When a paper says two people "contributed equally," the signal is usually a tiny asterisk next to two names and a footnote at the bottom of the page. Citation indexes routinely lose it and fall back to author order, so the second co-first author is recorded as a secondary contributor. Published corrections show how much this matters to the people involved. For example, Vishnu et al. note that a missing co-first marker affects grant eligibility and career advancement. This proposal treats the recovery of these markers as a test of a broader machine learning claim: models generalize to new conditions because their training data is varied, not because it is large. The task is a good test because the visual form of the signal changes sharply between publishers. The marker might be an IEEE asterisk, a Nature dagger, or a Springer unnumbered footnote, and the parser has to link that glyph to specific names on the page.
The concrete question is as follows. With a fixed budget for labeling paper headers, should those labels be spent on many pages from one source, on pages spread evenly across publishers, or on pages chosen to maximize a measured geometric diversity of header layouts? The measure is PLDiv, which summarizes the shape of the data as gaps, clusters and loops in an embedding space. We will compare these options on how well a layout-aware document model identifies co-first and co-corresponding authors in publisher layouts it has never seen. The answer gives a useful extraction tool and also a controlled test of whether shape-based diversity predicts out-of-distribution transfer better than simpler proxies.
Core Claim
We claim that, at equal labeling budget, a parser's author-level accuracy at identifying co-first authors on held-out publisher layouts depends more on the geometric diversity of its training headers than on page count. We further claim that selecting training pages greedily by PLDiv beats selecting them stratified by publisher. The second claim is the one that could fail, and we treat it as an open hypothesis rather than an assumption.
The work is new in three respects. It defines a task that existing document benchmarks collapse away. It studies how training sets should be composed for that task. It also uses the published-correction literature as a natural source of labels and as a measure of how often the marker is lost upstream, before any parser sees the page.
Positioning
This section groups prior work by the role each paper plays in the proposal.
Two papers motivate the work by showing the gap. A Review of Microsoft Academic Services describes a web-scale scholarly graph that assigns equal credit to all authors as a stated simplification and never parses contribution markers. Cleaning Noisy and Heterogeneous Metadata links noisy CiteSeerX headers to clean records by title, author string and year. That paper shows that headers drawn from many sources are noisy enough to break simple matching, but it never extracts author roles.
Two papers supply the starting method and expose its blind spot. DocBank builds 500,000 arXiv pages labeled automatically by coloring LaTeX source. On those pages LayoutLM, a transformer that reads text tokens together with their positions on the page, reaches strong overall scores, but it struggles specifically on the Author category. DocBank labels all author material as one coarse box, and every page comes from a single source. Document AI: Benchmarks, Models and Applications surveys such models on forms and receipts, not on how credit is attributed in research papers. We reuse DocBank's weak-labeling idea and its models but split the author region into finer pieces and draw pages from many publishers.
Several papers support the diversity hypothesis in other modalities. Le Scao et al. find that a curated multi-source text corpus beats larger uniform web crawls at equal compute. Larrazabal et al. find that a gender-balanced chest X-ray training set beats larger imbalanced ones. WIDER FACE shows that face detectors trained on controlled data fail on extreme variation. None of these papers separates two different kinds of diversity: diversity counted by source labels, and diversity measured geometrically. Measuring Dataset Diversity from a Geometric Perspective supplies the geometric measure, PLDiv. Its authors validated it against human judgments of diversity and against synthetic ground truth, but never as a predictor of downstream generalization. That is the use we test.
One paper constrains the scope. The authorship dilemma models credit splitting between two authors as a game and shows that equal splitting encourages matched effort. That result explains why a recorded split of 1/2 should not silently become first-author-takes-most. We use it only to motivate a credit-distortion metric. We do not attempt to extend the game, and we drop the data-valuation framing of authorship sketched in the research file, because it would require a separate project.
Research Program
The first step is to build a header benchmark with fine-grained labels. The token classes are author name, affiliation mark, equal-contribution symbol, equal-contribution statement and corresponding-author symbol. For each paper we also record author-level labels saying who is co-first and who is co-corresponding. Pages come from arXiv, where the DocBank-style LaTeX coloring gives labels at scale, and from several publisher families: IEEE, Springer, Nature-family, ACM and Elsevier. Publisher pages are labeled by hand under a fixed budget, with two annotators per page and adjudication of disagreements.
The second step uses the correction literature. Correction notices that add a missing co-first marker give verified ground truth for the corrected versions of the papers named in them, such as Zhao et al., Wiesenfarth et al., Park et al., Lennon et al., Lu et al. and Zhang et al.. These notices also force an honest distinction. In these cases the original published layout lacked the marker, so no parser could have recovered it. We therefore report two separate failure rates: markers lost by the parser, and markers lost upstream by the publisher. This keeps us from overstating what better extraction can fix.
The third step is to embed every header region with a frozen layout encoder and compute PLDiv for any candidate training set. Using a fixed budget of labeled pages, we build training pools by five rules. The first takes arXiv pages only. The second takes random pages from all sources. The third stratifies evenly across publishers. The fourth adds pages greedily to maximize PLDiv. The fifth adds pages greedily to maximize Vendi Score, a measure of diversity based on how spread out the embeddings are. The same parser is trained on each pool and evaluated with leave-one-publisher-family-out splits. Finally, we regress held-out accuracy on page count, publisher count, Vendi Score and PLDiv across many pool sizes and random seeds.
