acceptodds
Withdrawn from ICLR 2027

RAP: A Dataset for Paper-to-Poster Generation via Declarative Markup from Human-Crafted Exemplars

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Making a conference poster is visual coding in miniature: an entire paper must be compressed onto a single page, which exercises summarization, visual design, and code implementation at once, using nothing beyond the paper itself. Yet benchmarks for this task in the context of visual coding remain scarce: existing resources are small and annotated only with layout boxes, so evaluation must fall back on VLM-as-judge or human checklists rather than objective render-and-compare, and the generation methods they support design layouts from scratch, leaving human-crafted posters unusable as either a retrievable design library or as verified supervision. We introduce RAP, a dataset and benchmark for paper-to-poster generation as a visual coding task. RAP contains 25K+ verified paper-poster pairs from various Machine Learning conferences published between 2022 and 2026, annotated with a three-level content taxonomy, plus a held-out benchmark of 150 test papers each paired with its human poster. A fully automated derendering pipeline turns each poster into styled HTML/CSS of typed, refillable slots, so every sample is a paper-poster-markup triplet verifiable by rendering against the source poster. The corpus serves both inference and training: PosterRAG, a training-free multi-agent pipeline, retrieves an exemplar and refills its slots without touching layout or styling, while an SFT+RL recipe trains a dedicated model from the same triplets.

Primary area:
datasets and benchmarks
Keywords:
Paper-Poster Dataset, Poster Derendering, Academic Poster Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SsL1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper constructs RAP, a fairly large dataset for paper-to-poster generation. It includes an automated data construction pipeline that converts posters into HTML/CSS with fine-grained layout, style, and content-slot information. Based on these data, the authors explore two applications: first, PosterRAG, which retrieves an existing template and fills it with content from a new paper; second, SFT and RL on Qwen3.8-27B, using factors that affect poster quality, such as layout, content completeness, and readability, as training rewards. The experiments evaluate reconstruction quality during data construction, as well as the visual diversity of the posters generated by the two approaches and their distributional similarity to human-crafted posters.

Critical strengths

1.The data construction pipeline is quite interesting and provides a clear process that can scale up. In terms of data processing, some of the fine-grained representations are helpful for subsequent training. 2.Using the factors that determine poster quality as a rubric for training is a good idea and a good design choice.

Critical weaknesses

1.The paper does not clearly describe manual spot-checking in the data construction pipeline. 2.I noticed that the authors state that template information from the corresponding original poster is provided at test time. However, in the comparison in Figure 6, I could not clearly see the reference template’s style reflected in the outputs from Astra and Gemini. 3.Table 3 does not specify the parameter count of Qwen3.5. 4.There is a lot of blank space on page 21. I suggest addressing this.

Questions for authors

Ask only questions whose answers could materially affect your assessment. Distinguish clarification requests from suggestions for future work. Was any manual spot-checking carried out during data construction? If so, how many samples were checked, what was checked, and what were the results? In Figure 6, Astra and Gemini do not seem to follow the reference template’s style, even though the template information is provided. Does the prompt explicitly ask them to follow the template? Normally, when we ask AI to help us make a poster and provide a template, we would also ask it to follow that template. If this requirement is made more explicit in the prompt, would the gap still be there? Which size of Qwen3.5 was used in Table 3? Maybe the largest one?

Reviewer_SkVDOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces RAP, a 25.4k paper–poster dataset derendered into HTML and CSS, and demonstrates that both a template-retrieval pipeline (PosterRAG) and an SFT+GRPO post-trained model achieve state-of-the-art poster diversity and human-likeness on a 150-paper benchmark.

Critical strengths
  1. A massive and genuinely editable dataset. With over 25k pairs converted into structured HTML or CSS with typed slots, RAP significantly scales up existing resources (e.g., Paper2Poster's 100 pairs) while providing truly renderable and editable markup. The derendering pipeline is technically sound, combining layout detection with OCR-derived font constraints to achieve reliable reconstruction. Beyond the specific experiments in this paper, the corpus itself represents a valuable standalone artifact for layout generation and document understanding.

  2. Practical workflows with smart engineering safeguards. Both applications introduce sensible constraints against typical failure modes in automated design. In PosterRAG, programmatic slot-filling with CSS guardrails effectively prevents the destructive layout corruption common to unconstrained regeneration loops. Similarly, the post-training recipe visibly resolves backbone collapse, and demonstrating the utility of the derendered supervision.

Critical weaknesses
  1. The central premise of reference-based evaluation is missing. The abstract, introduction, and Section 2 heavily pitch "objective render-and-compare" against paired human references as an alternative to subjective LLM judges. Yet this evaluation is entirely absent from the generation benchmark. Tables 2 and 7 report only pairwise diversity and dataset-level FID/FDiD against RAP-25K. The 150 withheld ground-truth posters are never compared against the generations. Actual render-and-compare appears only in Table 3's derendering ablation, which evaluates layout reconstruction from an existing image, not paper-to-poster generation.

