Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers

• Academic posters are harder than slide decks because they must compress long multimodal papers into one coherent page while interleaving text and graphics under strict spatial constraints.These requirements make VLM- or LLM-only approaches insufficient without explicit visual feedback for layout, logical flow, legibility, and aesthetics.Source: §1 · p. 1
• Paper2Poster introduces the first benchmark and metric suite for poster generation, comparing generated posters with author-designed posters across visual, textual, holistic, and comprehension criteria.PaperQuiz measures whether posters convey core paper content by using VLMs to answer questions derived from the paper.Source: §1 · pp. 1–2, Figure 1 · p. 2
• PosterAgent compresses full papers into editable posters through a Parser, Planner, and Painter–Commenter loop with visual feedback for content organization, layout, and panel refinement.The Planner builds binary-tree layouts preserving reading order and spatial balance, while the Commenter targets text overflow and alignment errors.Source: §1 · p. 2
• 87% fewer tokens: open-source PosterAgent variants outperform existing GPT-4o-based multi-agent approaches on nearly all metrics.The open-source toolbox includes models such as Qwen-2.5-VL-7B.Source: §1 · p. 2
• GPT-4o outputs appear visually appealing but contain noisy or incoherent text and achieve poor PaperQuiz performance, while Engagement is the primary aesthetic bottleneck in human-designed posters.PaperQuiz scores correlate closely with human evaluations, and more capable VLMs score higher on well-designed posters.Source: §1 · p. 2

2.1 Visual Design Automation

• Paper2Poster addresses scientific visual design from long-context, interleaved multimodal academic papers, extending beyond poster generation and complex layout automation while requiring computationally efficient processing.Diffusion methods can synthesize detailed visuals but often produce blurry, unreadable embedded text, whereas code-based workflows assemble complex outputs through tools.Source: §2.1 · p. 2

2.2 Vision-Language Agents

• Unlike slide agents that translate brief text prompts into executable code, Paper2Poster converts full-length academic papers into compact, well-structured posters and therefore requires specialized metrics and practical automation workflows.General-purpose agents can interact through reasoning, coding, APIs, or user interfaces, but still struggle with professional tasks requiring accurate, domain-specific interaction.Source: §2.2 · p. 3

3.1 Task Definition

• Paper2Poster evaluates whether a poster faithfully conveys a long scientific paper through coherent, spatially efficient text–visual layouts.The task requires hierarchical abstraction, multimodal extraction and alignment, and joint reasoning over language, visual content, and constrained layout.Source: §3.1 · p. 3

3.2 Data Curation

• Paper2Poster contains 100 paper–poster pairs from ICML, NeurIPS, and ICLR, balanced across 2022–2024 and covering 280 topics.The dataset includes 33, 33, and 34 pairs from 2022, 2023, and 2024, respectively, and 35 NeurIPS, 37 ICML, and 28 ICLR papers.Source: §3.2 · p. 3
• The benchmark compresses papers averaging 12155.7 words and 22.6 pages into author posters averaging 774.1 words and 8.7 figures.This corresponds to approximately 14.4× textual compression and 2.6× figure reduction; input papers average 22.59 figures.Source: §3.2 · p. 3, Figure 2 · p. 4

3.3 Evaluation Metrics

• The evaluation framework measures visual quality, textual coherence, VLM-as-judge quality, and PaperQuiz content conveyance.Visual quality uses Visual Similarity and Figure Relevance; textual coherence uses poster-text perplexity under Llama-2-7b-hf; holistic assessment scores six aesthetic and informational criteria from 1–5.Source: §3.3 · pp. 4–5, Figure 3 · p. 4
• PaperQuiz generates 100 paper-specific multiple-choice questions and has six VLMs answer them from each poster to measure communicated knowledge.Each paper receives 50 verbatim and 50 interpretive questions, while the protocol compares poster variants across reader standards ranging from casual to expert.Source: Figure 3 · p. 4, §3.3 · p. 5
• PaperQuiz applies a human-poster-length penalty that discourages verbosity, avoids harsh punishment, and caps the brevity bonus at twice the raw score.The adjustment uses poster text length L relative to the median human-designed poster length W, with raw scores sr in [0, 100] and adjusted scores sa in [0, 200].Source: §3.3 · p. 5

