cost and 3.3 × faster median latency, with several times tighter variance. The residual gap is confined to entity-level scoring; a substrate-swap decomposition shows it is corpus-dependent — roughly half substrate and half architecture on heavily degraded text, but essentially all architecture on cleaner text, where the frontier substrate adds no measurable quality. A stage-level pipeline autopsy localizes the architecture gap to never-emitted entities rather than resolver errors, and negative probes (intra-document rescan, prompt ensembling, higher reasoning effort) do not recover it. SOLE is not a frontier replacement; it is an operational point that makes open-weight NER cheap, stable, fast, and sovereign while making the remaining quality gap explicit and localized.

The BIBFRAME ontology, introduced by the Library of Congress, is a bibliographic metadata standard of classes and properties for describing library collections. Unlike its predecessor MARC, the BIBFRAME standard is described in the Web Ontology Language (OWL), a fundamental component of linked data. Metadata professionals struggle to effectively create BIBFRAME records without training in linked data principles. To address this sociotechnical challenge, this paper describes the BF Vibe system, which provides support for semi-automated metadata creation. BF Vibe contextualizes technical requirements with natural language explanations backed by authoritative Library of Congress sources. Grounded in social signaling theory, BF Vibe employs honest signaling through validation loops that check outputs against community SHACL shapes and ground repairs in Library of Congress exemplars, reducing the cost of probing, the manual effort of checking a metadata description against ontology rules and examples. The intervention reduced errors from 6 per description to 1. An interface analysis showed a 23% reduction in predicted cataloger task time.

Scientific data repositories can be FAIR yet unsearchable: column names such as NEE_CUT_REF, SSA637, or lel carry meaning only for specialists, so a non-expert cannot find, filter, or plot the data at the variable level. We generate natural-language descriptions for opaque columns with a language model at ingestion, from their names, statistics, and sample values, and store them as curator-reviewable catalog metadata for search and a natural-language visual-discovery interface. Our clearest result is on findability: across three collections from a live repository, the descriptions make opaque columns retrievable by natural-language query, raising pooled mean reciprocal rank from 0.46 to 0.79, and a control shows the gain comes from the descriptions’ content—generic or shuffled text does not reproduce it. Downstream, in a within-model NL2Vis ablation, the benefit is conditional: enrichment raises translation accuracy in two of six model–collection cells and leaves the rest flat or slightly negative, in a pattern that tracks a column’s opacity to the model and the model’s accuracy headroom, so the single-collection lesson that “enrichment helps weaker models” does not generalize. A quality evaluation against curator references finds description accuracy governed by how familiar the model is with a naming convention; how opaque the name looks to a person does not predict it, so a human-unreadable but standard schema is described well while errors concentrate on unfamiliar specialized vocabulary, where curator review should focus. We release the questions, descriptions, and evaluation harness as a benchmark.

Subject indexing assigns descriptors from a controlled vocabulary to bibliographic records, a task characterized by thousands of labels, few positive examples per label, and a long tail of rarely used descriptors. Existing extreme multi-label classifiers treat the vocabulary as a flat list, ignoring the semantic relations that thesauri encode.

We present TAGHN, a heterogeneous graph re-ranker over a per-document subgraph of retrieved candidates augmented with the thesaurus neighborhood. The relevance of a candidate descriptor is estimated jointly from its similarity with the document and also the degree to which the candidate’s thesaurus neighborhood matches the document content. A decomposed scoring head combines interpretable graph-derived features with retrieval signals.

On STW, an economic thesaurus of 6,198 descriptors, TAGHN reaches 0.380 macro F1 against 0.346 for a fine-tuned cross-encoder and 0.278 for a fixed-vocabulary XMLC classifier. In terms of micro F1, it outperforms re-ranking models and stays competitive (-3.0) with the best fixed-vocabulary XMLC model. Qualitative analysis shows the graph resolving errors that per-label scoring cannot, selecting a narrower gold descriptor where an unstructured re-ranker settles for its broader parent.

An AI-generated podcast companion article to augment the Paper Session 5: Scientific Document Processing Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

URLs in scholarly documents link to rich external resources such as datasets, software, publications, and websites. Extracting these URLs is crucial in the data preparation stage of many downstream tasks, such as link rot analysis, web crawling, and building knowledge graphs. However, existing studies often downplay this phase, simply extracting URLs from a single format, usually text directly converted from PDFs. We present a systematic study evaluating URL extraction across six input formats (text with annotation layer, LaTeX, HTML, XML, Markdown, and PNG converted from PDF). To support the evaluation, we compiled a benchmark dataset consisting of 2,338 manually annotated URLs from 200 arXiv papers spanning a wide range of domains over a 33-year period. In addition to evaluating individual file formats, we also compared 63 composite input-format combinations. Our extensive evaluations indicate that TEXTWAL achieves the best performance among single-format inputs, while TEXTWAL+LaTeX achieves the best overall URL extraction performance. The same trend is observed for URLs linking to open-access datasets and software. To further validate these findings, we apply our format-specific URL extraction pipelines to a longitudinal random sample of 364,744 arXiv papers spanning 33 years. We observe a sharp increase in URL density after 2015, along with remarkable differences in URL extraction across file formats over time. Overall, our study highlights the importance of selecting an appropriate format for URL extraction from scholarly documents. The dataset and code are publicly available at: https://github.com/lamps-lab/arxiv-url-bench.

Sequential sentence classification (SSC) is an essential task for structuring scientific publications, and extending SSC research to languages other than English can improve accessibility to scientific knowledge in multilingual digital libraries. Cross-lingual transfer is a promising approach to address the scarcity of training data in non-English languages. Prior work on other natural language processing tasks has shown the benefits of capturing linguistic similarity between source and target languages. However, SSC inherently depends on patterns at the discourse level, such as label sequences and positional regularities, which appear consistently across languages regardless of linguistic differences. To examine the factors that determine transfer success in SSC, we constructed a multilingual SSC dataset covering 13 non-English languages collected from five academic databases. Our cross-lingual transfer experiments, using both encoder-based and generative models, show that linguistic proximity has no consistent predictive power for transfer performance, whereas structural similarity in rhetorical organization shows a weak but consistent positive correlation across models. After controlling for source-language performance, the similarity of label distributions is the most consistent predictor. Building on this finding, we propose a set of three methods that explicitly leverage structural information using generative models. In the in-domain evaluation, the best combination reaches parity with the strongest encoder baselines, and in transfer to languages unseen during training, it outperforms the strongest encoder baseline.

Accurate table recognition requires complete recognition of cell content and logical structure, which provides a foundation for table understanding and reasoning. However, complex table recognition remains challenging: existing training datasets contain limited complex examples and insufficient diversity, while commonly used similarity metrics mainly measure local matching and do not directly indicate whether an entire table is functionally correct. In this work, functional correctness means that the evaluated table content, cell order, and logical topology exactly match the reference after canonicalization. To address these problems, we introduce StrucTab, a data resource comprising StrucTab-Train, with approximately 633K bilingual image–HTML pairs, and StrucTab-Bench, a benchmark of 425 real complex scientific tables. We also introduce Table-Level Correctness (TLC), which formalizes this functional-correctness criterion as a canonicalized whole-table exact-match measure. Experiments show that StrucTab-Train effectively improves complex scientific table recognition: after fine-tuning Qwen2.5-VL-7B and Qwen3.5-9B, TLC exceeds the best alternative training dataset by 5.65 percentage points for each backbone. Nevertheless, the highest TLC on StrucTab-Bench is only 24.94%, despite much higher similarity scores. This gap shows that partial structural or content agreement does not guarantee a fully correct table and that complex table recognition still has substantial room for improvement.

Many information needs are difficult to express as keyword queries, since users often lack the domain-specific terminology needed to retrieve relevant documents. We propose a Summaries-as-Queries (SaQ) framework that enables contextual document discovery by transforming a user-provided document into an effective retrieval representation. Rather than relying on a single summarization strategy, SaQ employs a hybrid summarization pipeline that first identifies salient concepts through extractive summarization and then guides abstractive summarization to generate concise, contextually faithful summaries. To improve retrieval coverage beyond a single summary query, the framework decomposes the generated summary into multiple topical subqueries using a large language model (LLM), allowing distinct semantic aspects of the source document to be searched independently. The resulting ranked lists are integrated through rank fusion to produce a unified set of relevant documents. By combining hybrid summarization, LLM-guided query decomposition, and rank fusion, SaQ bridges document understanding and information retrieval, providing an effective framework for contextual document discovery in digital libraries and large document collections. Experimental results on a large dataset demonstrate that SaQ consistently outperforms competitive retrieval baselines. These findings demonstrate that SaQ provides an effective new paradigm for contextual document discovery in digital libraries.

Historical digital libraries hold millions of scanned pages spanning centuries of human knowledge. However, high-fidelity content accessibility through automated question answering remains largely unsolved. Large language models offer a promising path but historical documents pose particular challenges such as OCR noise, archaic spelling, mixed layouts and domain shift from modern training corpora. Moreover, large language models exhibit sharply divergent competence across these challenges. Selecting the wrong model silently degrades retrieval quality at scale, with no easy way for digital library practitioners to know which model to trust for a given document. To address this challenge, we investigate adaptive LLM selection as a core problem for historical digital libraries. We construct a large-scale benchmark of 84,441 multiple-choice questions derived from 5,941 scanned pages of bilingual English and Chinese historical periodicals. We systematically evaluate 50 open-source LLMs ranging from 0.5B to 9B parameters, producing a page-level binary correctness matrix of 50 × 5, 941 entries. Our analysis reveals that model performance is highly page-dependent: the per-page oracle accuracy reaches 86.1% while the static best model achieves only 42.8%. We further show that our routing models include the oracle-best model in the top-3 candidate models for more than 61% of pages, offering a practical path toward efficient deployment. Gains are largest on the most challenging pages, including tabular content, Chinese-language text and hard pages where the dominant model consistently fails. Expert validation further confirms that adaptive selection decisions derived from automatic MCQs are consistent with historian judgement. Our benchmark and methodology establish a reproducible framework for studying and deploying adaptive LLM selection in historical digital library pipelines.

Chinese government work reports form a large civic archive whose formulaic language and uneven historical coverage complicate thematic organization. We index 6,839 central, provincial, and prefectural reports spanning 1954–2025 and analyze 6,316 retained documents and 2.39 million sentences from 1980–2025. We develop EnvTextSim, a Chinese sentence encoder adapted with an adaptive circle loss to an author-defined environmental taxonomy, and a coverage-aware topic-evaluation protocol based on Hungarian soft alignment. On distantly labelled test sentences, intra-/inter-topic cosine similarity is 0.994/0.016, compared with 0.797/0.768 for text2vec-base-chinese and 0.786/0.826 for bge-large-zh-v1.5. In a guided BERTopic run, the adapted encoder recovers 13 of 14 taxonomy topics, compared with seven for text2vec. These results measure taxonomy consistency: training, topic guidance, and alignment all reuse the same dictionary. Keyword masking reduces EnvTextSim’s nearest-centroid accuracy from 0.993 to 0.471, showing substantial dependence on the supervision vocabulary. Separately, dictionary-based panels describe environmental attention and first observed textual appearances across administrative levels. Removing eight broad keywords preserves the overall provincial attention pattern (r = 0.966) but changes several topic-specific timing estimates. Retrieval examples illustrate topical access without establishing ranking quality or usability. The accompanying labelled sentences, analysis panels, taxonomy, and reproducibility materials support further collection research and independent validation.

