PosterForest: Hierarchical Multi-Agent Collaboration for Scientific Poster Generation

• Scientific posters make complex technical information more concise and visually intuitive, while manual poster creation demands substantial time, domain knowledge, and design expertise.Source: §1 · p. 1
• Earlier systems arranged extracted text and figures with heuristics, while newer multi-agent pipelines such as P2P (Sun et al., 2025) and Paper2Poster (Pang et al., 2025) decompose generation but typically require training.Examples of earlier systems include PGM (Qiang et al., 2016, 2019), NCE (Xu and Wan, 2021), and PostDoc (Jaisankar et al., 2024).Source: §1 · pp. 1–2
• Existing methods lack hierarchical document understanding, separate content from layout, and may misplace or truncate content, disrupt logical flow, oversimplify posters, and incur training overhead.Figure 1 illustrates an experiment table placed in a conclusion and a poster reduced to summarized paragraphs in fixed panels.Source: Figure 1 · p. 1, §1 · p. 2
• The Poster Tree prunes and merges content across section, subsection, and paragraph levels, links text to figures and tables, and jointly encodes content and layout attributes.It is designed to preserve salient information, reduce redundancy, maintain logical flow, and strengthen text–visual associations.Source: §1 · p. 2
• Specialized agents iteratively coordinate and refine the Poster Tree across content summarization, layout planning, and visual-material placement to jointly optimize content and structure.The pipeline uses standard APIs and publicly available checkpoints without training.Source: §1 · p. 2
• 59.2% human-study preference versus 27.2% for prior work, with PosterForest also outperforming prior approaches in automatic evaluations while remaining training-free.Source: §1 · p. 2

Scientific Poster Generation

• Prior systems use layout-driven summarization, content extraction, or modular agents, yet mapping sections to panels often loses semantic flow and hierarchy (Qiang et al., 2016, 2019; Xu and Wan, 2021; Sun et al., 2025; Pang et al., 2025).Source: §2 · pp. 2–3

Hierarchical Document Parsing and Understanding

• Hierarchical parsers, multimodal models, graph representations, and OWL's multi-LLM reasoning support document segmentation and salient-content extraction (Rausch et al., 2021; Zhang et al., 2024a; Huang et al., 2022; Gemelli et al., 2022; Hu et al., 2025).Source: §2 · p. 3

Multi-Agent Reasoning and Collaboration

• Multi-agent systems extend single-path reasoning through specialized roles, iterative feedback, critique, and coordination, including poster agents for analysis, summarization, and layout (Li et al., 2023; Zhang et al., 2024b; Sun et al., 2025; Pang et al., 2025).Source: §2 · p. 3

3.1 Preliminaries

• Existing approaches underuse document structure and text-visual links; PosterForest addresses this with a training-free framework that builds and collaboratively refines a hierarchical Poster Tree.The framework has two stages: constructing the tree, then iteratively refining it with specialized content and layout agents.Source: §3.1 · p. 4

3.2 Hierarchical Paper-to-Poster

• The Poster Tree transforms document hierarchy into concise content, derives spatial regions from that content, and unifies semantic and layout attributes in each poster node.The raw tree represents elements such as sections, paragraphs, figures, and tables; refinement prunes, merges, and summarizes content, while figures and tables remain linked to the text that references them.Source: §3.2 · pp. 4–5

3.3 Hierarchical Modification Planning

• PosterForest jointly optimizes content quality and layout organization by traversing Poster Tree nodes with local attributes, hierarchical context, and global evaluation.Unlike prior methods that fix layout before adjusting content, its updates can address both dimensions.Source: §3.3 · p. 5

3.3.1 Poster Tree Traversal (Node-level)

• In top-down node traversal, the Layout Agent adjusts geometry using descendant statistics, while the Content Agent refines textual density and abstraction against parent layout and descendant context.Parent-level changes propagate through the tree, producing coherent updates across spatial and semantic dimensions.Source: §3.3.1 · p. 5

3.3.2 Iterative Tree Refinement (Tree-level)

• Each tree-level iteration follows a complete node-update pass, and global MLLM feedback can trigger another traversal; the method uses a maximum of K = 2 iterations in practice.The feedback evaluates visual organization, textual structure, and hierarchical balance; refinement stops when another pass is not indicated.Source: §3.3.2 · pp. 5–6