  2. The evaluation metrics are entirely paper-agnostic. Neither internal diversity nor dataset FID or FDiD measures alignment with the input paper. A degenerate baseline that ignores the paper and samples random posters from RAP-25K would achieve near-optimal FID or FDiD and high diversity. Similarly, returning the retrieved template unmodified would yield scores comparable to PosterRAG. Without evaluating these trivial baselines or reporting the ground-truth human posters as an upper bound, Table 2 cannot prove that any system generates a faithful poster for a given paper. Consequently, the HTML-vs-image ablation in Section 5.4 primarily tracks how much template markup is preserved rather than synthesis quality.

  3. Circularity in the reference distribution. PosterRAG explicitly preserves the retrieved template's layout, CSS, palette, typography, and decorative assets. Because the reference distribution for FID and FDiD is RAP-25K, which is the exact repository from which these templates are drawn, achieving low distributional distance is guaranteed by construction.

  4. Zero measurement of content fidelity, legibility, or layout integrity. While criticizing prior benchmarks (P2PEval, Paper2Poster) as subjective, the paper replaces their verifiable checks with no content evaluation at all. There is no human study, no text-fidelity or factuality metric, and no reporting of text clipping or overflow rates.

  5. Complete lack of failure analysis. The qualitative evaluation is limited to cherry-picked successes (one example per table). A credible systems paper needs transparent failure cases: what happens when LayoutLMv3 missegments dense blocks, when template retrieval selects an incompatible narrative layout, or when the post-trained model outputs overlapping DOM elements?

Questions for authors
  1. What do the following score on the five Table 2 metrics: (a) a random RAP-25K poster per test paper, (b) the retrieved template with no filling, (c) the 150 human ground-truth posters themselves?

  2. Of the 25,435 posters, how many derendered successfully, under what acceptance criterion, and what is the distribution of per-poster SSIM and mIoU?

Reviewer_pStDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors performed derendering on 25K posters and decomposed them into HTML code and assets such as figures and charts, thereby preparing the data so that poster generation could be treated as visual code generation. The experimental results confirmed that using human-created poster templates allows for the generation of more diverse layout and design than generating them from images, and that training using the collected data improves performance compared to the baseline.

Critical strengths
  • By annotating the 25K poster dataset with diverse information—including boxes, OCR, HTML, figure assets, design captions, and taxonomy—this dataset contains richer information than other datasets.
  • They conducted evaluations using two methods: agent-based evaluation within a retrieval-augmented generation framework and a method involving the custom fine-tuning of a VLM.
Critical weaknesses
  • Regarding quantitative evaluation, they only assessed metrics such as diversity and human-likeness; there is a lack of quantitative evaluation regarding how well the model generated posters based on actual inputs. For example, quantitatively evaluating factors such as content alignment and layout quality relative to the input would allow for a more concrete assessment of the quality of the generated posters. Additionally, incorporating human evaluation at this stage would be beneficial.
  • There were some unclear points regarding the accuracy of the dataset construction (details are provided in the “Questions For Authors” section).
Questions for authors
  • Regarding the accuracy of the poster rendering process itself during dataset construction, while an automatic evaluation is performed in the ablation study in Section 5.4, do these evaluation results indicate that the data has been reproduced with sufficient accuracy? For example, I assume there is a possibility of OCR errors occurring—has the accuracy of the OCR been verified separately? Also, regarding layout detection, I assume errors are present—are these limited to minor ones? I’d also like to know if there were any typical errors encountered during automatic derendering and whether a quality filter was applied. Additionally, regarding the test set, please let me know if errors were manually checked and corrected.
  • Tables could be represented directly as text-like content using HTML structure rather than being treated as images—is there a specific reason they were treated as images in this study?
  • The information in the paper and the poster does not correspond on a one-to-one basis, and I imagine there are cases where figures and tables exist only in the poster. I’d like to know how common such cases were. Also, I understand that figures and tables found only in the poster cannot be generated using this method—is that correct?
  • When fine-tuning a VLM, how are images handled? Is the process such that IDs and other identifiers are first generated as strings, and then replaced with links to the actual images corresponding to those IDs?
  • In Sections 5.3.1 and 5.4, the analysis showed that HTML guidance resulted in greater diversity compared to image guidance, but I didn’t quite understand the reason for this. Is the difference between HTML guidance and image guidance simply whether Stage 1 template retrieval fetches HTML or an image? If it fetches an image, does that mean it tends to retrieve similar ones? Could you explain in a bit more detail why this difference in diversity occurs?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.