Research data infrastructures are designed, among other objectives, to improve the discoverability and accessibility of research data. However, this is often accompanied by an implicit assumption that data become perceptible to scholars simply by existing and being made accessible. Building on the premise that visibility is not a stable property of data itself, we argue that the visibility of research data is instead shaped by several conditions that determine whether and how data become perceptible to scholars.

Using the humanities as an illustrative domain, we focus on the modes of access to humanities research data, understood as the ways through which researchers become aware of research data and argue that each access mode is associated with a distinct Dimension of Visibility, shaped by different factors that influence how and whether research data become visible within these Modes of Access. To capture these dynamics, we introduce the Humanities Research Data Visibility Model (HRDVM), a literature-driven model that identifies four central Modes of Access: Social Interaction, Information Systems, Physical Access and Scholarly Literature. For each Mode of Access, the model introduces a corresponding Dimension of Visibility: Social Visibility, Digital Visibility, Physical Visibility, and Literature Visibility. Building on this model, we derive implications for the design of humanities research data infrastructures. These include, among others, supporting multiple modes of access to research data, improving metadata to better represent reuse-relevant information, and accounting for the social and institutional practices through which researchers discover and access data.

Although we focus on the humanities, the proposed framework and the implications can be adapted to other research domains and non-academic contexts, as well as to other information resources.

Libraries worldwide have begun to uncover the presence of arsenic in their collections, posing risks to collections, staff, and the broader public. KB National Library of the Netherlands is at the forefront of addressing this issue, having tested a large sample of items for arsenic. Building on these testing practices and results, this study adopts a digital library and information science approach to explore how Exploratory Data Analysis (EDA) and Machine Learning (ML) can be integrated into library workflows to optimise the detection of toxic materials. Using a mixed method approach, the current workflow is traced via interviews with 12 staff members, revealing key issues including selection bias and fragmentation of data across non-interoperable systems. The final dataset of arsenic test results is then enriched with metadata, including 1,418 items with 9 features per item. EDA reveals patterns including a higher positive arsenic rate among 17th- and 18th-century items and notable regional differences. Random Forest (RF) and Gradient Boosting (GB) classifiers were trained on this dataset, both showing promising recall (70% and 69%, respectively) for arsenic-positive items. Based on these findings, we propose integrating data visualisations and a human-in-the-loop (HITL) ML system into existing workflows to support decision-making. This study demonstrates that digital methods such as EDA and ML applied to metadata can meaningfully support the detection of toxic materials in physical library collections.

Accurate webpage validity timestamp determination is critical for Retrieval-Augmented Generation (RAG) and web archiving, yet existing methods face two persistent limitations: symbolic extractors are brittle across heterogeneous web layouts, while standalone LLM prompting is vulnerable to hallucination and attention drift in long, noisy pages. We introduce NSTS, a Neuro-Symbolic Timestamping System that combines a neural Prioritized Semantic Reasoner (PSR), a symbolic Cascaded Pattern Extractor (CPE), and an Uncertainty-Driven Governance (UDG) protocol. PSR uses structured prompting to classify explicit timestamps and prioritize evidence from visible text, metadata, and URLs. CPE is not designed as the strongest standalone symbolic extractor; instead, it is specialized as a complementary symbolic partner that selects the latest valid timestamp candidate, a behavior especially useful on high-entropy forum and Q&A pages with many competing timestamps. UDG then uses PSR confidence and conflict signals to decide when to trust semantic reasoning and when to fall back to CPE. We further introduce Iterative Greedy Denoising (IGD), a corpus-level exact-match denoising method for removing recurring timestamp-bearing boilerplate. On a dataset of 33,964 webpages, NSTS achieves 92.49% overall accuracy and 91.77% average domain accuracy, outperforming a simple standalone LLM baseline (84.39% overall; 78.35% average) and established symbolic baselines. Additional ablations show that symbolic-partner quality differs from standalone symbolic accuracy: CPE is the most effective partner for PSR under UDG, despite not being uniformly strongest as a standalone extractor.

Optical character recognition (OCR) quality assessment has depended on reference-based metrics (CER/WER), which require costly transcription. We ask: can an LLM supply a reference-free, text-only signal for OCR readability? Using 90 stratified passages from 53 scans of 17 English works in RETAS, we score OCR text with Claude Sonnet 4.6. Scores correlate with CER (Spearman ρ = −0.55, − 0.60) and WER (− 0.65, − 0.69); Pearson coefficients are weaker. Prompts rank passages almost identically (ρ = 0.92). A trivial OOV baseline reaches ρ = 0.49 against CER and 0.61 against WER; paired bootstrap intervals for the LLM advantage include zero. Inspecting seven highest-CER passages without page images, we found one reference-side mismatch and six mixing real errors with irregularities. We treat disagreement as an audit trigger rather than evidence that CER is wrong.

Evaluating whether scholarly research contributes to a Sustainable Development Goal (SDG) requires an interpretive framework that combines the goal definition, its targets, contribution criteria, and the document itself. Instruction-tuned large language models can instantiate this framework in context, producing a consistent evaluative function that can be applied to every document in a collection. We present Goal-Conditioned Multi-Label Distillation (GCMLD), which distills this evaluative function into a compact multi-label model. A Qwen2.5-7B-Instruct teacher evaluates each document–SDG pair by constraining its generative output to two tokens ("0" and "1"), and the logit margin defines the teacher probability used to supervise the student. An 86-million-parameter DeBERTa-v3-base student learns to approximate the teacher’s evaluative function by predicting all seventeen SDG scores in a single model invocation. We test whether reducing the influence of intermediate teacher evaluations improves or impairs approximation by comparing plain binary cross-entropy with weighting, masking, and hard-label objectives. Masking small-margin evaluations reduced teacher fidelity with the teacher targets and temperature held fixed, and the reduction was largest for SDG–document pairs whose own margins are small. The loss of fidelity shows that intermediate teacher evaluations provide useful supervision for approximating the evaluative function. The student trained on publication abstracts continued to approximate the teacher on electronic thesis and dissertation (ETD) abstracts, with ETD-specific distillation producing modest additional gains. Replacing seventeen prompted teacher evaluations with one model invocation makes collection-scale evaluation practical. Aggregating document-level contribution probabilities by academic department demonstrates how the distilled model supports collection-scale analysis of institutional research.

Electronic theses and dissertations (ETDs) often state research gaps, limitations, and directions for future work. These statements can support literature review and research planning, but they are difficult to find in long, structurally varied documents. We present a provenance-first pipeline that turns such statements into source-linked records. A replaceable proposal model reads bounded source windows and returns a proposed statement type, a short synopsis, and a verbatim source passage. A deterministic aligner then locates that passage in the compiled text and records its physical PDF page or page span. The record contract stores source-derived fields separately from the interpretive overlay, so type, synopsis, and rhetorical ownership can change without altering the evidence address. Three questions structure the evaluation. A development study across bundled input configurations compares LLM extraction from structure-selected regions, model-selected chapters, whole-thesis and per-chapter inputs, and bounded sentence-numbered windows, including additional sampling and recovery passes. It also compares cue-based and supervised sentence-classification baselines. Single-pass bounded windows provide the best recall–precision balance in the directly comparable panels and become the fixed input interface. Deterministic alignment then locates 93.9% of proposed passages and excludes unresolved proposals before serialization. In the main-evaluation portion of a stratified record audit, 99.5% of records point to the exact physical PDF page; proposed-type mismatches are concentrated at the gap–limitation boundary. Under the fixed input and provenance contract, four 4B-class open-weight backends adapted only on 27B proposal outputs retain 90.3–92.1% of the selective 27B configuration’s reference-set coverage, have precision estimates 2.8–3.8 percentage points higher, and emit fewer passages per ETD. Distilling the precision filter into a second 4B model yields 81.7% audit-only precision at 71.4% reference-set coverage, showing that both the proposal and filtering stages can use 4B-class models. The resulting system is a source-linked discovery and review layer over ETD collections. The release contains prompts, schemas, run manifests, source-linked ledgers, and audit decisions.

Electronic Theses and Dissertations (ETDs) provide broad access to student scholarship and support the long-term preservation and dissemination of academic research. Despite their central role in digital libraries, little is known about their usability for people with blindness or visual impairments (BVIs) who access scholarly documents using screen readers. We address this gap through a mixed-methods study in which BVI participants completed representative ETD interaction tasks followed by semi-structured interviews. Our findings reveal a substantial gap between accessibility and usability. Although the evaluated ETDs were technically accessible, participants struggled with information-seeking and navigation tasks and experienced considerable cognitive load, even with AI assistance. These challenges reflected limited screen reader AI capabilities, fragmented workflows, and poor integration between screen readers and more capable third-party AI systems. Based on these findings, we derive design considerations for assistive technologies that better support usable interaction with long-form scholarly documents.

An AI-generated podcast companion article to augment the Paper Session 10: Science of Science and Scholarly Communication Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

The same underlying computational problem is solved across unrelated fields under different names: recursive Bayesian state estimation appears as a “Kalman filter” in control, “Bayesian forecasting” in pharmacokinetics, and “data assimilation” in geoscience. Topical and citation-based scientific embeddings cannot see this shared problem. We distill each paper once into a domain- and method-name-stripped faceted computational fingerprint, a free-text mechanism skeleton plus controlled computational facets. We define a tunable, facet-selectable similarity over it. The goal is solution import: surface cross-field pairs solving the same problem, so a bespoke implementation can be swapped for another field’s standard, specialized solver. On a benchmark of 18 method families across 109 papers, the skeleton lifts cross-domain retrieval average precision over the abstract from 0.222 to 0.513, and the whole fingerprint reaches 0.557. Strikingly, four trained scientific embedders all fall below plain abstract+TF-IDF: they encode topical and citation similarity, the wrong signal for this task. The gain is the representation: the abstract-to-skeleton swap lifts every embedder, and the pipeline is one cached LLM call per paper plus a cheap embedder. An interventional re-skin / math-edit test shows the fingerprint tracks the computation, not the field. On a 501-paper wild corpus, known twins dominate the top of the ranking (23 of the top 30); with planted pairs excluded from the results, three blind LLM judges rate 3 of the top 5 and 8 of the top 30 pairs genuine import candidates, and 0 of 30 random ones. The human verification is the four executed imports: in one, an open standard solver reproduces a bespoke clinical dosing engine’s output. We release the benchmark, the code, and the distillation prompt.

This study examined the relationships between Data Availability Statement (DAS) categories, citation counts, and the evolution of research data-sharing practices. We analysed 653,490 research articles published between 2010 and 2025 in the ten open-access journals with the most DAS-containing articles, using PubMed Central full-text records matched to Web of Science citation data. DASs were classified into seven categories with a validated large language model (LLM) framework, and associations with citations were evaluated by non-parametric tests and regressions controlling for journal and publication year. The associations varied across journals (DAS × journal interaction: joint Wald test F₅₄ = 30.49, p < 0.001). After adjustment they were generally small: relative to articles reporting no data generated or not applicable (N), reuse of public third-party data (R) was the only category positive and significant in all four specifications; in-article or identifier-free availability (P) was also associated with slightly higher counts in most specifications; and full repository deposition (F) showed no clear advantage over N and lay slightly below author-request-only statements (A; − 0.029, 95% CI: − 0.036 to − 0.023). Longitudinally, rising DAS adoption was accompanied by a more heterogeneous pooled composition of sharing mechanisms rather than a single pathway towards open data, a shift driven largely by the staggered entry of journals. DAS categories thus provide a practical, scalable framework for characterising data-sharing practices, and journal context matters when interpreting them.

