The Workshop on Artificial Intelligence for Scientific Publications (WASP) brings together research on methods that read, organize, connect, and evaluate the scientific literature. Its fourth edition was held online as part of AACL-IJCNLP 2026 on November 10, 2026. It received 43 submissions and accepted 25: 15 main-track papers and 10 system papers for the AstroCLIMB shared task on linking astronomy figures and captions to the literature. This overview aims to give the community, including those who did not attend, an accessible account of what the workshop presented and found. For each main-track paper, it summarizes the approach, evaluation, main results, and the authors’ interpretation and next steps. We then draw out shared themes and future directions. Many contributions are deployed systems built with archives and curators. They combine language models with deterministic validation, keep source evidence attached to extracted values, and design evaluations around scientific use. Several papers find that lean, focused systems perform as well as more elaborate ones. The authors highlight directions such as automatic verification of grounding, shared gold standards, normalization of quantities and entities, figure understanding, and checking that citations are relevant and real.

Scientists rely on figures to share their discoveries, but this makes the information contained in the figures difficult to parse, archive, and search. Multimodal neural-network models, designed to handle multiple types of data, promise to extract this information, but have not yet been widely tested and adopted by the astronomy community. In partnership with astroexplorer.org, we propose a novel dataset with over 100K figures and captions, and an associated task as a benchmark to evaluate a model’s multimodal capabilities. The task is to reconstruct the citation graph of astronomy papers solely from their figures and captions. AstroCLIMB was the shared task for WASP 2026. It attracted 15 participants on the open competition platform Kaggle, of which 10 submitted system papers. Results show that zero-shot methods and general purpose metrics underperform compared to finetuned methods, indicating that scientific figures are still hard to understand for currently available multimodal models. The dataset, benchmark task, and associated code are publicly available online.

GCN Circulars distribute crucial information on astronomical observations of transient phenomena, but their flexible prose mixes observations, limits, metadata, contacts, and interpretation. We present an auditable pipeline that converts these reports into event-linked, structured records. Regular expressions provide domain hints; a large language model returns strict schema-constrained JSON; deterministic stages normalize values and emit review flags. Seven domain reviewers assessed 231 outputs from 210 distinct Circulars, fully confirming 223 reports and marking eight partly correct. They confirmed 25,827 of 25,880 definite field decisions (99.80%); errors concentrated in timing and facility attribution. Comparison with independent redshift datasets found semantic catalog disagreements but no values misread by our pipeline. Applied to the GCN Circular archive back to 2016, the pipeline has structured 26,811 reports into 68,393 observations, now served in real time by the Astro-COLIBRI platform. We release the extraction pipeline as an open-source package and the derived records as a continuously updated resource.

Researchers need efficient and reliable methods for collecting high-quality information from trusted sources, but modern tools for AI-assisted research still suffer from the tendency of Large Language Models (LLMs) to produce factually inaccurate or nonsensical output, commonly referred to as hallucinations. We apply the extractive question answering system VerbatimRAG (CITATION) to research papers in the ACL Anthology[<https://aclanthology.org/>], directly mapping user queries to verbatim text spans in retrieved documents. We contribute a novel ground truth dataset for the task of mapping user queries to relevant text spans in research papers, and use it to train and evaluate a variety of extractive models. Human annotation is performed by NLP researchers and is based on synthetic user queries generated using a custom pipeline based on the ScIRGen methodology (CITATION), paired with chunks of research papers retrieved by VerbatimRAG. On this benchmark, a 150M-parameter ModernBERT token classifier trained on silver supervision from our pipeline achieves the best word-level F1 (53.6), ahead of the strongest evaluated LLM extractor (48.7).

We study AstroCLIMB, a four-way relationship-classification task over pairs of astronomy figure/caption endpoints. Starting from a leakage-audited lexical and low-level image-statistics baseline (M7), we evaluate a hosted multimodal judge under a fixed, auditable protocol. GPT-5.6 Luna receives only the two endpoint contents and produces integer class probabilities for three views in both endpoint directions. On 2,000 frozen validation pairs, all 12,000 calls succeeded. Equal-weight fusion reaches macro-F1 0.67, versus M7’s 0.40 (paired component-bootstrap delta +0.27, 95% CI [+0.24,+0.29]); Luna-only reaches 0.68. These are conditional local-validation results, not an official Luna leaderboard score. We therefore report an auditable system study without claiming paper-level generalization.

