💻 computer science

PosterForest: Hierarchical Multi-Agent Collaboration for Scientific Poster Generation

PosterForest is a training-free framework that leverages a structured "Poster Tree" representation and hierarchical multi-agent collaboration to jointly optimize content and layout, thereby generating scientifically coherent and visually balanced posters without the need for additional training or domain-specific supervision.

Original authors: Jiho Choi, Seojeong Park, Seongjong Song, Hyunjung Shim

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiho Choi, Seojeong Park, Seongjong Song, Hyunjung Shim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have just written a massive, 30-page scientific research paper. It's full of complex data, dense paragraphs, and dozens of charts. Now, you need to present this at a conference, but you only have a single large poster board to work with.

The Problem:
Trying to squeeze that whole book onto one board is a nightmare. If you just copy-paste, it's a mess. If you try to summarize it yourself, you might accidentally put the wrong chart next to the wrong paragraph, or make the text so small no one can read it.

Existing computer programs that try to do this automatically are like clumsy interns. They might grab the right words but put them in the wrong order, or they might shrink a giant graph until it looks like a pixelated smudge. They often treat the paper like a flat list of words rather than a structured story.

The Solution: PosterForest
The authors of this paper created a new system called PosterForest. Think of it not as a single robot, but as a team of expert architects and editors working together to build a perfect poster.

Here is how it works, using a simple analogy:

1. The "Tree" Blueprint (The Poster Tree)

Instead of looking at the paper as a flat wall of text, PosterForest first builds a hierarchical tree.

The Analogy: Imagine the research paper is a giant oak tree. The trunk is the main title, the big branches are the main sections (like "Methods" or "Results"), the smaller twigs are paragraphs, and the leaves are the specific facts, charts, and tables.
What it does: The system maps out this entire tree structure. It understands that a specific chart belongs to a specific branch (section) and shouldn't be floating around randomly. This ensures the "logic" of the paper is preserved.

2. The Team of Agents (The Collaboration)

Once the tree is built, the system doesn't just print it. It sends the blueprint to a team of specialized "agents" (AI workers) who argue and refine the design together.

The Content Agent (The Editor):
- Role: This agent is like a strict editor. It looks at the "leaves" of the tree and says, "This paragraph is too long for a poster; let's cut the fluff." or "This section is missing a key point; let's add it."
- Goal: To make sure the text is concise and fits the space without losing the main message.
The Layout Agent (The Architect):
- Role: This agent is like an interior designer. It looks at the "branches" and asks, "If we put this huge chart here, the text will be squished. Let's move this section over there and make the panel bigger."
- Goal: To ensure the poster looks balanced, the charts are readable, and the white space is used well.
The Feedback Agent (The Critic):
- Role: This agent acts like a museum curator walking through the gallery. It looks at the whole poster and says, "Hey, the top half looks crowded, and the bottom looks empty. Let's fix that."
- Goal: It gives the team a "thumbs up" or "thumbs down." If it says "thumbs down," the team goes back and tweaks the tree again.

3. The "Forest" Process (Iterative Refinement)

The magic of PosterForest is that these agents don't just work once. They work in a loop, like a sculptor chipping away at a block of marble.

They build a rough draft.
The Critic says, "Too much text on the left."
The Editor cuts the text; the Architect moves the chart.
The Critic checks again.
They repeat this until the poster is perfect.

Why is this better?

Old methods were like a photocopier that just shrinks the whole page. The result is often blurry, disorganized, or missing key details.
PosterForest is like a human design team. It understands the story of the paper. It knows that the "Conclusion" should come after the "Results," and it knows that a complex graph needs more space than a simple bullet point.

The Result

The paper shows that this "team of agents" approach creates posters that are:

Easier to read: The text isn't tiny.
Logically sound: The story flows naturally from top to bottom.
Visually balanced: No awkward empty spaces or crammed corners.

In short, PosterForest turns the chaotic job of turning a dense research paper into a clear, beautiful poster into a collaborative, intelligent process, ensuring that the science shines through without the clutter.

1. Problem Statement

The generation of scientific posters from academic papers is a critical task for disseminating complex research, yet it remains a labor-intensive process requiring both domain expertise and design skills. Existing automated methods suffer from three primary limitations:

Shallow Document Understanding: Current approaches often rely on surface-level text features, failing to grasp the deep hierarchical structure of scientific documents (e.g., section-subsection-paragraph relationships) and the semantic alignment between text and visual elements (figures/tables). This leads to interrupted logical flows and misplaced content.
Weak Content-Layout Integration: Most methods use a sequential pipeline where layout is determined before content placement, or they optimize content and layout separately. This decoupling ignores the intrinsic interdependence between the two, resulting in truncated content, poor visual balance, and overly simplified posters that fail to capture the original document's complexity.
Training Overhead: State-of-the-art methods often require instruction tuning or regression-based training, which adds complexity and limits practical deployment.