An AI-generated podcast companion article to augment the Paper Session 11: Research Integrity and Scholarly Records Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

The integrity of the peer-review process relies on the independence of its stakeholders, particularly handling editors. We introduce EditorCheck, an automated pipeline for detecting guideline-violating editorial assignments arising from prior collaborations between handling editors and article authors. Our pipeline ingests JATS XML files, disambiguates editor and author identities via the Dimensions API, and identifies offending coauthorships within a configurable time window to account for any prior submitted article. We further propose a 3-level scoring scheme—collaborations, article, and editor—designed to surface the most problematic cases. We applied EditorCheck to two large corpora: 340,409 research articles from PLOS and 172,614 from MDPI, uncovering more than 8000 and 2500 potentially problematic articles, respectively. The PLOS publication ethics team annotated the 365 top-ranked flagged cases. On this corpus, EditorCheck has a precision of 1.0 with no false positives among annotated cases. This work demonstrates that automated, large-scale auditing of the editor–author conflicts of interest is both feasible and effective, and highlights the need for publishers to enforce their own declared guidelines more systematically.

Open Large Language Model (LLM) research increasingly produces compact and derived artifacts whose evidence is distributed across papers, model hubs, code repositories, and release records. Although these artifacts are often publicly visible, the evidence needed to identify, cite, and preserve them as coherent scholarly objects remains fragmented. We introduce curatability as a record-level property of distributed scholarly records and operationalize it through four evidence dimensions: artifact identity, scholarly linkage, upstream evidence, and release assets. Using a May 2026 snapshot of 191,375 public Hugging Face repositories and 2,214 scholarly paper records, we characterize open LLM curatability at collection scale. The results reveal a pronounced visibility-to-curatability funnel: 90.7% of paper records contain at least one useful curation signal, but only 18.1% combine usable upstream and release evidence, and only 6.1% provide sufficiently coordinated evidence for high-curatability records. Based on these findings, we derive a minimal seven-field curatable record and outline complementary responsibilities for model hubs, scholarly indexes, and digital libraries. The code and results are available at: .

An AI-generated podcast companion article to augment the Paper Session 12: Text and Language Processing Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

Retrieval-augmented generation (RAG) reduces hallucination in large language models by conditioning generation on retrieved evidence, but retrieved contexts are often compressed before inference to reduce cost and latency. Prior work evaluates compression primarily through answer accuracy, leaving open whether different compression strategies affect how faithfully model outputs are grounded in the provided evidence.

We present a controlled study of hallucination behavior under context compression and introduce Attention-Based Semantic Preservation (ABSP), a lightweight query-agnostic method that retains tokens according to structural centrality in a proxy model’s self-attention patterns.

Across 2,400 inference runs comparing no compression, random pruning, LLMLingua-2, and ABSP on TriviaQA and HotpotQA, structural compression consistently improves grounding faithfulness, measured by K-Precision. ABSP achieves the highest K-Precision among compressed methods at every compression ratio, outperforming LLMLingua-2 by + 0.111 on average and + 0.131 under aggressive compression, while matching its answer accuracy on single-hop retrieval. Cross-sample consistency is stable across methods, suggesting this hallucination signal is driven more by the generation model than by the compression strategy.

An AI-generated podcast companion article to augment the Paper Session 13: Recommendation and User Engagement Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

Reading plays a critical role in adolescent cognitive development, literacy, and long-term educational success. Yet leisure reading among teenagers has declined substantially in recent decades, driven in part by increasing digital distractions and the difficulty of identifying engaging and meaningful reading materials. To address this challenge, we propose a personalized book recommender that jointly models the emotional alignment and STEM relevance of books for teenage readers. Our approach represents books using emotion and STEM feature vectors and employs Bilinear Pooling to capture interactions between affective and topical characteristics. By integrating teenagers’ emotional preferences with educational interests, the proposed framework promotes both reader engagement and exposure to STEM-related content. Experiments on a large-scale book dataset demonstrate that our method consistently outperforms strong recommendation baselines across multiple evaluation metrics. The results suggest that jointly modeling emotional and topical alignment can improve recommendation quality while supporting sustained reading engagement and educational development. The proposed framework offers a scalable approach for connecting young readers with books that are both personally meaningful and academically enriching.

This study focuses on Chinese researchers and systematically investigates their awareness, usage behaviors, and attitudinal characteristics toward generative artificial intelligence (GenAI) through a large-scale questionnaire survey (N=4,788). Findings show near-universal adoption (99.8%) and high-frequency use among the sampled researchers, with 47.62% using GenAI almost every day and an additional 46.01% using it two to four days per week. Cognitive evaluations are highly asymmetric, with 88.61% reporting strong perceived benefits but only 48.46% acknowledging risks. Besides, GenAI is embedded across research stages, with perceived usefulness ranging from 75.58% in academic communication to 84.80% in inspiration generation, while negative evaluations remain below 6%. Despite ethical and epistemic concerns, satisfaction is high (86.02%) and only 0.36% report unwillingness to use GenAI in the future.

Libraries have long used social media mainly to broadcast announcements and promote services. Users now encounter information, learning, and community directly on platforms, making social media a venue for service delivery rather than promotion alone. In this paper, we present a layered Content Delivery Model to rethink library services on social media. We first draw on Uses and Gratifications Theory, Connected Learning, and the concept of platform vernacular to build an analytical framework of five archetypes of library social media practice: Digital Bulletin Board, Institutional Personality, Librarian-as-Creator, Micro-Learning Channel, and Community Hub. Reading the archetypes through the three theoretical lenses shows why announcement-driven posting is often misaligned with how users now encounter information on platforms. The model then organizes service-oriented activity into four layers: Discovery, Value Delivery, Connection, and Bridge. Together, the archetypes and layers show how libraries can shift from visibility-oriented posting to service-oriented content delivered through platform-native formats. An exploratory descriptive analysis of 372 TikTok and Instagram posts from 25 libraries and librarian-creators illustrates the archetypes. It shows how they classify communicative purposes, identify degrees of service orientation, and clarify patterns in current practice. In this paper, we contribute a theory-grounded vocabulary and planning framework for aligning social media activity with information discovery, learning, community participation, and pathways into the broader library service ecosystem.

Access to digital information depends on a user’s ability to express an information need before a retrieval or answer system can respond. Text-centred query interfaces can make this entry point difficult for adults with learning disabilities when interaction depends on spelling, lexical recall, and abstract keyword construction. We combined community fieldwork with an exploratory study of an icon-supported multimodal query interface paired with a generative answer backend. Fieldwork observations informed design requirements for recognition-based input, visual composition, guided suggestions, and simplified interaction. Seven adults completed two structured activities under instruction-defined icon and text conditions within the same multimodal system, followed by a self-directed free-choice task. Screen and session recordings were analysed to trace query trajectories, assistance, input-resource crossovers, response outcomes, and observable evidence that participants followed the responses. Retrospective ratings of the two structured conditions and synchronised physiological traces provided secondary descriptive context. The observed patterns did not indicate a single best input resource. Most participants selected icon input for the self-directed task, but responses addressing the intended need were observed with both input resources. A crossover from text to icon input was also observed during query repair, while icon input sometimes introduced ambiguity and required assistance. We contribute a traceable account linking field observations to interface design and derive implications for inclusive digital-library query interfaces: providing multiple input resources, supporting movement between them, and making the system’s interpretation of a query visible to users. Given the small sample, fixed condition order, and use of generated rather than collection-retrieved answers, the findings should be treated as exploratory design evidence.

Antiretroviral therapy has made HIV medically manageable, yet long-term treatment continually brings HIV-related information into medication routines, digital devices, institutional records, and social relationships. Prior research shows that disclosure is relational and sometimes involuntary, but less is known about how the possibility of identity linkage shapes everyday information practices before HIV status is actually recognized, or why information that people understand and trust may remain unacted upon. We conducted a qualitative field study in a large Chinese city, combining participant observation, 22 semi-structured interview sessions with 20 participants, including 16 people living with HIV, retrospective accounts, and exploratory online observation. Inductive thematic analysis identified three interconnected practices: regulating engagement with information, navigating identity-linked information flows, and translating information into everyday life. We develop identity-linkage potential as the situated possibility that a particular audience can access a cue and interpret it as evidence of a person’s HIV status. This potential shaped participants’ engagement with information, management of artifacts and records, access to support, and ways of acting on information. Within a situation, the same action could alter all three practices; across situations, its consequences could reshape later sources, audiences, channels, support resources, and everyday arrangements. Accurate and trusted information could therefore be acted on directly, discreetly, or through others, but could also be deferred or abandoned. Lower identity linkage sometimes protected ordinary life while limiting HIV-specific support. For digital library research, these findings highlight the need for privacy-sensitive access, intelligible information flows, low-visibility participation, and multiple viable routes from information to action.

Generative Artificial Intelligence (GenAI) is rapidly transforming academic research, yet the mechanisms underlying researchers’ adoption decisions remain insufficiently understood. This study employs systematic semi-structured interviews and a grounded theory approach to explore how researchers make adoption decisions when using GenAI. The findings reveal that GenAI adoption is a dynamic decision-making process involving five interconnected mechanisms: dynamic fit, comprehensive evaluation, decision-making, contextual constraint, and feedback. Researchers continuously assess the fit between research needs, task requirements, and perceived GenAI capabilities while balancing academic risks, self-efficacy, and contextual conditions. Based on these findings, this study develops a grounded theory model of researchers’ GenAI adoption decision-making. The study extends technology adoption research by conceptualizing GenAI adoption as a dynamic process of human-AI collaborative decision-making and provides practical implications for researchers and academic institutions.

Scholarly metadata describe scientific careers at scale, but their potential as inputs to simulation remains underexplored. This poster presents a method for studying scientist mobility through large language model (LLM) agents embedded in an explicit institutional environment. Publication and affiliation histories define annual states; quantitative priors inform decisions; retrieval limits institution choices; and staff quotas, offer competition, and waitlists determine feasible outcomes. LLMs form country preferences and application portfolios, with a separate, narrowly bounded role in institutional review. Policy interventions can enter at the information, institution-accessibility, or post-offer stage, making their propagation observable. A proof of concept follows 3,795 scientists under four prospective policy scenarios that share one simulated starting state. The reported runs show substantial contraction before offers are issued, illustrating what stage-level analysis adds to aggregate mobility counts. The contribution is a method for turning scholarly records into an inspectable simulation environment.

