PosterForest: Automated Scientific Poster Generation

• PosterForest is a training-free framework that integrates hierarchical representation and multi-agent collaboration for automated scientific poster generation.
• The system employs a multi-stage process using a Poster Tree structure to merge semantic content and layout design, ensuring high content fidelity and visual coherence.
• Empirical evaluations demonstrate that PosterForest outperforms traditional methods by achieving superior logical structuring, design balance, and overall poster quality.

PosterForest is a training-free framework for automated generation of scientific posters that integrates hierarchical document representation and multi-agent collaboration. The system is engineered to produce posters that closely match expert-designed ground truth in terms of information preservation, visual clarity, and user preference by jointly optimizing logical document structure, content fidelity, and aesthetic coherence.

1. System Overview

PosterForest's architecture contrasts with earlier approaches, which generally employ linear text extraction followed by isolated layout prediction. Instead, PosterForest introduces a hierarchical, multi-agent method wherein specialized agents for content summarization and layout planning coordinate iteratively to refine both semantic and visual elements. This coordinated workflow obviates the need for explicit predictive modeling or end-to-end regression pipelines, yielding a unified framework that better captures complex relationships between document sections, figures, textual summaries, and layout attributes.

2. Poster Tree: Hierarchical Intermediate Representation

A central innovation of PosterForest is the Poster Tree, a hierarchical structure encoding document layout and semantic relations at multiple levels. Construction begins with the Raw Document Tree (𝒯_raw), parsed into sections, paragraphs, and visual assets. A summarization agent prunes and merges details to form the Content Tree (𝒯_content), while a Layout Initialization agent generates a Layout Tree (𝒯_layout) specifying hierarchical arrangement of panels and regions. These are merged:

$$\mathcal{T}_{\text{poster}} = \operatorname{Merge}(\mathcal{T}_{\text{content}}, \mathcal{T}_{\text{layout}})$$

Each node in $\mathcal{T}_{\text{poster}}$ carries semantic (e.g., text, figure details) and layout (e.g., size, placement, aspect ratio) attributes, enabling joint consideration of information value and spatial organization. This mechanism provides a natural mapping between the original document’s logical structure and the visual hierarchy of the poster.

3. Multi-Agent Iterative Coordination

At each node of the Poster Tree, specialized agents analyze and refine both content and layout in a coordinated fashion:

• Stage 1 (Analysis):
  • Layout Agent evaluates spatial variables: panel balance, aspect ratios, figure allocation.
  • Content Agent checks text information density vis-à-vis layout constraints.
• Stage 2 (Collaboration):
  • Agents exchange feedback, iteratively updating opinions:
  • Content agent generates $O_c' = \mathcal{A}_{\text{Content}}(O_l)$.
  • Layout agent updates $O_l' = \mathcal{A}_{\text{Layout}}(O_c)$.
  • Joint negotiation adjusts summaries and layout to resolve overflow or underutilization, preserving structural coherence.
• Stage 3 (Finalization):
  • Consensus attributes $\{c^*, l^*\}$ are committed for each node after one or more rounds. Breadth-first traversal ensures consistency across the tree.

This iterative agent negotiation, formalized at each hierarchical level, achieves a coupled optimization of logic and aesthetics absent in purely sequential methods.

4. Logical, Semantic, and Visual Optimization

PosterForest is explicitly optimized for:

• Logical Consistency: Hierarchical relationships (sections, figures, panels) from the source document are maintained in the generated tree.
• Content Fidelity: Summarization preserves domain-essential detail and research contributions at each node.
• Visual Coherence: Layout planning dynamically adapts element position, grouping, and sizing to maximize clarity and information accessibility.

Rather than optimizing these objectives independently, the agents utilize ongoing feedback to reach global optimality at both the micro (panel/section) and macro (poster-wide) levels. By updating nodes through feedback from both content and layout perspectives, the system harmonizes information density with space allocation and layout aesthetics.

5. Evaluation Across Domains

PosterForest was benchmarked against state-of-the-art automated poster generation systems on multiple academic domains (computer vision, NLP, reinforcement learning). Evaluations using MLLM-as-Judge protocols (such as GPT-4o) rated posters for element quality, layout balance, content completeness, and aesthetic coherence. PosterForest achieved:

• Comparable or superior scores in visual design metrics.
• Outperformance in informativeness and content completeness relative to systems like P2P and Paper2Poster.
• Balanced use of space, clear grouping, and reduced figure-text misalignment in qualitative analysis.

User studies among graduate researchers corroborated quantitative findings: PosterForest posters ranked highest in content fidelity, structural clarity, and overall quality, with a notably high proportion of first-rank preferences.

6. User Preferences and System Impact

User feedback demonstrates that the preservation of document hierarchy and the iterative agent collaboration underpin PosterForest’s advantage in readability and professional presentation. Participants consistently cited logical structuring and visual distribution of information as decisive factors. This engineering approach—tightly coupling summarization and layout refinement—addresses key shortcomings in prior systems such as fragmented structure and informational loss.

7. Significance and Advancement

PosterForest advances automated poster generation by introducing the Poster Tree as a joint semantic-visual scaffold and deploying hierarchical, multi-agent negotiation for integrated optimization. The approach aligns closely with how expert designers balance content and aesthetic constraints in manual poster creation. Empirical and user-driven evaluation demonstrates that this joint methodology produces results of high informational value, clarity, and visual appeal.

This represents a departure from pipeline-based designs and regression-based layout prediction, illustrating the potential of hierarchical representation and collaborative agent frameworks for complex compositional tasks in scientific communication.