4 PosterAgent

• PosterAgent uses a top-down multi-agent pipeline combining global document restructuring with local panel-level refinement through Parser, Planner, and Painter–Commenter components.The Parser extracts text and visual content, the Planner arranges coherent layouts, and Painter–Commenter renders panels while checking layout coherence and overflow.Source: §4 · p. 6

4.1 Parser: global organization

• The Parser converts PDFs into a compact multimodal asset library containing section-level text synopses and caption-keyed extracted figures or tables.MARKER[21] and DOCLING[14] produce Markdown that an LLM transforms into a structured JSON-like outline, preserving essential semantics while reducing size.Source: §4.1 · p. 6

4.2 Planner: local organization

• The Planner selects content iteratively by matching visual assets to relevant sections and completing the poster section by section rather than in one shot.This process mirrors template-based human poster design and produces section–figure pairs for subsequent panel construction.Source: §4.2 · p. 6
• The Planner uses a binary-tree layout strategy [30] to estimate panel bounding boxes from content length and figure size while preserving reading order and aspect ratio.This avoids unstable direct numerical coordinate prediction and ensures each poster section maps to a defined panel.Source: §4.2 · p. 6

4.3 Painter–Commenter: local refinement

• The Painter converts each section–figure pair into concise bullet points and executable python-pptx rendering code for a draft panel image.An LLM performs synopsis distillation, while a deterministic code generator invokes predefined helper functions and executes the presentation code.Source: §4.3 · p. 6
• The Painter–Commenter loop uses zoomed panel views and in-context overflow and ideal-layout references to guide VLM feedback until success or an iteration limit.Feedback such as “overflow,” “too blank,” or “good to go” drives revisions intended to improve accuracy, readability, and visual balance.Source: §4.3 · pp. 6–7

5.1 Baselines and Settings

• The evaluation compares oracle, end-to-end, multi-agent, and PosterAgent baselines, including GPT-4o and fully open-source Qwen variants.Oracle methods use the original paper and author-designed poster as upper bounds, while OWL and PPTAgent represent specialized multi-agent workflows.Source: §5.1 · p. 7

5.2 Main Results

• PosterAgent-4o reaches a 3.72 average VLM-as-Judge score, while PosterAgent ranks second in visual similarity and highest in figure relevance against PPTAgent.GPT-4o commenters outperform Qwen2.5-VL-7B commenters during panel refinement, especially for preventing text overflow; engagement remains the primary bottleneck, with all variants below 3.Source: §5.2 · p. 7, Table 1 · p. 8
• PosterAgent variants consistently achieve the best PaperQuiz scores, with PosterAgent-Qwen surpassing resource-intensive baselines through structured compression and logically ordered layouts.PaperQuiz rewards concise, well-designed posters rather than sheer content volume; PosterAgent-Qwen outperforms 4o-HTML and OWL-4o despite weaker reader-model backbones and no brevity penalty for those baselines.Source: §5.2 · p. 8, Table 2 · p. 9

5.3 Qualitative Analysis

• PosterAgent produces structurally coherent, readable posters with the highest qualitative scores, unlike pixel-generated, blog-like, or template-filled baselines that impair readability or omit panels.GPT-4o pixel generation renders fine text poorly, 4o-HTML and OWL are text-dense, and PPTAgent often produces missing panels; PosterAgent still trails human-designed versions.Source: §5.3 · p. 9

5.4 Efficiency Analysis

• PosterAgent uses 101.1K 4o-based or 47.6K Qwen-based tokens, reducing cost by 60%–87% versus OWL-4o and costing $0.55 or $0.0045 per poster.Parallel panel generation further reduces runtime by 40.7%, making PosterAgent-4o-Parallel more competitive in speed.Source: Figure 7 · p. 9, §5.4 · p. 9

6 Conclusions

• Paper2Poster is a new benchmark for generating posters from academic papers.Source: §6 · p. 9
• The benchmark highlights challenges and limitations of current generative models and agents on long-context, layout-sensitive tasks.Source: §6 · p. 9
• Figure 8 compares author-designed, PosterAgent, multi-agent, pixel-generative, and website-generative poster variants using PaperQuiz’s augmented score.Source: Figure 8 · p. 10
• PosterAgent combines structured parsing, hierarchical planning, and visual feedback to enhance poster-generation quality.Source: §6 · p. 10
• PosterAgent narrows the performance gap with human-designed posters.Source: §6 · p. 10
• PosterAgent establishes an efficiency standard for scientific communication.Source: §6 · p. 10
• PosterAgent provides a practical and scalable approach to scientific communication.Source: §6 · p. 10