4.1 Experimental Setup

• The evaluation compared oracle, end-to-end GPT-4o, multi-agent workflow, and poster-specialized baselines, using 100 benchmark pairs quantitatively and 15 additional pairs for qualitative results and user studies.Poster colors and fonts were standardized to prevent visual factors from influencing qualitative evaluations and user studies.Source: §4.1 · p. 6

4.2 Qualitative Evaluation

• Across diverse AI fields, PosterForest used space more efficiently and retained both the result table and graph omitted by P2P and Paper2Poster for the VAR paper.The comparisons used GPT-4o across methods and included examples from 3D vision, language processing, and reinforcement learning.Source: §4.2 · p. 6, Figure 4 · p. 7

4.3 Quantitative Evaluation

• MLLM judges rated PosterForest comparable to baselines in aesthetics and superior in informativeness, with scores closest to author-created ground-truth posters.Six GPT-4o judges scored six criteria from 1 to 5; the authors note that MLLM evaluation may miss subjective preferences, motivating a complementary user study.Source: §4.3 · pp. 6–7, Table 1 · p. 7

4.4 User Study

• Participants strongly preferred PosterForest over the other state-of-the-art baselines across content fidelity, aesthetic quality, structural clarity, and overall quality.The study recruited 25 AI graduate students and used 10 sets totaling 40 questions.Source: §4.4 · pp. 7–8, Table 2 · p. 8

4.5.1 Effect of Hierarchical Content Tree

• Hierarchical parsing preserved section relationships and grouped related content, improving poster readability and spatial cohesion compared with omitting hierarchy.Without hierarchy, sections and subsections were often disordered or mixed, weakening semantic grouping and the spatial coherence of figures and explanatory text.Source: §4.5.1 · p. 8, Figure 6 · p. 8

4.5.2 Effect of Content and Layout Agents

• Using both agents addressed content redundancy and layout imbalance, producing organized posters with appropriate information density and visual harmony.The Content Agent pruned redundancy and resized text but could leave unbalanced panels; the Layout Agent improved spatial organization but could cause poor figure scaling and text overflow.Source: §4.5.2 · pp. 8–9

5 Conclusion

• PosterForest substantially outperforms prior methods in informativeness, clarity, and structural quality, according to empirical results and user studies.Source: §5 · p. 9
• The Poster Tree unifies semantic and spatial attributes, enabling integrated reasoning about document structure and layout design.Source: §5 · p. 9
• Hierarchical refinement by Content and Layout Agents dynamically balances information density and visual harmony without training or dataset-specific tuning.Source: §5 · p. 9

• Generated posters may have suboptimal content density, leading to less efficient space utilization.Source: Limitations · p. 9
• The lack of robust quality metrics may limit comprehensive quantitative evaluation, motivating advanced evaluation methodologies in future research.Source: Limitations · p. 9

Ethical Considerations

• The framework uses officially released models and APIs without fine-tuning, operates under provider licenses, and calls GPT-4o through OpenAI’s official interface alongside public Qwen checkpoints.The work is training-free.Source: Ethical Considerations · p. 9
• Source papers and posters are analyzed solely for non-commercial research under academic fair-use practices, with original-source references included to respect creator rights.Source: Ethical Considerations · p. 9
• Human evaluation involved 25 graduate-level participants with prior conference experience, assessed posters only, and collected or analyzed no personal attributes; an AI assistant helped refine sentence-level drafts.Source: Ethical Considerations · p. 9

A Additional Qualitative Results

• Additional scientific poster-generation results are presented in Figures A1 and A2.The supplied passage identifies the figures but does not describe their results.Source: A Additional Qualitative Results · p. 12

B.1 Qualitative Experiments Setup

• Posters were standardized at 48 × 36 inches with unified color schemes and font styles to focus qualitative evaluation on content and layout quality.These controls were applied across methods to eliminate visual biases related to design choices.Source: §B.1 · p. 12

B.2 Quantitative Experiments Setup

• Paper2Poster contains 100 full-length papers and corresponding author-designed posters from NeurIPS, ICML, and ICLR (2022–2024).It is described as the first large-scale dataset for scientific poster generation.Source: §B.2 · p. 12
• Papers average 22.6 pages, 12,156 words, and 22.6 figures, while posters average 774 words and 8.7 figures, compressing textual and visual content.Source: §B.2 · p. 12
• To avoid leakage, benchmark papers were drawn only from the POSTERSUM test split (Saxena et al., 2025), filtered for camera-ready versions and broad year and venue coverage.The benchmark is intended to evaluate both informativeness and layout quality.Source: §B.2 · p. 12

