P2P: Automated Paper-to-Poster Generation and Fine-Grained Benchmark

1. Introduction

• Manual poster creation is time-consuming and skill-intensive, while existing automated methods struggle with semantic richness, structural nuance, quality control, and standardized evaluation.These limitations motivate an automated framework and benchmark for academic poster generation.Source: §1 · p. 3
• P2P uses Figure, Section, and Orchestrate Agents with dedicated checker modules for iterative academic poster generation.The agents process visual elements, generate structured content, and assemble HTML-rendered posters.Source: §1 · p. 3, Figure 1 · p. 4
• Over 30,000 instruction examples in P2PInstruct support the academic paper-to-poster generation task.The dataset is designed to advance model development for this task.Source: §1 · p. 3
• P2PEval provides 121 paper-poster pairs and dual Universal and Fine-Grained evaluations using LLM-as-a-Judge and human-annotated checklists.The benchmark targets comprehensive assessment of poster quality across multiple dimensions.Source: §1 · p. 3

2.1. P 2 P: Multi-agent for Paper-to-Poster Generation

• P2P coordinates Figure, Section, and Orchestrate Agents, each paired with a checker module for iterative refinement.The workflow covers visual processing, textual generation, and final poster assembly.Source: §2.1 · pp. 4–5
• The Figure Agent extracts figures and tables, pairs them with captions, generates descriptions, and checks duplication, coverage, and visual-caption alignment.Its confidence threshold is lowered iteratively when detected figure and caption counts disagree.Source: §2.1 · p. 4
• The Section Agent infers a poster schema and generates content from the paper, schema, and described visual elements while checking coherence, completeness, faithfulness, and references.A reflection loop revises structure or content when the checker detects inadequacies.Source: §2.1 · p. 5
• The Orchestrate Agent renders Markdown content and visuals as HTML/CSS, using modular styling, institutional color alignment, and responsive flexbox layouts.The Poster Checker triggers adjustments for spacing, alignment, and structural integrity.Source: §2.1 · p. 5
• Figure 2 maps titles, figures, and sections from a multi-page input paper to corresponding locations in the generated poster.The visualization presents the paper on the left and the P2P poster on the right.Source: §2.1 · p. 5, Figure 2 · p. 6

2.2. P 2 P I N S T R U C T: A Large-Scale Instruction Dataset for Paper-to-Poster Generation

• P2PInstruct contains 30,460 instruction-response pairs spanning the complete poster-generation workflow.The dataset includes 16,848 figure-description pairs and 13,612 textual-generation examples; responses average over 3,300 tokens for textual content generation.Source: §2.2 · p. 5

3.1. Benchmark Constructing

• Four annotators create checklist criteria covering visual elements, content consistency, structure, research emphasis, and importance-based scores from 1 to 5.Each poster is independently reviewed by three annotators, with a fourth verifying annotations.Source: §3.1 · p. 6
• P2PEval contains 121 paper-poster pairs from ACL and SciPostLayout, preserving papers and posters in PDF formats and posters additionally in PNG.The collection spans diverse research categories and supports evaluation of textual content and visual layout.Source: §3.1 · pp. 6–7
• The benchmark contains 1,738 checklist items, including 775 associated with visual elements, with an average per-item maximum score of 4.07.Its distribution spans broad research categories and disciplines.Source: §3.1 · p. 7, Figure 4 · p. 7

3.2. Poster Evaluation Framework

• P2PEval evaluates academic posters through complementary Universal and Fine-Grained Poster Evaluation pipelines.The benchmark is designed to assess overall quality alongside fidelity to specific content and visual elements.Source: §3 · p. 6, §3.2 · p. 7, §3.2.2 · p. 8

3.2.1. Universal Poster Evaluation

• Universal evaluation scores ten criteria from 0 to 5, including title accuracy, image quality, white space, content fidelity, information flow, and self-contained explanation.Three human annotators provide cumulative 0–50 ratings, producing 1,701 independent annotation scores.Source: §3.2.1 · pp. 7–8
• XGBoost models nonlinear interactions between universal criteria and human annotations, achieving R2 = 0.92 with 10-fold cross-validation and 200 trees.This exceeds Ordinary Least Squares (R2 = 0.66), Random Forest (R2 = 0.83), and regularization methods (R2 = 0.89).Source: §3.2.1 · p. 8

3.2.2. Fine-Grained Poster Evaluation

• Fine-Grained evaluation normalizes checklist scores to 0–100 using annotator-consensus maximums from 1 for minor elements to 5 for core research components.The score emphasizes fidelity to authors’ intended communication goals and enables comparisons across poster-generation methods.Source: §3.2.2 · p. 8

4.2. Evaluation Metrics

• P2PEval combines Universal and Fine-Grained human-validated scores with ROUGE, BERTScore, and a Judge preference metric after removing image links from text.Universal evaluation has a maximum of 50, while Fine-Grained evaluation has a maximum of 100.Source: §4.2 · p. 10