Meta-analysis is the synthesis of information from multiple sources to arrive at an overarching conclusion. There is a large need for meta-analysis in agricultural research to synthesize what is known and analyze overarching patterns. Extracting data from published literature is, however, labor-intensive, time-consuming, and tedious, and is impeded by a lack of standardization in research design, units of measurement, and terminology. These challenges are particularly evident in the domain of crop species mixtures, also called intercropping. With the growing capabilities of LLMs, many recent attempts have focused on building systems and tools to automate data collection, yet rigorous assessment against human-labeled ground truth is often missing. In this research, we evaluate three LLM-based approaches—direct zero-shot prompting, a staged workflow, and a multi-agent system—with six open-weight models to extract data from the intercropping literature. The results are evaluated against the manually curated ground truth and through a downstream statistical analysis. Overall, direct zero-shot prompting is the strongest and most consistent approach, achieving the highest mean similarity-adjusted F1 of 0.577, although none of the approaches is close to fully accurate. In the downstream analysis, most model–approach combinations recover the direction of the relationship between the predictor and outcome variables, but do not estimate its magnitude accurately.

AstroCLIMB asks whether two objects taken from astronomy papers, each a figure or a caption, show the same figure, come from the same paper, come from papers linked by a citation, or are unrelated. We describe the two-stage system of team HCMUS_Beyonders, which reached a macro-F1 of 0.7421 on the full test set. In the first stage, cross-encoders read both objects of a pair jointly. Each figure is represented by the text that optical character recognition (OCR) extracts from it, an input that scored higher than every image-embedding and image-patch input we tried. Because test papers are unseen, the cross-encoders are pretrained on 600K pairs generated from the citation graph of the public AstroCLIMB corpus. An error analysis shows that this stage most often confuses pairs from the same paper with pairs from citing papers, and that none of the cues we tested separates the two. The second stage therefore adds a different kind of model rather than more features: a Qwen3-8B large language model (LLM) fine-tuned with low-rank adaptation (LoRA). On its own it scores lower than the first stage, but over a third of each stage’s errors are not made by the other. Averaging the two raises held-out macro-F1 by 0.016 (95% confidence interval [+0.002, +0.030]) and test macro-F1 from 0.7072 to 0.7329. Adding a second LLM, Gemma-4-12B, without held-out validation brings the test score to 0.7421, a gain within test noise.

Resolving correct citations can be a time-consuming task for research writers. We address local citation recommendation (LCR), where the goal is to retrieve appropriate references for a given sentence. Web tools generally depend on the existing citation graph, which may bias results toward citation count or more recent publications. LCR research using full paper texts has been limited by the lack of publicly available data. We present a new dataset, the Local Citation Recommendation Corpus (LCRC), containing 14,735 scientific statements derived from astrophysics review papers, each with one or more citation targets whose full text is accessible via API. Using this dataset we systematically compare retrieval methods and develop a citation recommendation system that avoids citation graph bias while achieving 8.0% improvement in Recall@25 and 6.9% improvement in NDCG@25 over dense retrieval baselines without model fine-tuning. To support reproducibility, we make our code and dataset publicly available.

Tracing how knowledge diffuses through scientific networks reveals how research builds on prior findings, but doing so at scale remains challenging. We present an LLM-based approach, Citation-Aware Research Diffusion Summarization (CARDS), which describes how knowledge diffuses through the scientific landscape. We trace citations from research funded by the U.S. National Institutes of Health through the landscape quantified by CSET’s Map of Science cluster dataset to generate Research Diffusion Pathways, and compare LLM-generated summaries based on (1) cluster-level metadata and (2) citation-aware connections between research clusters. We find that CARDS outperforms cluster-based summaries under both human annotation and LLM-as-a-judge evaluation. Our work outlines a methodology for generating research diffusion paths, and for summarizing them with greater specificity.