Evidence synthesis methods, such as systematic reviews (SR), require research teams to find, select, evaluate, and synthesize all available evidence to answer a specific research question in a reproducible and transparent, bias-reduced manner. Early in the process, teams must screen all unique titles and abstracts gathered from a comprehensive search, which can result in thousands of records. Therefore, the process often takes months to years to complete. In this research, we conduct a set of experiments using open-source Large Language Models (LLMs) with small parameters hosted on local servers to screen titles and abstracts against pre-defined eligibility criteria respective to each topic. We use two cases from real SRs for which titles and abstracts screening has already been completed by dual, independent human screeners, which we use as our “ground truth”.

The goal of this work is to contribute to developing an AI-enabled workflow to reduce the time and labor required for the relevance screening phase of SRs. Therefore, we start by focusing on metrics such as recall to ensure we don’t miss any relevant results during automatic screening. As the value of these methods depends on a comprehensive corpus and missing relevant records can impair the quality of results, recall is the topmost priority. Our results suggest that AI may indeed aid with initial relevance screening, but more research is needed before full adoption. Repeated runs also demonstrated that models have deterministic results across different prompt variations.

Research data reuse is an objective of Open Science, and data journals have emerged as an important publication format for supporting reuse of research data. However, it remains unclear how this objective is operationalized in humanities-oriented data journals and whether reuse-related information is made accessible through their search interfaces. This paper investigates the support for research data reuse in humanities-oriented data journals. We synthesize prior work on research data reuse and identify different reuse purposes and forms. In addition, we analyze how ten humanities-oriented data journals implement support for reuse in their documentation structures and search interfaces. Our analyses show that reuse is frequently recognized as an objective in humanities-oriented data journals, but reuse-related information is not directly represented in the search interfaces. Building on these findings, we derive implications that address the representation of reuse-related information and the design of search systems.

Aspects and opinions in reviews are often left unstated, and these implicit cases remain the hardest part of Aspect-Category-Opinion-Sentiment (ACOS) quadruple extraction. We reimplement the generative GEN-SCL-NAT model on a T5-Base backbone and add two modifications aimed at those cases. Instead of a single overall score, we report results separately for the four ways an aspect and an opinion can be written or left out. Seen this way, one modification helps on implicit aspects and hurts on implicit opinions, while the overall score barely changes. Gains appear only where either the aspect or the opinion still remains in the sentence. Where neither is written, nothing improves beyond the variation we see between training runs.

Classical documents in digital libraries—such as Classical Chinese texts—are not only valuable cultural heritage but also core educational resources at the foundational education level. However, existing digital library systems predominantly focus on full-text retrieval and basic machine translation, making it difficult to support learners' deep cognitive construction. Meanwhile, the direct integration of large language models (LLMs) tends to trap learners in an answer-seeking mode, thereby diminishing the inquiry-based thinking essential to the humanities. This paper proposes and implements Classical-Tutor, an intelligent tutoring system for Classical Chinese digital resources. The system deeply integrates Retrieval-Augmented Generation (RAG) with a curriculum knowledge graph and introduces a graph-based dual-role dynamic scheduling engine (Dual-Role State Machine). Through a multi-trigger mechanism, the system adaptively switches between knowledge lecturing and heuristic guiding. In an LLM-simulated pilot evaluation the system sustains longer instructional interaction than a baseline RAG system, adapts retrieval granularity to learner profiles, and avoids the deadlocks of static Socratic prompting; the results characterize orchestration behavior rather than learning gains. The study offers an architectural reference for the intelligent utilization of humanities special collections in digital libraries.

In this paper, we propose a method for estimating the name recognition of public figures, that is, the proportion of the population that recognizes the individual. It is typically measured through surveys. The proposed method estimates changes in name recognition relative to the previous survey results, enabling us to infer current values without conducting repeated surveys. It uses two types of information: (1) the number of recent page views of the individual’s Wikipedia page, and (2) the topics of recent news articles about the individual. The former serves as a proxy for the scale of events that made people newly aware of the individual, while the latter captures the types of the events. We conducted an experiment on a dataset of 967 individuals. The result shows that the combination of these two types of information improves estimation accuracy.

This poster presents the initial steps toward ARCHIGRAPHY as a global platform for digital photographs of architecture and urban spaces. Since 2009, its original Portuguese-language version has been developed and maintained by a group of students, researchers and professionals at the University of São Paulo, Brazil. Building on the project's success in Brazil as both an educational resource and a discovery tool, we seek to expand the platform by incorporating architectural photograph collections from other regions of the world. This expansion presents several challenges, including enabling information access and retrieval across multiple languages, establishing meaningful relationships among image records, and ensuring the project's long-term funding and sustainability. To address these challenges, we have undertaken a number of strategies, including broadening collaboration with institutions in the United States and South Africa, testing cross-language information access solutions, and promoting our system as an open-source tool.

Electronic Theses and Dissertations (ETDs) represent a significant body of original scholarship, yet they remain underrepresented within the scholarly communication infrastructure that supports citation linking, discovery, and research assessment. Although many institutions assign persistent identifiers to ETDs, relatively few expose their bibliographic references through services such as Crossref, limiting the integration of graduate research into global citation networks. To address this gap, the University of North Texas Libraries developed a scalable workflow that enhances the visibility and interoperability of its ETD collection by combining retrospective Digital Object Identifier (DOI) registration with large-scale reference extraction and reference deposit to Crossref. This poster describes the end-to-end implementation, including DOI assignment for tens of thousands of ETDs, automated extraction and normalization of bibliographic references, metadata quality assurance, and Crossref reference registration. By depositing references alongside ETD metadata, the initiative enables theses and dissertations to participate more fully in citation services and the broader scholarly citation ecosystem. The poster discusses the technical and metadata challenges encountered during implementation, lessons learned from processing a large institutional ETD corpus, and the broader implications of extending established scholarly publishing practices to theses and dissertations.

Evidence-based practice helps academic librarians understand users, evaluate services, and make decisions in changing information environments. This ongoing content analysis examines how evidence is used in articles published from 2020 through 2025 in College & Research Libraries, The Journal of Academic Librarianship, and Reference Services Review. Of 420 coded articles, 339 (80.7%) were evidence-based. The leading topics were information literacy and user education, librarians and the profession, and information and reference services. Local evidence was most often combined with research evidence, while studies of librarians and the profession more frequently incorporated professional knowledge. Questionnaires and interviews were the most common data-collection methods; survey and evaluation research were the leading research strategies. Mixed, quantitative, and qualitative approaches were all represented, although descriptive statistics were more common than inferential techniques. AI-related studies addressed staff perceptions and literacy, reference-question answering, virtual-chat assessment, service analysis, and library-focused language models. The findings identify patterns in evidence use, research methods, and emerging AI applications that can inform professional development and LIS education.

Japanese classical texts are an important resource for literary and historical studies. Although large-scale digitization projects have made many classical texts available, OCR-generated texts still contain recognition errors and historical orthographic variations that reduce retrieval performance. This paper investigates the impact of OCR correction and text normalization on example-based retrieval of Japanese classical texts. We propose an evaluation framework that systematically compares multiple text collections, including OCR-generated texts, rule-based normalized texts, LLM-corrected texts, and manually corrected texts. Using an existing dense retrieval model, we evaluate how different correction and normalization methods influence retrieval performance. Our experiments show that simple rule-based normalization methods consistently improve retrieval performance with different normalization rules benefiting different types of queries. These findings suggest that retrieval performance depends not only on transcription accuracy but also on retrieval-oriented text normalization.

Provenance research investigates the ownership histories of museum objects, but records are typically documented as semi-structured or unstructured free text, which limits computational and comparative analysis within and across institutions. This paper presents a reusable LLM-in-the-loop workflow—with all scripts and prompts shared—that transforms such records into structured, authority-linked data through schema-constrained extraction, source-grounding validation, entity pooling, Wikidata reconciliation, and disambiguation, with audit metadata preserved throughout to support traceability and human review. We applied the workflow to 1,829 provenance records of pre-1912 Chinese artworks at the Art Institute of Chicago. The workflow identified 314 distinct collecting actors, of which 114 were matched to Wikidata. Newly created Wikidata items were also incorporated in subsequent runs, showing that the workflow supports the participation of Libraries, Archives, and Museums (LAMs) in building shared authority data. Beyond the pilot dataset, the workflow contributes to digital scholarship infrastructure for non-Western collections, enabling researchers to map historical collecting networks and formulate new questions made possible by linked data.

We have developed state-based semantic programs that describe scientific research phenomena. However, implementing interactive research reports requires more than a model of the phenomenon: it also requires discourse elements such as rhetorical and genre structures. We use the term discourse broadly to cover how material is presented. This paper treats those structures as an overlay on the semantic program of the base phenomenon. We discuss two examples related to the Snowball Earth Hypothesis.

Translations are surprisingly difficult to identify from legacy library records such as those in the HathiTrust Digital Library catalog. This poster describes a custom pipeline for building a corpus of probable English translations from HathiTrust records: five scripts run in sequence, each taking the output of the previous as input. The script accepts a list of candidate HathiTrust IDs as input; extracts MARC metadata from the HathiTrust Bibliographic API for each volume; downloads the corresponding HathiTrust Research Center Extracted Features files via rsync; sends sample words from those files to the fastText language detector to infer the likely language of the text. Selected metadata fields, both structured and free-text, are searched for translation-related terms and indicators, and a record is retained as a probable English translation if it has a positive translation indicator in the MARC 041 field or contains a translation-specific term (e.g., “translated by”, “tr. from”, “rendered into”). Finally, quality filters are applied to infer source language; records are deduplicated; and a condensed public-facing dataset remains as output. This pipeline has proven more accurate and precise at identifying translations than other methods tested, including LLM-based methods. However, LLMs tend to achieve better overall recall, pulling in a wider pool of potential translations for consideration.

With the increasing use of automated methods for metadata creation and resource description, validation of programmatically generated data has become increasingly important. In digital humanities and cultural heritage work, validation of these outputs often takes on further significance due to the personal, subjective, or historical nature of the information being processed. Our demo seeks to approach the issue of validation through the prototyping of a Wikipedia Edit-a-thon companion application, which combines the community-centered practice of crowdsourcing with the use of computationally generated metadata. Initially designed for use with the oral history interview collection of the Niels Bohr Library and Archives at the American Institute of Physics (AIP), our application assists users in identifying interview subjects within a corpus who may lack description on Wikipedia. The application does so by presenting possible matching Wikipedia pages for oral history subjects, then facilitating user input to identify candidates who may be in need of additional description. Our demo highlights a potential workflow with this application, along with broader use cases for crowdsourcing practices, reparative description efforts, and collective public history.

TURATH (Arabic: تراث, “heritage”) is an open-access web platform that combines large language models with domain-specific glossaries, structured extraction schemas, and human-in-the-loop review to transcribe, translate, and index historical handwritten documents. TURATH allows researchers to configure a collection-specific AI reader by providing annotated examples from their collection. The system then generates a tailored configuration—including prompts, glossaries, and output schemas specific to the collection—designed to support transcription of collection-specific language and terminology. The platform has been tested with Arabic, French, and English manuscripts and applied to collections spanning 1910–1960, including archaeological excavation diaries, business correspondence, and personal recipe notebooks. A live demonstration is available at . A video walkthrough is available at .