Baselines
The do-nothing reference is positional credit, which is what indexes effectively use now. It labels the first author as first author and nobody as co-first, so its co-first recall is zero by construction. It anchors the credit-distortion metric.
The rule-based baseline extracts PDF text, searches the page for phrases like "contributed equally," and resolves symbols by matching each glyph to the nearest preceding name. It is the cheapest serious competitor and is tuned on the same labeled development pages as the learned models.
The strongest existing method is LayoutLM fine-tuned as in DocBank. That paper reports macro F1 of 0.9350 for LayoutLM-Large, and 0.9488 for a Faster R-CNN plus LayoutLM ensemble, across its 12 coarse categories. These are not author-role numbers. No reported result was found on Lacuna for co-first-author extraction on any benchmark. We will run LayoutLM, and the ensemble, with the same architecture, the same compute and the same hyperparameter search for every training pool. That way the only thing that changes between runs is which pages are in the pool.
There are two further comparisons. The volume-scaled arXiv pool uses ten times the weakly labeled arXiv pages and tests whether quantity alone closes the gap. The key ablation compares PLDiv-greedy selection with publisher-stratified selection at an identical budget, which isolates whether geometric selection adds anything beyond counting sources.
We define beating the baselines as follows. On held-out publishers, the PLDiv-greedy pool must exceed the stratified pool in author-level co-first F1 by more than the seed-to-seed standard deviation, and it must also exceed the ten-times arXiv pool.
Why not just…?
Why not just use regex on footnote text? For deciding whether a paper has any equal-contribution statement at all, regex may be nearly perfect, and we expect to report that honestly. The hard part is deciding which authors the statement covers. Standard PDF text extraction separates superscripts from names, as the DocBank pipeline discussion notes, and Sefid et al. document how badly raw extracted headers are corrupted. If regex matches the learned parser at the author level on held-out publishers, the extraction motivation collapses, and we would report the diversity study as a negative result for this task.
Why not just stratify by publisher? This is the real threat. Counting publishers might capture almost all of the useful diversity. Template variation within a publisher, such as journal-specific styles, two-column versus one-column layouts and changes across eras, is the only place geometric selection can add value. The ablation decides the question directly. We will also check whether PLDiv adds explanatory power beyond publisher count in the regression. If stratification ties PLDiv selection, the practical recommendation becomes "stratify," which is still a useful finding.
Why not just scale arXiv volume? Le Scao et al. and Larrazabal et al. give evidence from other modalities that composition beats volume. However, layout is a different setting, and arXiv's LaTeX classes may already vary a lot. The ten-times arXiv pool tests this directly.
Why not just assign 1/N credit? Equal division across all authors erases the distinction between leads and contributors. The funding and promotion bodies cited in Kosta et al. and Vishnu et al. require exactly that distinction. Equal division also does not answer the generalization question this study is designed to test.
Proposed Experiments
The primary metric is author-level F1 for co-first-author identification on each held-out publisher family, with corresponding-author F1 reported as a secondary metric. We also report a credit-distortion score. For each paper it measures the absolute difference between the credit share the system assigns to the second listed co-first author and the agreed equal share. Each paper is then weighted by how many co-first authors it has.
The correction-notice set is evaluated separately. The corrected versions test the parser. The original versions measure the upstream loss rate, which should come out near total by construction and is reported as a check on the pipeline rather than as a result.
The decisive checks are three. First, across pools of matched size, PLDiv should predict held-out F1 better than publisher count and better than Vendi Score. Second, PLDiv-greedy selection should beat stratified selection on at least most held-out families. Third, the ten-times arXiv pool should not close the gap.
If the first check fails, geometric diversity measures are not shown to be useful for choosing training data in this task. If only the second fails, the practical recommendation becomes "stratify by publisher." In either case the benchmark and the measured split between parser losses and publisher losses stand as contributions.
References
- Correction to "Improving Multi-Agent Trajectory Prediction Using Traffic States on Interactive Driving Scenarios"
- A Review of Microsoft Academic Services for Science of Science Studies
- Cleaning Noisy and Heterogeneous Metadata for Record Linking across Scholarly Big Datasets
- DocBank: A Benchmark Dataset for Document Layout Analysis
- Document AI: Benchmarks, Models and Applications
- What Language Model to Train if You Have One Million GPU Hours?
- Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis
- WIDER FACE: A Face Detection Benchmark
- Measuring Dataset Diversity from a Geometric Perspective
- The authorship dilemma: alphabetical or contribution?
- Correction: Multi-source-free Domain Adaptive Object Detection
- Author Correction: Methods and open-source toolkit for analyzing and visualizing challenge results
- Author Correction: Cell segmentation-free inference of cell types from in situ transcriptomics data
- Publisher Correction: Efficiently measuring a quantum device using machine learning
- Author Correction: Contactless 12-lead electrocardiogram via deep computational radar
- Correction to "Nonlinear Regression via Deep Negative Correlation Learning"
- Erratum to "CASH-RAM: Enabling In-Memory Computations for Edge Inference Using Charge Accumulation and Sharing in Standard 8T-SRAM Arrays"