A Limitations and Future Work

• 4.5 minutes per document results from serially executing structurally independent panel refinement cycles, limiting batch-generation and interactive scalability.Panel-level parallelism could substantially reduce runtime while preserving modularity and output quality.Source: A Limitations and Future Work · p. 15
• Panel-level parallelism is proposed to reduce end-to-end poster-generation time and improve scalability for batch and real-time workflows.Source: A Limitations and Future Work · p. 15
• Future workflows should incorporate external knowledge, including community feedback and institutional or conference assets, beyond the source paper.Examples include OpenReview comments, social-media reactions, institutional icons, and conference logos.Source: A Limitations and Future Work · p. 15
• A human–AI workflow could have the agent draft a poster, solicit user feedback, and iteratively refine the output to meet requirements.Source: A Limitations and Future Work · p. 15

B Example Visualization

• The Paper2Poster dataset contains 100 paper–poster pairs, and examples demonstrate PosterAgent’s multimodal extraction, summarization, and single-page layout generation.The examples cover reinforcement learning, anomaly detection, and neuroscience, including varied layouts, compression ratios, and figure–text integration.Source: B Example Visualization · p. 15
• Figures 9–11 present posters for reinforcement learning, industrial anomaly classification, and brain-data pretraining applications.Source: Figure 9 · p. 15, Figure 10 · p. 15, Figure 11 · p. 16
• Figures 12–13 present posters for conformal semantic keypoint detection and neural tangent kernels for axis-aligned tree ensembles.Source: Figure 12 · p. 16, Figure 13 · p. 16

C Ablation Study

• The ablation study isolates binary-tree layout planning, the commenter module, and in-context examples while holding other PosterAgent-4o components fixed across five randomly selected papers.The evaluated variants are Direct, Tree, Tree + Commenter, and the full Tree + Commenter + IC system.Source: C Ablation Study · pp. 16, 18
• Tree+Commenter+IC achieves the best ablation results across five papers, outperforming direct layout generation and variants lacking the commenter or in-context examples.Direct generation causes structural defects, missing text boxes, poor hierarchy, and incorrect figure aspect ratios; Tree produces severe text overflow, while Tree+Commenter still leaves overflow and unused whitespace.Source: C Ablation Study · p. 18

D Abbreviations

• Table 3 provides the full names of the abbreviations used throughout the paper.The abbreviation reference supports interpretation of the model and system comparisons in the surrounding analyses.Source: D Abbreviations · p. 23, Table 3 · p. 23

E.1 Efficiency Analysis

• PosterAgent-Qwen-2.5-7B uses 47.6K tokens and 192.0s, compared with PPTAgent’s 255.7K tokens and 230.7s, while maintaining output quality.Although 4o-Image is most efficient by avoiding multi-turn reasoning, it lacks layout awareness.Source: §E.1 · p. 23

E.1.1 Runtime Analysis and Parallelization

• Parallelized content generation reduces generation time from 176.69s to 54.16s, a 69.3% reduction, lowering total runtime to 166.80s, 40.7% faster than sequential and near OWL-4o’s 158.97s.The implementation maintains superior output quality across all metrics.Source: §E.1.1 · p. 23, Table 5 · p. 24

E.2 Cost Analysis

• The cost analysis compares GPT-4o and Qwen-2.5 backbones using their respective API prices and reports text tokens, vision tokens, and computation times.GPT-4o prices are based on OpenAI rates from May 22, 2025, while Qwen-2.5-7B and Qwen-2.5-VL-7B prices use OpenRouter rates from May 26, 2025.Source: §E.2 · p. 23, Table 4 · p. 24

E.3 Impact of Backbone Choices

• Backbone combinations are largely robust: PaperQuiz ranges from 114.09 to 118.25, while visual similarity and figure relevance vary by less than 0.01.The fully open-source Qwen–Qwen stack scores 114.65 on PaperQuiz, only 3.6 below the best closed-source variant.Source: §E.3 · p. 24
• VLM substitution improves PaperQuiz by +2.1 for 4o-Qwen over 4o-4o and +0.56 for Qwen-Qwen over Qwen-4o, whereas 4o-4o receives higher VLM-as-judge ratings of 3.72 versus 3.58.The paper attributes this trade-off to GPT-4o’s stricter layout criticism and the VLM’s influence on aesthetic scoring; 4o-4o offers the best balance.Source: §E.3 · p. 24