Conversational agents in digital library and information-access systems must revise a session-level constraint when a user overrides an earlier stated preference. We ask a more specific question, with direct implications for system design: when the model produces its final response, does it depend more on the turns before the user’s correction, or the turns after it? We answer this with a causal, context-masking audit: for each user turn we measure how much removing it changes the model’s teacher-forced probability of its own final response, and compare the influence of pre-correction turns to post-correction turns via a primacy influence ratio ρ_mask. Across five Instruct-tuned LLMs (Llama-3 8B/70B, Qwen2.5-7B/72B, Mixtral-8x7B) on 308 contradiction-injected MultiWOZ dialogues, four of five models show a statistically significant primacy bias (ρ_mask > 1, p < 0.001): pre-correction turns causally shape the final response more than the correction and everything said afterward. The exception, Qwen2.5-7B, is balanced (p = 0.51). Its 72B counterpart is not, revealing that scale inverts the bias direction within one model family (Qwen2.5) while leaving it unchanged within another (Llama-3). We argue that digital library systems using LLM-driven session models should not assume a single, architecture-independent fix for primacy effects.

Keyphrases are core metadata for digital libraries, yet progress in keyphrase extraction and generation is hard to judge from the literature: each project rebuilds its own evaluation (dataset readers, matching rules, cost bookkeeping, significance tests), so published numbers are hard to reproduce and to compare. We demonstrate KPViz, an open-source tool that turns a folder of datasets, model outputs and cost descriptions into one reproducible evaluation and an interactive dashboard. Users describe each dataset, model and machine once in a short JSON card; KPViz then aligns every prediction with its gold keyphrases and its cost through these cards, so adding a dataset, a model or a run requires no code. It computes corpus statistics, present/absent keyphrase classes, exact-match scores and significance tests, and recomputes only what changed after an edit. Every table and figure exports to camera-ready LaTeX with an editable caption, so the figure a researcher sees on screen is the one that lands in the paper. The demonstration follows one question end to end: does a model’s context window limit what it can extract? Demo data and a demo video are public, and the code is open source at .

Generative artificial intelligence (AI) systems, specifically Large Language Models (LLMs), are increasingly used as collaborative problem-solving tools, yet their efficacy in physical environments like makerspaces remains unexplored. This study presents a cross-generational UX analysis of adults and parent-child dyads interacting with conversational AI (Copilot, ChatGPT) during physical engineering tasks. Through episodic qualitative coding of 54 interaction episodes across two multi-layer maker activities (paper roller coasters and squishy circuits), we identify the "Physicality Gap". We introduce the "Prompt Repair Taxonomy", that captures "Context Grounding," "Constraint Reframing," "Granularity Adjustment," and "Logic Interrogation", which demonstrates that while AI excels at text-based ideation, it lacks spatial and material awareness. Users are forced into a high cognitive state of "Critical Verification" to debug AI outputs against physical constraints. Finally, we outline UX design implications for future digital library infrastructures and argue for multi-modal, spatially aware AI interfaces.

Large Language Model (LLM) workflows increasingly decompose knowledge-intensive generation into planning, role assignment, subtask execution, and synthesis. Although this structure may support specialization, it also adds coordination and computational overhead. We present a preliminary task-level comparison of a Direct Generation Workflow and a Decomposed Generation Workflow under matched inputs and evaluation conditions. The study covers three grounded knowledge-generation tasks: public-event narrative synthesis, project-grounded research ideation, and scientific related-work generation. Across the evaluated configurations, decomposition produces the largest mean Task Quality Score (TQS) gain for research ideation (+0.053), a smaller gain for related-work generation (+0.025), and only a modest gain for public-event narrative synthesis (+0.013). These improvements require 95.1–126.2 seconds of additional runtime on average. The results provide preliminary evidence that the value of multi-agent decomposition is task-dependent.

Historical maps are becoming increasingly accessible; however, users often struggle to search and navigate the content. Much of the embedded textual and spatial information within scanned maps is not machine-readable, and therefore largely inaccessible. In this research, we investigate computer vision and AI-assisted methods to make metadata and map information available to (a) improve searchability and (b) streamline the metadata creation process. We apply AI tools and machine learning to extract text from the mapframe and the mapborder, and thus aim to provide a better user experience when interacting with maps within the digital library. Additionally, we classify the extracted mapborder information to help facilitate metadata creation through AI-integrated workflows and reduce data inconsistencies arising from a completely manual workflow.

Digital libraries enable systematic scrutiny of scholarly work, including retraction, a safeguard for flagging flawed or fraudulent publications. However, existing retraction studies primarily focus on the reliability of a paper’s own findings and offer little insight into how it engages with the literature it draws on. We investigate the referencing behavior of retracted papers and formulate citation selectivity as a novel research problem grounded in motivated reasoning, where authors may curate reference lists that echo their own claims while under-citing equally relevant but less convenient work. We introduce the reference selectivity score, which compares how closely a paper’s cited references align with its content against an independently constructed pool of expected references. Using a large-scale dataset of 42,330 matched pairs of retracted and control papers, we find that retracted papers exhibit significantly higher citation selectivity than their matched controls.

Executable works produced in Java in the late 1990s are threatened by the disappearance of applets and the obsolescence of their software interfaces. This article presents an exploratory study of their restoration from bytecode and surviving assets, without access to the original source code. The protocol combines two conventional decompilers, a coding assistant based on a large language model, historical references, and iterative human validation. Following a methodological pilot on John Maeda’s Flora Calendar, it is applied to Forward (Komplex, 1998), a complex and obfuscated Java demo. In these two cases, reconstruction in the original language provides a more reliable foundation than direct porting: it allows us to restore execution, examine the algorithmic logic scene by scene, and then produce documented derivative versions. The contribution is a traceable heritage-restoration protocol that restores access to some of the aesthetic and technical know-how buried in the code. AI acts as an assistant for interpretation and repair; the fidelity of the reconstruction remains subject to external evidence and human expertise.

Digital libraries may encounter AI-transformed texts without reliable links to the collection records that guided their generation. We study post-hoc source-link recovery: using a generated derivative to rank its conditioning source within a known collection. Our controlled benchmark contains 60 passages generated from three King James Version chapters by two chat systems under content-preserving and style-oriented instructions. Each passage queries all 260 New Testament chapters using transparent lexical, character-based, stylometric, and hybrid retrieval methods. Word TF–IDF achieves mean reciprocal rank (MRR) of.967 for content adaptations and.956 for style-oriented transformations. After masking words participating in source-shared 3-grams, MRR remains.944 for content passages but falls to.598 for style-oriented passages. Stylometric MRR for style-oriented passages falls from.540 to.018, indicating substantial dependence on retained source-local phrasing. These results provide exploratory evidence that lightweight retrieval can support source-link suggestions in a controlled collection. The small, homogeneous dataset does not establish transferability to larger or more diverse digital libraries.

Scholarly search is dominated by top-k retrieval, which returns a ranked list of documents judged relevant in isolation, leaving the relationships among them for the reader to reconstruct. Retrieval-augmented generation leaves this structure implicit as well, since it synthesizes summaries over a top-k result set without exposing the structure that connects the retrieved work. We argue that this missing structure is central to scholarly sensemaking, where understanding literature means grasping how its ideas connect and develop. Building on the Narrative Trails algorithm, we present a narrative extraction pipeline that imposes this structure on the retrieved set. The proposed method uses a large language model to decompose a natural-language query into two narrative endpoints, retrieves a candidate set of documents, constructs a document graph, and extracts the maximum-capacity path between the endpoints as a coherent storyline, which a language model then synthesizes into a structured summary. Code for a demo of our system can be found at

Galleries, libraries, archives, and museums increasingly publish their catalogs as linked open data structured by CIDOC-CRM, an event-centric ontology in which people, objects, places, and documents are connected through the events of their shared histories. Making these event chains inspectable alongside collection records is an access-design challenge. We investigate immersive access as an alternative: an XR client that renders a CIDOC-CRM collection as an explorable spatial graph, retrieved on demand from a live, configured SPARQL endpoint. We contribute (1) six design requirements for immersive access to event-centric heritage linked data, grounded in the digital-library literature; (2) an ontology-driven mapping from CIDOC-CRM constructs to a spatial-graph representation, partially implemented in a working prototype; and (3) an author-conducted analytical evaluation that combines a requirements walkthrough with a quantitative characterization of a live heritage endpoint, identifying where the design holds and where collection-scale realities strain it. No user study was conducted.

Generative AI has brought a new wave of automated visitors to digital libraries, and libraries now try to allow, slow or block particular crawlers. The tools they use, such as robots.txt rules and block lists, act on the name a crawler gives itself. That name is a text string any client can send, so it cannot prove identity on its own. We ask how far a library can check a declared crawler identity against independent evidence: the IP ranges that operators publish. We apply this to the requests that OpenEdition, a large multilingual humanities and social sciences platform, identified as declared-bot traffic. Three lessons emerge. Among requests claiming a generative-AI identity, a minority is consistent with published operator infrastructure. Most come from families that publish no range at all, so they cannot be checked. And a mismatch is not necessarily spoofing: it can reflect one address posing as many bots, or an operator’s own list lagging behind its infrastructure. A library should therefore read a declared identity against the infrastructure documented at the time of the request. We release our pipeline and dated registry so that other libraries can run the same audit, and future studies can compare crawler traffic on a common footing.

Digital libraries increasingly host spoken content containing mathematical expressions. Making this audio searchable and accessible requires converting spoken formulas into structured representations such as LaTeX. While traditional dual-stage pipelines (transcribing audio to text, then text to LaTeX) suffer from compounding errors, direct speech-to- LaTeX models eliminate intermediate text transcription. This paper evaluates two unified end-to-end models alongside dual-stage pipeline baselines on an extended version of the MathBridge-Audio dataset, augmented with 1,055 additional samples beyond the original 1,000. Models are evaluated on word error rate, character error rate, and hallucination frequency. Results show that unified models consistently achieve lower error rates, produce more exact matches, and hallucinate far less often than dual-stage pipelines, with a fine-tuned Whisper model achieving the strongest performance. Overall, direct speech-to- LaTeX modeling provides a more reliable foundation for indexing and rendering spoken math in digital library systems.

Digital libraries increasingly provide access to large collections of mathematical content, including textbooks, research articles, educational resources, and community question-answering forums. Effectively searching these collections remains challenging because mathematical queries combine natural language with symbolic notation, where symbols often have context-dependent meanings and essential assumptions are frequently left implicit. These characteristics make query ambiguity difficult to identify, reducing retrieval effectiveness and increasing user effort. Despite its importance for mathematical digital libraries, mathematical query ambiguity detection has not previously been studied as a benchmark task. This paper introduces MathDisAm, the first benchmark for mathematical query ambiguity detection. The benchmark contains 1,000 queries, evenly divided between ambiguous and unambiguous instances, derived from Math Stack Exchange. Ambiguous variants are created through targeted omission operations, while corresponding unambiguous versions provide explicit clarification. All instances are manually validated by human annotators to ensure annotation quality. Baseline experiments with three transformer-based classifiers show that the domain-adapted MathBERT achieves 90.0% accuracy, outperforming the general-purpose RoBERTa-base (86.5%) and BERT-base-uncased (86.0%). Error analysis indicates that domain-specific pre-training substantially reduces false positives, although all models continue to struggle with missing mathematical context and inconsistencies that are difficult to recognize from symbolic notation alone.

