Research map / Vision-language and multimodal / Multimodal LLMs

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

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Interleaved image-text generation with MLLMs

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes P2P, an LLM-based multi-agent framework that turns research papers into HTML academic posters, plus the P2Pinstruct training set (30K+ pairs), a fine-grained P2Peval benchmark, and a fine-tuned Qwen3-P2P-8B.

Problem

Automated academic poster generation struggles to preserve scientific detail and visual-textual integration, and lacks standardized, fine-grained benchmarks and instruction data for evaluation and training.

Approach

A three-agent (Figure, Section, Orchestrate) pipeline with checker modules and reflection loops that renders posters as HTML/CSS, paired with checklist-based fine-grained evaluation and an XGBoost-calibrated universal evaluator fitted to human scores.

Key results

• Claude-3.7-Sonnet with P2P reaches FineGrain 65.40 and Universal 37.25 on P2Peval, the top Universal score in the table.
• Qwen3-P2P-8B achieves the highest ROUGE-1/2/L among all evaluated models (0.2882/0.0955/0.2675) and improves FineGrain from 45.03 to 57.66 over base Qwen3 (T setting).
• Human pairwise study: P2P posters are strictly preferred over original author posters in 35.59% of comparisons, with 57.63% preferred or tied.

Surprising findings

• Qwen3-P2P-8B achieves the highest ROUGE scores but not the highest fine-grained or universal scores, showing lexical overlap does not track fidelity.
• Ablating the Figure Describer and passing raw images to the MLLM degrades performance, indicating textual figure descriptions help more than raw visuals.

Limitations

• Benchmark is small (121 paper-poster pairs) and drawn from ACL papers and SciPostLayout.
• Universal evaluation relies on GPT-4o judgments and an XGBoost model calibrated on 1,701 human scores.
• Some baseline comparisons mix settings (e.g., Claude used as figure describer for text-only models).

Experimental setup

Datasets: P2Pinstruct SciPostLayout ACL conference papers 2022-2024

Benchmarks: P2Peval

Metrics: ROUGE-1 ROUGE-2 ROUGE-L BERTScore F1 Fine-Grained score Universal score LLM-as-judge preference rate human pairwise preference

Baselines: Claude-3.7-Sonnet GPT-4o GPT-4.1 Qwen3 Qwen2.5-VL Deepseek-R1 YuanBao Gemma-3 InternVL3

Builds on: DocLayout-YOLO SciPostLayout Design2Code

Closest papers by meaning

• Paper2Poster: Towards Multimodal Poster Automation from Scientific Papers (NeurIPS 2025 · similarity 0.94)
• POSTA: A Go-to Framework for Customized Artistic Poster Generation (CVPR 2025 · similarity 0.86)
• GlyphDraw2: Automatic Generation of Complex Glyph Posters with Diffusion Models and Large Language Models (AAAI 2025 · similarity 0.86)
• Presenting a Paper is an Art: Self-Improvement Aesthetic Agents for Academic Presentations (ICLR 2026 · similarity 0.85)
• OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.85)
• PosterO: Structuring Layout Trees to Enable Language Models in Generalized Content-Aware Layout Generation (CVPR 2025 · similarity 0.84)