E.4 Additional Backbone Evaluations

• PosterAgent variants score 3.56–3.72 on overall VLM-as-Judge versus 2.33 for 4o-Image, a 1.23–1.39 absolute improvement, while Table 7 reports the complete comparison.All variants substantially outperform the baseline across nearly all metrics.Source: §E.4 · p. 24, Table 7 · p. 25
• Information scores cluster at 3.83–3.87 versus 1.77 for the baseline, while Aesthetic scores reach 3.26–3.58 versus 2.90, indicating backbone-insensitive improvements.Source: §E.4 · p. 25
• The results confirm that PosterAgent’s multi-agent design generalizes across reasoning closed-source models and larger open-source alternatives.The evaluation covers o4-mini and Qwen-2.5-72B alongside the original 4o-Image baseline.Source: §E.4 · pp. 24–25

E.5 Poster Generation Paradigm Comparison

• PosterAgent combines code-generated coordinates, sizes, and layering with PPTX rendering and visual feedback to iteratively refine layout and content.This hybrid paradigm is designed to satisfy both content accuracy and visual layout quality.Source: §E.5 · p. 25
• Coding-only methods provide precise, reproducible placement but lack natural rendered-result inspection and produce artifacts that are cumbersome to edit.They therefore struggle to correct issues such as text overflow and alignment problems.Source: §E.5 · p. 25
• GUI-only pipelines ease editing and rendered feedback but require inefficient low-level operations for precise large-scale automated adjustments.Typical operations include clicking and dragging.Source: §E.5 · p. 25
• Template retrieval offers editable outputs but depends on scarce suitable templates and is not true from-scratch generation; PPTAgent-4o still underperformed PosterAgent despite six manually selected human templates.The comparison underscores template dependence as a limitation for scientific poster generation.Source: §E.5 · p. 25

E.4 Additional Backbone Evaluations

• PosterAgent-o4-mini achieves the highest raw PaperQuiz scores, reaching 61.33 overall and 121.91 augmented, while all PosterAgent variants substantially outperform the 4o-Image baseline across nearly all metrics.Source: §E.4 · p. 24, Table 8 · p. 26

E.6 VLM-as-Judge Robustness Analysis

• std < 0.024 across all metrics, while the overall average has std = 0.005 across five complete evaluations on 100 samples.Table 11 covers six fine-grained criteria: three aesthetic and three information dimensions, with extremely narrow 95% confidence intervals.Source: §E.6 · p. 26
• Coefficient of variation is < 1% for most measures, indicating high consistency across evaluation runs.All reported metrics exhibit high consistency across the independent runs.Source: §E.6 · p. 26
• Narrow confidence intervals indicate that a single evaluation run provides reliable estimates for model comparison.The stability results support using one run for practical comparisons while retaining confidence in the estimates.Source: §E.6 · p. 26
• The observed consistency validates the VLM-as-Judge approach as a robust automatic evaluation method for poster generation.This conclusion follows from the stable results across all six fine-grained criteria and the averaged scores.Source: §E.6 · p. 26
• Single-run evaluations are sufficient for practical model comparison, with periodic multi-run audits recommended to verify continued metric stability.The recommendation is based on the observed stability of the evaluation metrics.Source: §E.6 · p. 26

F Detailed Definition of Evaluation Metrics

• The study defines four types of evaluation metrics used to assess poster generation.This section introduces the detailed definitions of all four metric types applied in the study, without specifying them in the supplied passage.Source: F Detailed Definition of Evaluation… · p. 26

F.1 Visual Quality Metrics

• Visual quality uses CLIP-based Visual Similarity for instance-level semantic alignment and Figure Relevance for contextual appropriateness of poster figures.Visual Similarity compares generated and ground-truth poster image embeddings, while Figure Relevance compares figure-crop and corresponding-section-text embeddings.Source: §F.1 · pp. 26–27

F.2 Textual Coherence Metrics

• Textual coherence is measured with full-sequence perplexity under Llama-2-7b-hf, where lower PPL indicates more predictable and coherent poster text.The poster is tokenized as w1:n, and perplexity uses each token’s conditional probability p(wi | w<i).Source: §F.2 · p. 27