Educational digital libraries may grant large language model (LLM) tutors access to worked solutions, but access to repository content should not automatically authorize a model to reveal final answers. This study evaluates whether LLM tutors can leverage private solution context to generate pedagogical hints without disclosing final answers to students. Four models, two general instruction-following models and two math-specialized models, are evaluated across 5,000 aligned problem–message cases per condition, constructed from 500 MATH benchmark problems paired with ten fixed student prompts that ask for, or attempt to confirm, the final answer. Three prompting strategies are compared: a plain assistant prompt, a basic tutoring prompt, and a two-step method that generates a private solution before producing numbered hints. The basic tutoring prompt reduces detected answer leakage from 47.4% to 0.3% for GPT-4o-mini and from 30.2% to 1.0% for LLaMA-3.1, but produces no comparable improvement for Qwen2.5-Math or DeepSeek-Math. Under identical strict hint rules, adding private solution context has minimal effect on GPT but significantly improves tutoring compliance for LLaMA; Qwen produces the most accurate private solutions yet shows the highest leakage and solution-to-hint overlap. These results indicate that private solution access can support hint generation in some architectures but is not a reliable answer-withholding mechanism on its own. For educational digital libraries, access to solution content and authorization to disclose it should be treated as distinct access-control decisions.

Selecting which large language model to trust for querying a digitised historical collection is a concrete, unresolved problem for digital libraries. This is not only a collection-level question: model performance is highly heterogeneous even within a single collection, varying from page to page as layout, language, and content type shift, so that no single model can be trusted uniformly across a collection’s pages. We demonstrate, in this paper, an interactive tool for exploring the benchmark and the multi-model evaluation underlying our companion paper on adaptive LLM selection for historical document question answering. The demonstration exposes, page by page, 5,941 scanned pages of the Chinese Economic Journal (1923-1937), 84,441 multiple-choice questions automatically generated from them and the resulting 50 × 5, 941 binary correctness matrix obtained by evaluating 50 open-source LLMs. Rather than presenting model heterogeneity only through aggregate statistics, the demonstrator lets practitioners and researchers inspect, for any page and any question, which of the 50 models answered correctly and which did not, directly grounding the case for adaptive model routing in observable evidence. We describe the resource the demo is built on, walk through the views of the tool, and outline usage scenarios for digital humanists, historians, NLP researchers as well as digital library practitioners.

We present scalable-htr, a user-friendly open-source Python library for end-to-end segmentation and transcription of historical manuscripts. Built upon Kraken but extensible to other engines, the tool provides a framework for reproducible experimentation and efficient, large-scale inference on high-performance computing (HPC) environments. The tool facilitates systematic benchmarking across datasets and models by computing quantitative end-to-end metrics, and it generates interactive visualizations that allow users to qualitatively assess model adaptation to their target domains by inspecting segmentation and transcription results directly on the images. By combining systematic evaluation with scalable transcription, scalable-htr has been designed to assist researchers and cultural heritage professionals convert digitized manuscript collections into machine-readable data. The tool is demonstrated via a case study on medieval notarial manuscripts.

Processing technical terms in documents is essential for digital libraries and knowledge organisation, supporting applications such as scholarly search, information retrieval, and domain terminology construction. While document-level term identification, a sub-task of automatic term extraction (ATE), has been tackled as a sequence labelling task, the emergence of large language models (LLMs) opens new avenues for prompt-based extraction. However, evaluation practices have persistently relied on simple success or failure against a single gold standard, without analysing how model behaviours align with human annotations, especially on borderline cases where annotators disagree. To address this issue, this study investigates to what extent LLMs can perform term identification using prompt instructions mirroring human annotator conditions. By evaluating multiple commercial and open-weight LLMs against double human annotations in ACL RD-TEC 2.0, an ATE benchmark, we show that LLMs achieve performance close to human annotators (up to 76.6% F₁, within 5 percentage points of the inter-human agreement of 81.6%), with open-weight 30B-scale models outperforming larger commercial counterparts. Crucially, our analysis reveals a distinctly conservative nature: LLMs reliably extract terms on which humans consistently agree, but omit terms on which humans disagree. Our findings reveal the limitations of relying on a single gold standard, highlighting the need for annotator-aligned prompting and controllable precision–recall tuning.

Multimodal video collections increasingly require reliable access to visual content through textual descriptions, yet cross-modal retrieval and semantic matching can be sensitive to variations in textual input. In practical information access scenarios, textual descriptions may be incomplete, vague, or contaminated by irrelevant context, leading to unstable cross-modal similarity and unreliable identification of relevant video content. This paper presents RCA-VAD, a lightweight robust cross-modal alignment framework for reliable multimodal video information access under noisy textual descriptions. The method introduces prompt reliability weighting, which combines semantic agreement with video-conditioned evidence to down-weight unreliable textual representations, and a clean–noisy consistency objective that constrains model predictions to remain stable across textual perturbations. These components are integrated with a CLIP-based weakly supervised video anomaly detection framework without requiring additional frame-level annotations. We evaluate the approach on UCF-Crime using controlled lexical omission, semantic substitution, and context-insertion noise to simulate variations in textual descriptions. The results show that RCA-VAD maintains competitive performance with clean textual inputs while substantially reducing performance degradation as the noise ratio increases. The findings demonstrate the potential of robust cross-modal alignment for more reliable multimodal video information access and semantic discovery under noisy textual descriptions.

Citations in scholarly papers serve various functions and refer to diverse resources. Methods have been proposed to classify citation functions and the types of cited resources. In scholarly papers, citations take two formats, i.e., citations using citation tags (bibliographic citations) and citations using URLs (URL citations). To utilize information about citation functions and the types of cited resources, it is important to be able to classify them across both formats; however, previous studies have classified them independently for each format. This study investigated the effect of combining bibliographic and URL citations on the classification of citation functions and the types of cited resources. We constructed classifiers through transfer learning and multitask learning, and evaluated their performance. The results demonstrated that combining the two citation formats improved classification performance across various encoder models and learning approaches. The improvements were particularly large for cited resource type classification of bibliographic citations.

We present Journal Atlas, a public corpus of 399 curated venue profiles and two methods over it: one that scores how well a manuscript fits a venue and reports how much evidence that score rests on, and one that builds the corpus with a language model without letting it invent what nobody has recorded.

Existing recommenders match title and abstract against published articles and return topically similar venues, but what decides whether a submission survives is largely absent from any index: whether a stated word limit is enforced, whether a reviewer pool treats a method as a defect. The most systematic comparison of twenty such services closes by observing that background knowledge is still required for the final decision. Asking a language model instead substitutes fluency for evidence, where fluent-but-wrong is measured rather than suspected.

Hard constraints eliminate, six weighted dimensions rank what survives, and a dimension without evidence returns unknown rather than a midpoint. A score above neutral is then shrunk in proportion to the evidence missing beneath it, and missingness is never allowed to raise one. A recommendation therefore arrives as a score and an evidence-coverage figure. Corpus construction follows one rule: a fact with a source URL, or a blank with a recorded reason. Coverage measured across the corpus splits on the tier boundary, 163 curated entries at a median 100% against 236 at 40%, the latter missing the same four dimensions — the ones an automated pass could not recover reliably from public sources.

A live interface demonstrates the pipeline without account or installation. We name what the corpus does not contain, and make no claim about acceptance outcomes, which we have not measured.

Digitisation improves access to historical documents, but it does not, by itself, make them easy to navigate. Entity recognition and linking can identify recurring people, organizations, places, and events across an extended source. Yet the resulting dataset can appear authoritative before it has been reviewed, and translated or transcribed materials are especially affected by human or machine generated errors.

We present the HEXER Graph Explorer, a self-hosted browser-based application that addresses both challenges within a unified interface. First, it facilitates the construction and validation of a knowledge graph from a historical document. A selected language model generates candidate entities, relations, Wikidata matches, and each element is accompanied by its supporting text passage, confidence ranking, and explicit review status. Curators can compare ranked candidates, amend metadata, merge duplicate records, or leave uncertain entries unresolved. Their decisions are recorded in a non destructive manner, as appends to the original data, thus making them easily reversible. Furthermore, all decisions are logged and transparently displayed.

Second, the tool enables semantic rather than page-based exploration of the document. Readers navigate the corpus through thematic lenses, for example, vessels and military service, places and routes, peoples and languages. Filters based on entities or relations can trace any connection back to the sentences from which they are inferred.

The demonstration is based on a historical document, the journal of a Transylvanian sailor who was held as a prisoner of war in Japan from 1914 to 1919, where he wrote his memories, and whose account is available through Europeana.

Digital libraries provide access to research articles and related datasets and code, if available. However, even if a scientific statement is linked to supporting data, turning such relations into executable analyses requires separate workflows. This gap is increasingly urgent as computational reproducibility remains an imperative. Building on the TIB Knowledge Loom, which links scientific statements to supporting evidence, we present SciClaim2Explore, a system that turns statement-data links into entry points for natural-language exploration of research data supporting the claims conveyed by these statements. Starting from a selected scientific statement, in the proposed system a user asks a query about underlying linked, tabular data. The system uses the statement and schema context to generate and execute SQL, returns the resulting data table and, when appropriate, a data visualization. We evaluate the system on 84 claim-guided tasks from five scientific papers: 37 reproduce reported analytical results and 47 pose new analytical questions on the underlying data. In the grounding ablation, linked-dataset scoping reduces wrong-dataset queries from 14.3–28.6% to 0% and improves Answer Accuracy across three cloud models. These results suggest how scientific evidence links can provide executable context for exploring research data beyond artifact discovery. The source code is publicly available in the project’s repository¹. A video demonstrating the system is also publicly available ².

Libraries often experience two basic challenges when attempting to automate their traditional information organization processes: 1) overcoming practical barriers and 2) grappling with the intellectual and ethical compromises automation may impose. Generative AI applications have gained momentum as increasingly accessible tools for automating intellectual and creative tasks. Exploring how libraries are adopting AI to automate their real-world information organization work in relation to the challenges they face can illuminate areas where policymaking, development of library and information science curriculum, and further research efforts might focus. A panel for the JCDL 2026 conference is proposed to probe how academic libraries view the utility and capability of AI technologies to solve their information organization needs.

Scientific progress relies on acknowledging epistemic boundaries and charting forward directions. However, limitations and future work are systematically omitted or hedged to improve publication prospects. With escalating submission volumes, reviewers increasingly rely on Large Language Models (LLMs), yet zero-shot models generate superficial, hallucinated, or author-echoing feedback. This thesis presents a unified framework to discover, evaluate, recommend, and generate scholarly limitations and future work. First, we introduce LimTopic and FutureScope, using hybrid BERTopic, LLM pipelines to uncover thematic limitation patterns and model thirteen years of NLP future work trajectories. Second, to resolve metric inflation from surface metrics, we establish BAGELS and FutureGen, introducing atomic Limitation Units (LUs) for fine-grained PointWise evaluation and RAG-based future work generation with iterative critique loops. Third, we construct citation knowledge graphs with GraphSAGE to recommend paper-specific limitations. Finally, SciLimAgents couples intrinsic (LimitAgents) and novelty (NovAgents) specialist agents over a 120K-paper shared memory, aligned via on-policy preference optimization and composite process rewards. Across extensive benchmarks, our frameworks demonstrate substantial gains over frontier baselines, establishing a principled foundation for AI-assisted scientific quality assurance.