The majority of historical experimental data across the empirical sciences remains locked in unstructured legacy formats, which slows data-driven discovery. We introduce an automated multi-modal pipeline that processes large literature corpora, filters non-empirical noise, and outputs structured datasets. Layout-aware document parsing is paired with a lightweight semantic triage gatekeeper to substantially reduce the computational overhead of applying frontier multi-modal models at scale. A generative model reconstructs experimental trajectories from text and figures, while a schema-constrained prompt and a strictly typed programmatic validation layer check physical bounds before serialization. We evaluate the pipeline on a 60-document corpus of extreme-condition thermodynamic literature drawn from Low Energy Nuclear Reactions (LENR) research, reporting concrete throughput, triage, and constraint-violation statistics. We distinguish safeguards enforced in code (physical-bound validators) from those requested through the prompt, such as source grounding, and outline the evaluation needed to verify the latter. The architecture is domain-agnostic by design, and we outline how it could transfer to catalysis, energy storage, and preclinical biomedical research.

We present GRAIL, a system for classifying relations between astronomy figures and captions in the AstroCLIMB shared task. GRAIL combines frozen CLIP and SciNCL representations with optical character recognition (OCR), lexical overlap, and figure layout. A factorized gradient-boosted classifier predicts four relation classes from 190 symmetric pair features. The fixed system achieves 0.4802 macro-F1 under cross-validation that separates exact objects across folds and 0.48335 on Kaggle. A variant with nested classifier selection and decision offsets obtains 0.4769 and 0.48264, respectively. Removing OCR or corpus-weighted lexical features lowers local macro-F1 by 0.0081 and 0.0128. These development comparisons support the use of explicit textual evidence alongside semantic representations. Additional model selection gives no observed aggregate gain, and direct citation relations remain the hardest class for both variants.

We describe a submission to the AstroCLIMB shared task at WASP 2026. The task requires classifying a pair of astronomy figures and/or captions as the same figure, from the same paper, from citation-related papers, or from unrelated papers. Our system casts classification as constrained next-token prediction with Qwen3-VL and adapts the model using 4-bit quantization and low-rank adapters. On the 10,000-pair Kaggle test set, three Qwen3-VL-8B models with language-attention adapters obtain 0.73230–0.73307 macro-F1 individually. Averaging their four-class probabilities raises the score to 0.73604, and masking the impossible same-figure class for caption–caption and image–image pairs produces 0.73725. This paper documents empirical system engineering rather than proposing a new learning algorithm; broader visual adaptation, restricted-loss blending, and citation-focused auxiliary training do not improve the result.

The WASP 2026 (AstroCLIMB) task requires classifying the relation between a pair of objects drawn from astronomy papers — same_figure, same_paper, related_papers, or unrelated_papers — where each pair combines text (captions) and/or images (figures) in one of three disjoint modality configurations: image+text, text+text, and image+image. Since same_figure cannot occur outside the image+text subset, a single unified classifier is structurally sub-optimal; we instead adopt three independent per-modality classifiers whose predictions are pooled for the competition metric (pooled macro-F1). Starting from a hand-engineered baseline (CLIP, SciNCL and DINOv2 embeddings with lexical/OCR features and a LightGBM ensemble), we find that replacing the encoder with a joint vision-language embedding model (Qwen3-VL-Embedding-2B) produces the largest single improvement among the experiments we tested (public leaderboard 0.4808 → 0.62319), exceeding the combined effect of six feature-engineering ideas that were implemented and measured but produced less than 0.01 macro-F1 each. We further show that zero-shot cross-encoder rerankers (Qwen3-VL-Reranker-2B, Qwen3-Reranker-4B), applied with no task-specific fine-tuning, add a second large improvement on top of the engineered ensemble (leaderboard 0.64210 → 0.66299), and that an untrained text+text reranker alone already outperforms the fully engineered pipeline on that subset (macro-F1 0.7111 vs. 0.6704). We report a parallel distant-supervision branch that trains on 120,000 externally-labeled pairs mined from the competition’s source corpus and reaches a comparable but lower leaderboard score (0.57713) through a different mechanism; under our tested data and modeling regimes, this suggests that representation quality was a stronger bottleneck than either additional hand-engineered features or additional distant supervision. All reported numbers are traced to raw out-of-fold prediction arrays or literal notebook output under a shared leak-free StratifiedKFold(5) protocol, and we document two negative results (a broken nested-honest validation split, and several feature-engineering ideas that measurably hurt) as concrete engineering findings.