2. Methodology: PosterForest

The authors propose PosterForest, a training-free framework that utilizes hierarchical multi-agent collaboration to generate scientific posters. The method operates in two main stages:

A. Hierarchical Paper-to-Poster Representation (The Poster Tree)

Instead of treating the paper as a linear text sequence, PosterForest constructs a unified intermediate representation called the Poster Tree ( $T_{Poster}$ ). This is achieved through a three-step process:

Raw Doc Tree ( $T_{Raw}$ ): A parsing agent extracts the document's structural hierarchy (title, sections, subsections, paragraphs, figures, tables) and their referential links.
Content Tree ( $T_{Content}$ ): A refinement agent prunes redundant nodes, merges related content, and summarizes lengthy text while preserving the logical hierarchy. Crucially, this step explicitly links text nodes with their corresponding visual assets.
Layout Tree ( $T_{Layout}$ ): An initialization agent maps the content hierarchy to a spatial layout (rows, columns, panels), inheriting the hierarchical relationships from the content tree.
Merging: The Content and Layout trees are merged into the Poster Tree, where each node jointly encodes semantic attributes (what is displayed) and spatial attributes (where and how it is displayed).

B. Hierarchical Modification Planning (Multi-Agent Collaboration)

Once the initial Poster Tree is constructed, the framework employs an iterative refinement process driven by specialized agents:

Layout Agent ( $A_{Layout}$ ): Optimizes geometric attributes (region ratios, alignment, spatial distribution) based on the structural and semantic statistics of descendant nodes.
Content Agent ( $A_{Content}$ ): Refines textual density and semantic abstraction, ensuring the content fits the layout constraints and avoids redundancy by referencing sibling nodes.
Global Feedback Agent ( $A_{Feedback}$ ): A multimodal large language model (MLLM) evaluates the rendered poster at the tree level. It analyzes visual organization, textual structure, and hierarchical balance. Based on this evaluation, it provides structured feedback and a binary signal to either continue refinement or terminate.

This process involves node-level traversal (updating individual nodes based on parent/child context) and tree-level iteration (repeating the traversal based on global feedback) until the poster achieves optimal coherence and balance.

3. Key Contributions

Poster Tree Representation: A novel, structured intermediate representation that unifies document semantics and layout planning, explicitly modeling the hierarchy and text-visual associations inherent in scientific papers.
Hierarchical Multi-Agent Collaboration: A training-free framework that decomposes the complex task into specialized roles (Content, Layout, Feedback). These agents collaborate iteratively to jointly optimize content fidelity and visual harmony, overcoming the limitations of sequential pipelines.
Training-Free Deployment: The system relies solely on standard APIs (e.g., GPT-4o, Qwen) and public checkpoints, requiring no additional model training or domain-specific supervision, making it highly practical for deployment.

4. Experimental Results

The authors evaluated PosterForest against state-of-the-art baselines (P2P, Paper2Poster, and various end-to-end GPT-4o approaches) using both automatic and human evaluations.

Qualitative Evaluation: Visual comparisons demonstrate that PosterForest produces more balanced layouts, avoids oversized/undersized figures, and preserves critical information (tables/graphs) better than baselines, which often omit key data or create disjointed logical flows.
Quantitative Evaluation (MLLM-as-a-Judge): Using GPT-4o as a judge across six criteria (element quality, layout balance, engagement, clarity, content completeness, logical flow), PosterForest achieved scores comparable to or exceeding baselines, particularly in informativeness and logical flow, closely matching human-created ground truth posters.
Human Evaluation: In a user study with 25 AI graduate students, PosterForest was preferred over all baselines across all criteria. It achieved a 59.2% preference rate in overall quality, significantly outperforming the next best method (Paper2Poster at 27.2%).
Ablation Studies: Experiments confirmed that the hierarchical structure is essential for preserving logical order and that the combination of both Content and Layout agents is necessary to balance information density with visual harmony.

5. Significance

PosterForest represents a significant advancement in automated scientific communication. By moving away from flat summarization and decoupled optimization, it addresses the core challenge of maintaining the semantic integrity and visual coherence of complex research papers. The framework's training-free nature makes it immediately accessible to researchers without the need for extensive computational resources or data collection for fine-tuning. Ultimately, it lowers the barrier for researchers to create high-quality, publication-ready posters, thereby accelerating the dissemination of specialized scientific knowledge.