B.3 Implementation Details

• Experiments ran on Ubuntu 22.04 LTS with an AMD EPYC 7543 CPU and eight NVIDIA RTX A6000 GPUs, using PyTorch 2.6.0.Source: §B.3 · p. 12
• Four GPUs were assigned to each of Qwen-2.5-7B-Instruct and Qwen2.5-VL-7B-Instruct, and Python, NumPy, and PyTorch random seeds were fixed for reproducibility.Source: §B.3 · p. 12

B.4 Hyperparameters

• The maximum tree-refinement count was Tmax = 2 because pilot studies found no significant gains from additional agent exchanges, balancing quality and computational efficiency.Source: §B.4 · p. 12

B.5 Detailed Baseline Setup

• P2P figure extraction used DocLayout-YOLO (Zhao et al., 2024), and the Paper2Poster Poster-Agent was implemented with GPT-4o or open-source Qwen-2.5-7B and Qwen-2.5-VL-7B models.Source: §B.5 · p. 12

C.1 Reason for Metric Selection

• Human evaluation used content fidelity, visual harmony, structural effectiveness, and overall completeness to assess scientific poster generation.These criteria address accurate summarization, attractive and logical presentation, and professional appearance.Source: §C.1 · p. 12
• Human observers may detect discrepancies between ground-truth and generated posters better than MLLM-as-a-judge metrics such as GPT-4o.The paper expects MLLM-based metrics to become more reliable as multimodal evaluation advances.Source: §C.1 · p. 12

C.2 User Study Details

• The study recruited 25 graduate students with experience in academic conferences.Each participant evaluated 10 pairs of papers and posters generated by GPT-4o HTML, P2P, Paper2Poster, and PosterForest.Source: §C.2 · p. 12
• Posters from the four methods were standardized for font and color before evaluation.Source: §C.2 · p. 12
• Figure A1 presents qualitative comparisons of GPT-4o-framework baseline posters and PosterForest posters across AI-conference papers, alongside author-created originals.The displayed labels identify GT, P2P, Paper2Poster, and Ours.Source: Figure A1 · p. 13, §C.2 · p. 13
• Participants viewed the original poster and four generated posters in randomized order, judging content fidelity, aesthetic quality, structural clarity, and overall quality.Source: §C.2 · p. 13

D.1 Failure Cases

• Papers with many densely clustered or interleaved figures and tables can trigger parsing and matching errors, causing missing or misaligned visuals that reduce poster informativeness and coherence.The illustrated failure includes figures missing from the poster or matched to the wrong sections.Source: §D.1 · p. 13

D.2 Future Directions

• Future work targets robust parsing of complex layouts, better modeling of figure importance and integration, and more reliable automated poster-quality metrics.The modeling goal is to better account for figure information, perceived importance, context, font size, and spatial balance, narrowing the gap with human designers.Source: §D.2 · p. 14

E.1 Parser Agent, AParser.

• The Parser Agent organizes titles, sections, subsections, paragraphs, and visual assets from the raw document into a hierarchical tree for downstream agents.Source: §E.1 · p. 14

E.2 Summarizing Agent, ASumm.

• The Summarizing Agent prunes, merges, and summarizes information across document levels, removing nonessential details to produce content optimized for poster presentation.Source: §E.2 · pp. 14–15

E.3 Content Agent, AContent.

• The Content Agent refines text at each Poster Tree node by adjusting volume, resolving redundancy, and selecting panel information with the Layout Agent.Source: §E.3 · p. 15

E.4 Layout Agent, ALayout.

• The Layout Agent adjusts panel arrangement, visual balance, and aspect ratio in coordination with the Content Agent to allocate space and maintain a coherent layout.Source: §E.4 · p. 15

E.5 Feedback Agent, AFeedback.

• The Feedback Agent evaluates rendered posters for visual organization, textual structure, and hierarchical balance, then returns structured feedback and a binary decision on further refinement.Source: §E.5 · p. 15
• Figure A8 presents a prompt example for the Feedback Agent, AFeedback.Source: Figure A8 · p. 23