Supervised sequence-labeling approaches to keyphrase extraction operate at token level, but whether a phrase is relevant may depend on context beyond its sentence. Processing an entire document at token resolution can increase sequence length and introduce irrelevant information. We investigate a mixed-resolution context-fusion architecture that retains token representations for the target sentence while compressing available surrounding sentences into pooled representations. Both representation types are placed in document order and processed jointly by a self-attention module before IOB prediction. We evaluate the model on \textsf{kp20k}, \textsf{inspec}, and \textsf{kptimes} against SEKE, a recent state-of-the-art supervised system with approximately 60% more total parameters. The model achieves comparable F1 scores on the two scientific-domain benchmarks, while a larger performance gap remains on the journalistic benchmark. Error analysis shows that the model generates substantially more candidates than there are gold keyphrases. However, many predictions counted as false positives under exact-match evaluation partially overlap with gold keyphrases, while exact matches tend to receive higher scores than partial and non-matches.

Although generative AI increasingly reformulates scientific texts, evaluations based on fluency, readability, and semantic similarity may overlook changes in the epistemic and argumentative properties of the reformulated texts. This paper introduces a Claim-Preservation Framework (CPF) to assess whether AI-generated reformulations preserve not only propositional content, but also epistemic force, evidential relation, argumentative function, and qualification and attribution. The framework distinguishes stylistic variation from consequential shifts in modality, hedging, evidentiality, causal interpretation, and argumentative function; its analytical potential is illustrated through an exploratory proof-of-concept application to 40 SciFact claims reformulated under a standardized instruction. This application demonstrates how the proposed dimensions can be used to identify and classify potential shifts in scientific meaning, but it is not intended as a full validation of the framework or as a general estimate of LLM behavior. By conceptualizing reformulation as a transformation of scientific meaning, the paper proposes claim preservation as a distinct dimension of LLM evaluation and outlines future research on computational methods for detecting epistemic and argumentative shifts in scientific publications.

To assess the scientific impact of observational facilities, Astronomers analyze publications which have made use of the data by the facility. This requires bibliographies annotating and linking data products with the literature, allowing bibliometrics to measure impact of the data. Creating bibliographies is laborious, involves specialists searching the literature for names, acronyms and identifiers, then determining how/if observations were used in those publications.

These bibliographies are useful beyond metrics: from the perspective of a research scientist, the data-literature connections provide a critical path to data discovery and access. Thus, by leveraging the efforts of librarians and archivists, we can use telescope bibliographies to support the scientific inquiry process. To this end, we introduce the TRACS: Telescope Reference and Astronomy Categorization dataset and Shared task, consisting of over 89K public English texts from space telescope bibliographies, categorized using a new, simple taxonomy based on discussion with established bibliographers.

We hope TRACS will be used to train modern Machine Learning and Artificial Intelligence models that in turn will help data curators build bibliographies. We evaluate how off-the-shelf Large Language Models perform on automatic bibliography curation as a baseline analysis. Results show the difficulty of the task and the need for custom tools.

We adapt Qwen3.5-4B to the AstroCLIMB task, which asks whether two scientific figures or captions come from the same figure, the same paper, citation-linked papers, or unrelated papers. For caption–image pairs, we first adapt the model on 80,000 public pairs labeled automatically from figure, paper, and citation structure in the public corpus, then train on gold data. Across three gold-training seeds that share one public checkpoint, this initialization improves caption–image development macro-F1 by 0.067 on average over gold-only training under matched gold schedules, although with more total compute. Answer generation and the reinforcement-learning variants we tested do not beat the classifier. The combined system of three modality-specific classifiers reaches a pooled development macro-F1 of 0.756 and a Kaggle leaderboard score of 0.751. Repeated development selection and earlier project exposure to competition materials confine these findings to the evaluated setting, and we document both.

The repositories of the NASA Extragalactic Database (NED) and NASA Exoplanet Archive (NEA) provide the scientific community with regularly-updated records of published data on observed astronomical objects. Over their years of operation, both archives have seen increases in the rate of publications that are relevant to their holdings. We are developing an AI-powered data extraction tool to increase the rate at which NED and NEA can process published articles into their databases. This is accomplished through a multi-stage pipeline that combines large language models (LLMs) with programmatic techniques and human supervision to create formatted database load files. Pipeline evaluations on sampled human-made load files estimate a macro-level recall of 0.88 for NED and 0.79 for NEA with typical runtimes of 20 minutes per-paper. While this tool is still being developed (aided by human feedback), present results suggest that contemporary LLM-based approaches are useful for complex structured data extraction. Future work remains to improve the pipeline’s precision and further adapt it to the wide variety of articles present in the literature.