4.3. Results and Analysis

• Claude-3.7-Sonnet leads qualitative Universal and Fine-Grained evaluations, while Deepseek-R1 remains competitive among open-source models.Reasoning-enabled Claude and Qwen models perform especially strongly on Fine-Grained evaluation.Source: §4.3 · p. 10
• P2PInstruct fine-tuning substantially improves performance: Qwen3-P2P-8B achieves the highest ROUGE scores and outperforms its base, while Qwen2.5-VL-P2P-7B improves across all metrics.Qwen3-P2P-8B also gains in FineGrain and Universal scores over Qwen3.Source: §4.3 · p. 10
• P2P-generated posters are preferred over YuanBao outputs and competitive with original posters in pairwise human evaluations.The evaluation reports both preference-or-tie and strict-preference rates for method comparisons.Source: §4.3 · pp. 10–11, Table 2 · p. 10
• HTML outputs consistently outperform SVG and LaTeX on Fine-Grained and Universal metrics, supported by flexible CSS layout, browser rendering, and stronger LLM code-generation proficiency.The analysis associates HTML with fewer rendering errors and structural inconsistencies.Source: Table 3 · p. 10, §4.3 · p. 11
• Removing checker-based reflection moderately reduces Universal metrics, removing Figure Describer degrades performance, and removing all specialized components causes the largest Fine-Grained drop.The results support modular processing and semantic visual descriptions over direct raw-image integration.Source: §4.3 · p. 11, Table 4 · p. 11
• Claude produces more fragmented, column-heavy layouts with less balanced spatial use, whereas GPT produces more uniform and compact layouts.The comparison highlights blank-space utilization and column-height consistency as important layout dimensions.Source: §4.3 · pp. 11–12, Table 5 · p. 12

Poster Generation

• Academic poster methods decompose generation into extraction, panel inference, and layout, while template-based approaches lack flexible figure-text integration and fine-grained evaluation.Existing poster datasets also lack comprehensive quality metrics.Source: §5 · p. 12

HTML Code Generation and Multi-Agent

• HTML-generation research explores direct, structured, hierarchical, and divide-and-conquer approaches across interfaces, websites, and web design.Fine-tuning and multi-agent workflows are increasingly used for complex code-generation tasks.Source: §5 · p. 12

LLM as a Judge

• LLM-as-a-Judge research uses pairwise and single-response evaluation, with reported agreement with human assessments and growing application to multimodal outputs.Chain-of-Thought and training-free instruction-following methods aim to improve evaluator robustness.Source: §5 · pp. 12–13

6. Conclusion

• P2P transforms research papers into visually coherent and informationally faithful posters through three specialized agents and checker-enabled iterative refinement.The system renders structured posters using HTML and CSS.Source: §6 · p. 13
• P2PInstruct provides over 30,000 examples, while P2PEval contains dual evaluation methodologies for systematic multi-dimensional poster assessment.These resources support development and evaluation of paper-to-poster systems.Source: §6 · p. 13
• P2P posters approach or sometimes exceed human-created examples, particularly when reasoning-enhanced LLMs are used.The conclusion frames the work as a foundation for automated academic communication tools.Source: §6 · p. 13

A. Examples of Poster Generation

• Figures 5 and 6 show examples of generated academic posters.Source: A. Examples of Poster Generation · p. 19

• Checklist items reference specific elements in official posters and corresponding papers, aligning evaluation with the original authors’ intent.Source: B. The Features of Fine-Grained Poster… · p. 19
• The pipeline emphasizes domain-specific expectations and conventions that universal evaluation criteria may overlook.Source: B. The Features of Fine-Grained Poster… · p. 19
• Critical research components, including key figures, methodology details, and conclusions, receive explicit verification through detailed scoring mechanisms.Source: B. The Features of Fine-Grained Poster… · p. 19
• Four human annotators calibrate checklist scores, incorporating domain expertise and judgments about item significance and presentation quality.Source: B. The Features of Fine-Grained Poster… · p. 20

• HTML posters can be viewed on any device with a web browser without specialized compilation software or viewers.Source: C. The Features of HTML Format · p. 20
• HTML supports interactive elements such as hyperlinks, tooltips, expandable sections, and embedded multimedia through CSS and JavaScript.Source: C. The Features of HTML Format · p. 20
• CSS enables flexible, modern, responsive styling and makes HTML posters easy to embed in websites, share by link, and use on online conference platforms.Source: C. The Features of HTML Format · p. 20

Prompt design

• Paper analysis prompts identify sections to include and tailor emphasis to methodology, benchmark, or survey papers.Source: D. Prompt · p. 21

Figure processing

• Figure-analysis prompts request concise descriptions of academic figures, their purposes, and visible data or trends.Source: D. Prompt · p. 21

Text-based Poster Generation

• Text-based generation simplifies prose when figures convey the same content while retaining important mathematical formulas where useful.Source: D. Prompt · p. 21

Image-based Poster Generation

• Image-based generation analyzes paper content to select poster sections and inserts figures using integer indices, with each figure used once and placed accurately.Source: D. Prompt · pp. 21–22

HTML assembly

• The poster object maps section names to Markdown content and encodes image metadata including dimensions and aspect ratio.Source: D. Prompt · p. 22
• HTML output is restricted to body content, specified tags, prescribed poster containers, and images rendered with their indices and alternative text.Source: D. Prompt · p. 22

Styling

• Styling prompts require affiliation-based colors, gradient section titles, and matching title text and background colors.Source: D. Prompt · p. 22

Layout

• Layout prompts balance columns and image aspect ratios, rearrange content to equalize heights, and limit blank space to 5%.Source: D. Prompt · pp. 22–23