F.4 PaperQuiz

• PaperQuiz is curated by converting each paper PDF to markdown and prompting o3 to generate 100 multiple-choice questions per paper.The set contains 50 verbatim questions spanning 13 content aspects and 50 interpretive questions spanning 10 conceptual dimensions.Source: §F.4 · p. 33
• The evaluation generates exactly 50 factual questions answerable from poster text and exactly 50 high-level questions addressing purpose, novelty, approach, findings, implications, limitations, and conclusions.Source: §F.4 · pp. 33–34
• PaperQuiz queries six open- and closed-source VLMs with poster-grounded questions, requiring answers to rely only on visible poster content and cite supporting regions or return NA.The readers comprise three open-source models and three closed-source models; unsupported answers must provide NA for both answer and reference.Source: §F.4 · p. 35
• The scoring uses raw question-answering accuracy together with a density multiplier based on poster-text length relative to the median ground-truth poster length.The density multiplier is capped at 2, penalizing verbosity while rewarding concise, information-dense designs.Source: §F.4 · p. 36

Instructions

• Evaluators review each poster alongside six criterion files and assign scores according to those criteria.The criteria match those outlined in PaperQuiz F.4.Source: G Human Evaluation Protocol · p. 36

H Error Analysis

• Across five pipelines, the main recurring errors are corrupted text, visual or layout flaws, missing visuals, and content overflow.The evaluated systems are 4o-Image, 4o-HTML, OWL-4o, PPTAgent, and PosterAgent.Source: H Error Analysis · p. 37

H.1 Text Integrity Issues

• 4o-Image often produces garbled text, while PPTAgent can leave placeholders intact or partly overwrite content.These failures arise from fragile pixel-level typography in image-only generation and semantic corruption in template-based generation.Source: §H.1 · p. 37

H.2 Visual / Layout Flaws

• 4o-Image can truncate content or hallucinate figures, PPTAgent can make figures unusably small or leave blank regions, and OWL-4o can create empty areas without visual validation.The failures reflect missing hard spatial constraints, partial generation, and sequential code without iterative visual checking.Source: §H.2 · p. 37

H.3 Missing Visuals

• OWL-4o may fail to complete figure retrieval and insertion, while 4o-HTML is text-only by design, leaving their posters without figures.OWL-4o can invoke external extraction toolkits in principle but does not complete the full retrieval-insert cycle.Source: §H.3 · p. 37

H.4 Overflow Issues

• PPTAgent and PosterAgent sometimes let text spill beyond PPTX panel boundaries, with overflow more pronounced for PosterAgent’s Qwen variant.PPTX lacks HTML-like parent–child containment, and Qwen2.5-VL-7b provides weaker visual grounding than GPT-4o, reducing feedback-loop reliability.Source: §H.4 · p. 37

I.1 Baseline Prompts

• Baseline prompts instruct 4o-Image to generate a poster image, PPTAgent to execute python-pptx, and HTML agents to output a rendered HTML poster.The prompts require professional conference posters with summaries, relevant figures or tables, aligned layouts, and specified dimensions.Source: §I.1 · pp. 41–42

I.2 Parser Prompts

• Parser prompts reorganize papers into concise two-level JSON sections, preserve raw content, create poster metadata, and filter images or tables by relevance.The filtering output contains only image_information and table_information, with at most five entries per category.Source: §I.2 · pp. 42–44

J Planner Prompts

• Planner prompts assign at most one non-reusable image or table to each matching top-level section and generate structured bullet textboxes for poster content.Assignments require a content-fit reason, while text generation supports one or two textboxes with aligned bullet-item structures.Source: J Planner Prompts · pp. 44–46
• The planner prompts define structured textboxes and guide the commenter to classify target panels as overflow, too blank, or well-fitted using visual examples.The commenter returns “1” for overflow, “2” for excessive blank space, and “3” when text fits well within the bounding box.Source: J Planner Prompts · pp. 48–49

K Failure by Diffusion Models

• Diffusion models generate blurry or incoherent text, unreliable layouts, and uncontrollable styling, making them inadequate for academic poster generation.These failure cases are illustrated for Stable Diffusion Ultra in Fig. 28.Source: K Failure by Diffusion Models · p. 49, Figure 28 · p. 50

L Illustration of In-context reference for Commenter

• The commenter uses red-boxed in-context references to help the VLM identify panel failures such as overflow and excessive blank space.Fig. 29 illustrates these references during panel refinement.Source: L Illustration of In-context reference… · p. 49, Figure 29 · p. 50