Astronomers often have difficulty determining whether an object, such as a star or galaxy of interest, has been discussed in the literature. Existing tools enable object identification and linking in papers but do not record whether an object was genuinely discussed in a paper or just appears in a table or catalog. We address this by using language model agents to identify galaxies discussed in more than two sentences across a corpus of 156,006 arXiv astro-ph papers and extract those discussions. We extract 44,332 discussions of individual galaxies containing 85,314 quotes from 18,641 papers. With sky coordinates for 98.2% of discussions, we group them into 13,616 galaxies. We bring these results together into GalaxyWiki, where users can find literature discussing galaxies, read what was said about them, and follow links to the source papers. The code, dataset, and webpage are available at https://github.com/NolanKoblischke/galaxywiki-paper.

Figures are essential for scientific communication, and the field of Visual Question Answering (VQA) aims to test the figure understanding capabilities of large multimodal models (LMMs). Here, we investigate the accuracy of several low-cost LMMs on domain-specific VQA tasks for astronomical images of the sky. As instrument for this task we build a domain-specific VQA dataset comprising 2001 partially-synthetic figures divided evenly between contour plots, real telescope images drawn from NASA’s SkyView, and "fake" sky images which share the axes and angular scales of the real images but draw their colors from a gaussian mixture model, alongside 47357 questions. Each figure is released as pristine and with realistic page-aging degradations mimicking the historical astronomical corpus. Testing three low-cost proprietary LMMs, we find models frequently fail to answer domain-specific questions at high accuracy. An initial comparison with 104 human-verified, astrometrically tagged figures from historical journal publications shows similar median errors in angular extent. All code is available on GitHub: https://github.com/jnaiman/SkyImagesWASP2026 and dataset is available on HuggingFace: https://huggingface.co/datasets/ReadingTimeMachine/astro_sky_image_vqa.

We describe our system for the AstroCLIMB shared task (WASP 2026), which reconstructs the citation graph of astronomy papers from scientific figures and captions alone. The task requires identifying the relationship between pairs of figures and captions as same_figure, same_paper, related_papers, or unrelated_papers. Our system finetunes google/gemma-4-E4B-it using QLoRA on 8,200 competition training rows with a shared descriptive prompt and greedy decoding. A zero shot baseline achieves 0.4408 macro F1 on a balanced 1,800 pair validation split held out from the competition training file, and the finetuned system substantially improves this citation-relation classification for the challenging related_papers class in particular. Our submitted system achieves 0.64282 F1 on all 10,000 official test pairs. These results show that multimodal finetuning can improve citation-relation inference from scientific figures and captions when document metadata is unavailable.

Experts and laypeople alike increasingly encounter the scientific literature by posing questions to large language models (LLMs) rather than through database searches, making a model’s cited sources the reader’s main window into the original literature. Yet LLMs frequently suffer from citation hallucinations. Because source existence is objective and externally checkable, it provides an unusually clean signal for reinforcement learning with verifiable rewards (RLVR). We train LLaMA-3.1-8B-Instruct and LLaMA-3.2-3B-Instruct with GRPO and LoRA against a deterministic verifier that resolves each generated reference against scholarly databases CrossRef and OpenAlex. On 40 common medical questions in the MedAESQA dataset, the hallucination rate decreases from 85.5% for the base 8B model to 76.7% (p < 1e-4) for the best post-trained checkpoint. The post-trained model produces 71% more verified references at unchanged citation density, generates fewer misleading DOIs (down to 14% vs 54% for the base), and favors canonical sources such as the New England Journal of Medicine. A six-part audit finds no evidence of reward hacking. Ultimately, this deterministic verifier supplies a practical, objective signal that reduces hallucinated references, offering a novel method to improve citation accuracy and a stepping stone toward fully verifiable LLM answers for literature-based QA.

This paper describes our submission to the AstroCLIMB shared task, which evaluates whether multimodal models can identify citation relationships among astronomy papers from pairs of figures and captions. We explore an empirical progression spanning frozen embedding baselines, dual encoders, cross-encoders, and generative vision–language models (VLMs) adapted via parameter-efficient fine-tuning (LoRA with 4-bit quantization). Among our evaluated configurations, a fine-tuned Qwen3.8-27B achieves a test Macro-F1 of approximately 0.74. An auxiliary specialist adapter trained to re-rank the frequently confused same paper and related papers classes yields comparable performance within expected run-to-run variance. Across input modalities, Text–Text pairs achieve the highest classification accuracy, followed by Text–Figure pairs, while Figure–Figure pairs prove the most challenging. These findings demonstrate that visual similarity alone remains an insufficient proxy for semantic citation linkage in specialized scientific literature.