Electronic theses and dissertations (ETDs) hold detailed accounts of laboratory experiments, especially in the physical sciences. Digital libraries, however, expose ETDs mostly as bibliographic metadata and full-text PDFs. The experimental knowledge inside is scattered across chapters and hard to act on. My dissertation asks how digital libraries can turn these narratives into source-linked experimental records. The work proceeds in three stages, which identify experimental passages, represent them as structured records, and evaluate their utility for scientific tasks. Such records could speed up literature review and supply the machine-readable experimental knowledge that autonomous laboratories are envisioned to require. I report a completed pilot on eleven theses from a nanoplasmonics research group. It combines rule-based screening with a locally hosted LLM judge and yields 217 structured records, each linked to its source section and PDF page span. An expert from the source discipline reviewed 30 passages sampled across the theses. Eighteen were judged useful, nine too generic, and three unrelated to experimental methods. I seek consortium feedback on the scope of cross-domain evaluation, on validating the proposed readiness grading, and on bounding the representation work.

Running a language model on an edge device means fitting it into strict memory and latency budgets. Compression, chiefly pruning and quantization, is the standard way to achieve this, but it is capability-agnostic: it minimizes average error over the model’s overall behavior rather than the retention of any particular skill, and the aggregate metrics used to report compression quality conceal which skills were lost. The damage is amplified in exactly the small models edge deployment requires. This dissertation asks whether compression can be made capability-aware by borrowing the unit of analysis from mechanistic interpretability. The proposed pipeline localizes the circuits implementing a fixed capability profile, protects the parameters those circuits occupy while the remainder of the model is compressed, realizes that protection at a granularity the target hardware can execute, and verifies both behavioral retention and mechanistic faithfulness afterwards. I present the problem formulation, the gap identified in a targeted review of the compression and interpretability literatures, and the design of a first experiment measuring circuit damage under quantization.

Authors routinely point to the datasets and software their studies use through URLs in their papers, which we call OADS-URLs. Many no longer lead to the resource they point to. How much of the cited open access datasets and software (OADS) is still reachable is not well understood, leaving web archives no basis for deciding which endangered resources to preserve first. My dissertation measures the availability and discoverability of OADS-URLs at scale and how fast they decline. We built two benchmarks, sURL-3K and OADS-1K. On sURL-3K we evaluated our proposed extraction method, VisURLX, against heuristic baselines. VisURLX reads rendered page images with a multimodal large language model and is the most accurate, at an F1 score of 86.86%. On OADS-1K we developed EnSU, which classifies each URL by resource type and provider at a macro F1 of 90%. With the pipeline that best trades accuracy against cost and EnSU, we will build OADS Repo from about 16 million scholarly documents and measure how availability and discoverability differ across discipline, provider, host, venue, and institution, and how fast OADS-URLs disappear. The results will show how much of the endangered record can be preserved and which OADS to capture first.

To quantify web page longevity, our study samples from the Internet Archive’s Wayback Machine to curate the NYPW Sample, comprising 27.3 million HTML URLs spanning 26 years of web history (1996–2021). Our empirical findings indicate an average web page lifespan of 5.0 years (median 2.3 years), with a substantial disparity between root URLs (median 8.8 years) and deep links (median 1.3 years). Traditional quantitative analyses of webpage longevity generally consider a URL as a static identifier, which may not capture the complex lifecycles in which underlying content, ownership, or semantic purpose changes while a link remains active. To address this limitation, we proposed the Multi-Resident URL framework, which models six distinct lifecycle phases and formally separates structural URL stability from evolving content identity. Furthermore, our redirection analysis demonstrates the fragility of relocated content, revealing that only 50% of redirecting URIs resolve to active pages and illustrating how redirection shapes domain states and creates sink URIs. Through an integrated analysis of web resource survival, temporal stability, URL lifecycle evolution, and redirection behavior, we aim to improve the accuracy of web-archive studies, reduce reliance on detailed memento analysis, advance archiving practices, and deepen understanding of long-term web content persistence.

Social media content increasingly gets incorporated into U.S. television news. However, analyzing this cross-media flow remains difficult because social media references can appear in broadcasts through multiple modalities, including textual representations such as closed captions and OCR-extracted on-screen text, as well as visual representations such as platform logos and post screenshots. This research presents an infrastructure for identifying social media references in U.S. television news broadcasts. We used the Internet Archive’s TV News Archive and constructed Social2TV, a manually annotated 36-hour benchmark of CNN, Fox News, and MSNBC broadcasts from a single day from 2020 to 2024. We developed multimodal methods to detect social media logos and post screenshots, alongside text-based matching of social media references in closed captions. Preliminary evaluation for visual references achieved F1-scores ranging from 0.5053 to 0.8904 for social media logos and 0.3944 to 0.9057 for post screenshots, with recall as high as 0.9306 and 1.0000, respectively. Results show strong performance for textual detection, with F1-scores ranging from 0.9565 to 0.9941 across the three networks. These results demonstrate the effectiveness of multimodal methods for identifying social media references in U.S. television news. Building on these detection components, we propose a reference-level representation that aligns multimodal signals. The resulting infrastructure provides a foundation for future large-scale analysis of how, when, and in what forms information originating on social media enters television news.

Scientific data retrieval can identify potentially relevant datasets, but relevance alone does not establish whether those data are fit for reuse. Researchers may also need provenance and contextual information about data collection, methods, processing, prior use, and other conditions of data production to understand and evaluate candidate datasets. Although semantic, neural, and generative approaches are expanding possibilities for scientific data discovery, less is understood about how reuse-relevant evidence should be represented and surfaced as users move from retrieval to reuse assessment. My doctoral research investigates this transition by examining what provenance and contextual information scientific data reusers need, how scientific data repositories currently represent and surface such information during discovery and evaluation, and what design requirements follow for provenance-aware AI-mediated retrieval. The research combines comparative repository and interface analysis, semi- structured interviews primarily with scientific data reusers, with supplementary interviews with repository professionals, and evidence-to-requirements synthesis. The expected contribution is an empirically grounded set of design requirements for supporting the transition from scientific data discovery to informed reuse assessment.

Knowledge Graphs (KGs) offer a promising foundation for integrating heterogeneous resources, enabling complex querying and knowledge discovery. However, taking advantage of such knowledge infrastructures is often limited by the technical expertise required to construct, maintain, query, and explore them. Language Models (LMs) offer an opportunity to lower these technical barriers by providing natural-language interfaces and supporting the automation of otherwise complex knowledge management tasks. In this work, I present LibGraphLM, a modular language-model-based for framework combining domain-adaptable KG construction with human-usable exploration services. LibGraphLM assists in creating testbeds for investigating the interaction between KGs and LMs in the domain of digital libraries, which presents fundamental challenges such as sparse data, missing records, and heterogeneous resources. Preliminary modules address entity reconciliation and report creation based on the verbalization of connections between entities. Early results demonstrate the feasibility of these components, while highlighting open challenges related to evaluation, robustness, and generalization.

An AI-generated podcast companion article to augment the JCDL 2026 Undergraduate Research Symposium Papers Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

Recent scientific language models achieve strong performance by scaling model size and training on large domain-specific corpora; however, they primarily rely on raw textual data and lack explicit mechanisms for knowledge abstraction and reuse. In this paper, we present StemForce-LM, a framework for transforming unstructured scientific literature into structured, reusable supervision for language model training. The proposed approach generates concept-conditioned question–answer pairs that decouple underlying principles from document-specific context, enabling cross-document knowledge integration. The framework consists of four stages: domain corpus curation via citation graph expansion and prompt filtering, concept-conditioned knowledge extraction through constrained prompting, information integration via Q/A-based fine-tuning, and evaluation using debiased pairwise comparisons. We evaluate the framework on a stem cell mechanobiology testbed (StemForce-LM) and demonstrate improved domain-specific reasoning accuracy, strong performance against substantially larger models. These results indicate that structured knowledge extraction provides an effective and scalable alternative to raw data scaling for advancing scientific language models.

Large language models (LLMs) are increasingly used to summarize, annotate, and analyze literary texts, yet we know little about how their readings compare with a human interpretation. The open-ended nature of fiction makes that comparison hard: there is no single objective truth, and two careful readers can cite the same passages while organizing meaning differently. A single similarity score hides exactly this difference. We represent each complete interpretation as a structured record with four connected dimensions: the textual patterns a reader notices, the themes built from those patterns, how each theme develops across the story, and the global organization that holds the reading together, with every part linked to the passages that support it. Using this representation, we compare 46 independently produced interpretations of one 1,600-word story: one from a human reader and 45 from LLMs in a balanced design of 3 models × 3 reasoning settings × 5 independent runs. An alignment step lets one reader’s single element match several elements in another reading, so a difference in granularity is not scored as an error. The human and LLM readings cited largely the same passages and described matched ideas similarly (evidence similarity 0.68), but divided those ideas into different elements and connected them through different structures (relation overlap 0.04). Repeated runs agreed more within a condition (0.62) than across conditions (0.50), though conditions overlapped heavily. Some themes appeared in all 45 runs; others appeared only in the human reading or only in LLM readings. The human reading serves as a reference point, not necessarily as ground truth. The contribution is a layered comparison that shows where readings agree and where they differ: information a single score cannot provide.

Accurate representation of dermatologic conditions across skin tones is critical to equitable diagnosis and medical education because darker skin tones remain underrepresented in training imagery. Although generative artificial intelligence (AI) has shown increasing capability in synthesizing dermatologic images, systematic evaluations of whether these tools preserve both requested skin tone and condition-specific features when only skin tone changes remain scarce. To address this gap, we conduct a counterfactual evaluation of OpenAI’s ChatGPT and Google’s Gemini, holding all prompt attributes constant except the requested Monk Skin Tone (MST) classification, and generate 500 synthetic dermatology images across five conditions. Requested skin-tone accuracy was 52% for lighter tones (MST 01–05) and 44.8% for darker tones (MST 06–10). Among 242 correct-MST images, BioMedCLIP condition-matching accuracy was 68.5% for lighter tones and 57.1% for darker tones. These descriptive differences suggest that skin-tone fidelity and automated condition matching can vary across requested MST groups, motivating output-level auditing before synthetic images are considered for medical education. All code and data are available at:

Situational awareness during crises depends on synthesizing fragmented reports that vary across time, location, and source. Reliable crisis question answering must prioritize recent and nearby reports while reconciling updates that contradict earlier information. However, existing crisis datasets often provide insufficient evidence for systematically evaluating these capabilities. We audit evidence coverage across crisis-need queries and enrich under-covered cases with synthetic situational-awareness posts, constructing sequences of reports and updates that preserve spatiotemporal consistency. Using the enriched benchmark, we compare direct prompting, structured prompting, and ST-CrisisSynth, which applies query-dependent semantic, spatial, and temporal weights to re-rank evidence before generation. Human evaluation shows that ST-CrisisSynth achieves the highest correct-response rate (70.0%), improving by 3.3 percentage points over the strongest baseline. These findings suggest that selecting and organizing spatiotemporal evidence before generation supports more reliable crisis information synthesis.