In scientific articles, figures often convey experiments and observations, as well as physical intuition, to the readers. Yet, surface metadata alone can be insufficient to unpack how two scientific objects relate to each other. We explore this with AstroCLIMB, which infers relationships between pairs of figures and metafigures based on four classes: same figure, same paper, related papers and unrelated papers. We design a multimodal pairwise pipeline that consolidates the textual visual semantic, and domain-specific signal sources from object pairs. TF-IDF and SBERT text and image-based text-image feature pairings, combined with lexical overlap, entity, and instrument overlap, numerical overlap, and interaction features conditioned on methods, comprise the feature set fed into a weighted ensemble of LightGBM, XGBoost, and CatBoost trained with 313 engineered features on five-fold stratified cross-validation, optimizing for class-specific thresholds to produce a Macro-F1 of 0.5016 on out-of-fold data. Results show that yet stronger textual and multimodal counterpart representations exhibit complementary yet useful pairwise signals, although establishing related papers still remains a significant challenge.

We introduce Scientific Table Generation (STG), the task of extracting information from scientific documents to populate a table. We argue that STG can serve as a unifying task for structured knowledge extraction from scientific literature. To support this goal, we present the first multi-disciplinary benchmark for STG with predefined schemas and multi-record extraction, covering six scientific domains and comprising 20 datasets unified under our proposed formulation. We detail the benchmark’s construction, including dataset composition, and introduce an STG evaluation protocol. We analyse the results of zero-shot large language model (LLM) and vision-language model (VLM) pipelines with typical document processing AI methods for the STG task. In doing so, we identify key challenges of the STG task and propose a set of open challenges, including multimodality, table and figure understanding, normalisation of data, and treatment of reported quantities. All code and data accompanying this work are publicly available at[This repository will save researchers significant time. Instead of manually searching for 20 different datasets, researchers can use our repository to download all datasets directly, as a download link is provided for each paper. In addition, researchers can use our evaluation framework, which streamlines the process and improves overall efficiency.]: https://github.com/bolucunecva/table_generation

Repositories usually report metadata quality as completeness, the share of fields filled. We measure instead whether two deposited analyses record enough to be compared. Over all 167,331 pairs of gas chromatography analyses in Metabolomics Workbench, 120 pairs meet the necessary conditions for comparing retention indices and two meet those for comparing retention values; one field, the retention index reference series, recorded for 19 of 579 analyses, removes 99.9% of the pairs that share a stationary phase. The oven temperature programme is usually written in the protocol prose instead, and the series sometimes is, so we define their recovery as an extraction task with ordered output and explicit abstention. Audited by two blind LLM raters whose error detection we measure on planted errors, the programme extractor is 65% precise on 40 random MetaboLights recoveries with no fabricated value, recovers in full 49% of the programmes stated in a separate random sample of 40 protocols, and the series extractor is correct on 57 of 66. On the census corpus itself, recovered series raise the pairs that can compare indices from 120 to 171, or from 78 to 92 when the two series must be identical. Of 239 populated Workbench programme fields, 130 hold only a range, a single temperature or a pointer to a paper, yet a completeness audit counts all of them as filled. Every released value carries its source sentence and is attributed to the extracting pipeline, not to the depositor.

We present a modality-specific system for the AstroCLIMB shared task, which predicts relations among astronomical figures and captions. Caption–caption, image–image, and image–caption pairs use different combinations of frozen pretrained representations and supervised classifiers. Image–caption prediction first applies a Qwen3-VL 4B/8B exact correspondence gate before classifying the remaining paper relations. Subsequent improvements add astronomy-domain sentence embeddings to the caption branch and a learned visual-projection specialist to the image branch. Across eight Kaggle submissions, macro-F1 increases from 0.3641 to 0.6096. Adding astronomy-specific caption features improves five-fold pooled out-of-fold (OOF) macro-F1 from 0.6071 to 0.6388; the image specialist offers a smaller incremental test-set gain. The direct citation relation, related_papers, remains the most difficult. Results support selecting representations and decision mechanisms according to modality, while highlighting the limitations of recovering citation links from isolated content.