Cardiomyocyte (CM) yield (the fraction of cells that become cardiomyocytes) is widely used to evaluate and select cardiac-organoid protocols. Cross-study comparison could reveal which protocol choices are robust, but identical-looking percentages often describe different constructs, denominators, measurement methods, or experimental purposes. Existing resources organize protocols and reporting fields but do not provide an executable rule that determines comparability per experimental condition. We address this gap with an arm-level comparability framework built on a simple insight: comparability is a property of a protocol arm, not of the reported percentage alone. Given the metadata for one protocol arm, the framework determines whether its reported value can enter a method-matched comparison and otherwise assigns a primary failure code. We apply the framework to a source-linked resource of 114 papers and 367 protocol arms. Although 111 arms (30%) report a cardiomyocyte fraction, only 31 (8%) satisfy the construct, outcome-role, and measurement requirements, and they come from 17 papers. Most exclusions reflect legitimate study design; a further 8 arms define a narrow but addressable reporting-and-measurement gap. An independent blinded v1.1 re-annotation of a 54-arm sample achieved 87% exact primary-label agreement (κ = 0.81) and 89% binary eligibility agreement (κ = 0.78) against the adjudicated reference. The framework converts superficially similar percentages into an auditable candidate set and separates structural barriers from reporting barriers. Together, the source-linked corpus, failure-code schema, and proposed reporting checklist provide a transparent, inspectable, and reusable basis for method- and context-matched cross-study comparison.

Internal migration shifts dominate demographic changes and place asymmetric strain on infrastructure and public services. The fragmented, institution-specific nature of Mexican municipal data has limited the study of large-scale dyadic modeling in Mexico. We make two contributions: first, we construct a dyadic dataset harmonizing ten institutional sources across Mexico’s 2,457 municipalities. Second, we propose the Spatially-Gated Deep Graph Neural Network (SGDG), an architecture that incorporates spatial correlation between neighboring regions through graph-based aggregation while jointly modeling climate and income through an income-conditioned gating mechanism. The SGDG’s performance is competitive with the current strongest migration models, achieving a higher CPC while offering an interpretability boost.¹

Large Language Models (LLMs) can produce fluent and confident responses even when those responses lack sufficient support or contain factual errors. Reliable question answering therefore requires models not only to generate answers, but also to recognize when they cannot provide a well-supported response and abstain accordingly. This capability is particularly important in scientific domains, where inaccurate responses can cause real-world harm. This paper systematically investigates the abstention behavior of LLMs in challenging scientific multiple-choice answering tasks. We evaluate several LLMs on multiple modified versions of a scientific multiple-choice question answering dataset, varying both the availability of the correct answer and the wording of the abstention option. Our findings demonstrate that LLMs often struggle to abstain consistently when the correct answer is absent, and that their abstention behavior is sensitive to model family, abstention wording, and reasoning settings.

Predicting whether a juvenile will reoffend is an important problem in juvenile justice. The Positive Achievement Change Tool (PACT) is commonly used to assess juvenile risk and needs, but machine learning models based on PACT data often achieve Receiver Operating Characteristic - Area Under the Curve (ROC-AUC) values of only about 0.61–0.72. This paper investigates whether this performance plateau is caused by the machine learning pipeline or by limitations in the information available in the PACT assessment. We compared several machine learning algorithms and encoding strategies, expanded the feature set using Risk-Need-Responsivity (RNR) theory, and added county-level socioeconomic and community variables. We also used SHapley Additive exPlanations (SHAP) to examine which features contributed most to prediction. The results showed that changing algorithms, encoding strategies, and PACT feature sets produced only small changes in ROC-AUC. Adding county-level information resulted in a small but consistent improvement across repeated evaluations. SHAP analysis showed that attitudes/skills and criminal history were among the most important feature groups, while the overall PACT intake risk score contributed little once the individual assessment items were included. These results suggest that the prediction ceiling is mainly related to limitations in the information available at intake rather than the choice of machine learning method.

An AI-generated podcast companion article to augment the JCDL 2026 Resource Papers Session of the JCDL '26

NOTE: AI content can contain errors and should be double-checked before use.

Archives, libraries, and museums at universities typically operate with separate cataloguing workflows, metadata standards, and information systems, which makes it difficult to trace people, places, objects, and documents across institutional boundaries and over long historical periods. At the Jagiellonian University, cultural heritage data are distributed across the Jagiellonian Library, the University Archives, and the University Museum, each preserving different types of materials and serving distinct scholarly and public audiences. This paper introduces CHExGraph, a cultural heritage knowledge graph that integrates metadata from these three units within a shared linked data environment. CHExGraph focuses on the early history of the university (1364–1780) and unifies person-centric biographical records, bibliographic records, and object-centred descriptions. The graph adopts CIDOC-CRM as its core conceptual model and extends it to represent academic trajectories, institutional affiliations, and object-related roles. The resulting data set comprising 862,877 RDF triples was evaluated over selected competency questions. The evaluation shows that CHExGraph enables unified querying and analysis of Jagiellonian University’s distributed cultural heritage resources. CHExGraph is released with in a Zenodo repository . It is accompanied with documentation intended to facilitate reuse and adaptation in other digital library and cultural heritage projects.

Scholarly PDFs contain metadata essential for building digital libraries. Although various tools and methods have been developed for extracting structured information from scholarly documents, existing evaluations largely assess the extraction of author names and affiliations independently, while resources for explicitly evaluating author–affiliation linking remain limited. We present a benchmark for this task, consisting of 507 manually annotated CC-BY arXiv papers. We evaluate several extraction tools and additionally analyze their performance across different document-class families. Our results show that GPT-5.6-sol achieved the best overall performance with rendered-page input, while the benefit of visual and layout information varied substantially across models and document-class families. The benchmark dataset is publicly available.

AI chatbots change continuously, but snapshot evaluations cannot show when answers changed or whether models became more cautious, verbose, current, or restrictive. We present Ask Me Twice (AMT), a public, growing archive created by the Internet Archive’s Wayback Machine team. AMT submits curated questions daily to an evolving model roster and preserves outcomes with question, model, date, and operational metadata. As of September 17, 2026, the public API reports more than 267,000 responses to 130 questions from 42 models, covering July 18–September 17 and growing by about 5,000 responses per day. Access includes an Explorer, a read-only REST API, daily bulk exports, an OpenAPI schema, and a Model Context Protocol server. This paper describes the resource’s design, schema, access, research uses, and limitations. Unlike static benchmarks and naturally occurring chat logs, AMT supports repeated-measures comparisons of the same prompts across models and days.

Web archive collections are increasingly large and heterogeneous, making it challenging for practitioners to choose efficient representations and workflows for common analytical tasks. This half-day, hybrid tutorial provides a hands-on introduction to web archiving file formats (including WARC, CDX/CDXJ, Parquet, and derived formats such as WAT and WET) and demonstrates how each format affords different kinds of analysis. Using the End of Term (EOT) Archive Datasets (over 2.5 PB) as a real-world testbed, we present a three-tier model of analytical queries ranging from lightweight aggregate statistics over indexes and columnar files, to metadata- and text-centric analyses over derived formats, to compute-intensive studies requiring access to raw WARC payloads. Participants will learn how to access EOT data on AWS S3, convert and query CDX(J) data in Parquet, perform example analyses using SQL (DuckDB) and Python workflows, and carry out link-structure and basic NLP tasks using WAT/WET. The tutorial highlights practical trade-offs among formats (size, cost, and computational requirements) and equips attendees with concrete tools and patterns for scalable web archive collection analysis.

AI for Organizing Digital Collections is a half-day tutorial that introduces practical approaches for applying AI to organize textual, visual, and multimodal digital collections. The tutorial focuses on how Large Language Models (LLMs) can support classification, subject analysis, and vocabulary mapping, and how Vision-Language Models can support archival image description and visual subject term assignment. Through accessible explanations, demonstrations, and hands-on exercises, participants will learn how to train LLMs to support digital collection organizations, evaluate model outputs, and design AI-assisted workflows. The tutorial is intended for digital library researchers, metadata professionals, catalogers, LIS educators, students, and cultural heritage practitioners interested in using AI to improve collection organization and discoverability.

The 2026 edition of the Workshop on Web Archiving and Digital Libraries (WADL) will explore the integration of web archiving and digital libraries. The workshop aims to address the full life cycle of digital resources, including creation/authoring, uploading/publishing, crawling, indexing, exploration, and archiving. It will also examine processes and tools for archiving “non-traditional” resources, including social media, scholarly and government datasets, 3D objects, and digital online art.

Over the past decade, the field of artificial intelligence (AI) has witnessed significant advancements across various domains, such as natural language processing and computer vision, which have further accelerated discoveries, innovations, and breakthroughs in many interdisciplinary fields — including Science of Science, which relies on big data to uncover reproducible patterns governing the evolution of research topics, collaboration networks, knowledge diffusion, and the systemic factors that drive scientific progress. Today, Science of Science is incorporating AI techniques into the analysis of big data, yielding deep and compelling insights into both longstanding and emerging problems in the field. Building on the success of the first and second AI4SciSci workshops, held in conjunction with ICDM’23 and JCDL’25, respectively, we propose the Third International Workshop on Artificial Intelligence for the Science of Science (AI4SciSci) to further engage the AI and Science of Science communities in promoting research and discussion on frontier topics in this area. The full-day workshop will be held in person, co-located with the 2026 ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL’26).

Artificial intelligence (AI) is reshaping libraries from traditional resource repositories into intelligent, adaptive, and user-centered knowledge service hubs. Across public, academic, and special libraries, AI technologies are increasingly embedded in information discovery, reference services, knowledge organization, user interaction, decision support, and operational workflows. These developments create new opportunities for digital library innovation, while also raising critical questions about trust, accountability, data privacy, algorithmic fairness, professional ethics, and inclusive access. This workshop examines the future of libraries in the age of AI through the dual lenses of service transformation and governance challenges. Rather than viewing AI simply as a tool for improving efficiency, it explores how AI reconfigures library services, reshapes user–library interactions, and challenges existing models of institutional governance and professional responsibility. The workshop will bring together digital library researchers, LIS scholars, library practitioners, AI researchers, and policy-oriented stakeholders to discuss empirical findings, theoretical frameworks, case studies, design approaches, and governance models. By identifying emerging research agendas and practical pathways, the workshop aims to advance sustainable, equitable, accountable, and human-centered library development in the AI age.

Scholarly metadata is an essential element in the scholarly ecosystem, empowering discovery, interoperability, and provenance and its efficacy relies on metadata quality and coverage. However, with the substantial increase in the number and diversity of research artifacts over time, it has become a cumbersome task to harvest metadata features for scholarly publications, datasets, source code, and machine learning models. These artifacts generate heterogeneous metadata features, which are distributed across various repositories. In addition, large language models (LLMs) and agents have emerged as powerful tools and are now being extensively adopted to facilitate a wide range of tasks across domains, ranging from information extraction to knowledge graphs. Nevertheless, numerous open questions persist regarding their adoption in diversified research disciplines, especially concerning their accuracy, robustness, biases, and societal impact when applied in the context of research metadata, scholarly information extraction, and open science. This workshop aims to cover these exciting aspects centered around metadata features from research artifacts, with a focus on large language models and knowledge graphs.