AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations

Abstract: High-quality scientific illustrations are crucial for effectively communicating complex scientific and technical concepts, yet their manual creation remains a well-recognized bottleneck in both academia and industry. We present FigureBench, the first large-scale benchmark for generating scientific illustrations from long-form scientific texts. It contains 3,300 high-quality scientific text-figure pairs, covering diverse text-to-illustration tasks from scientific papers, surveys, blogs, and textbooks. Moreover, we propose AutoFigure, the first agentic framework that automatically generates high-quality scientific illustrations based on long-form scientific text. Specifically, before rendering the final result, AutoFigure engages in extensive thinking, recombination, and validation to produce a layout that is both structurally sound and aesthetically refined, outputting a scientific illustration that achieves both structural completeness and aesthetic appeal. Leveraging the high-quality data from FigureBench, we conduct extensive experiments to test the performance of AutoFigure against various baseline methods. The results demonstrate that AutoFigure consistently surpasses all baseline methods, producing publication-ready scientific illustrations. The code, dataset and huggingface space are released in https://github.com/ResearAI/AutoFigure.

• The paper introduces AutoFigure, a novel agentic framework that automates the creation of publication-ready scientific illustrations from long-form texts.
• It employs a decoupled pipeline with concept extraction, iterative critique-and-refine, and style-guided rendering to ensure semantic alignment and visual excellence.
• Quantitative and expert evaluations show its strong performance, with win rates up to 97.5% in textbook diagrams and significant acceptance for camera-ready publications.

AutoFigure: Automated Generation of Publication-Ready Scientific Illustrations

Introduction and Context

"AutoFigure: Generating and Refining Publication-Ready Scientific Illustrations" (2602.03828) addresses a persistent and impactful bottleneck in scientific communication: the manual creation of high-quality, publication-ready illustrations from long-form scientific texts. While existing automated methods focus primarily on extracting and recombining pre-existing figures or rendering captions into images, there is a notable deficit in automated systems capable of distilling the core methodologies and conceptual workflows from extensive documents (often exceeding 10,000 tokens) and synthesizing novel, semantically aligned, and visually compelling illustrations ab initio.

The work positions itself amid two major research threads. First, the development of datasets and benchmarks for scientific figure generation, which have until now relied upon succinct, structurally impoverished inputs such as captions or figure legends. Second, recent advances in agentic systems for multimodal scientific content synthesis (e.g., PosterAgent, PPTAgent) that perform extraction and reformatting but lack the capacity for true conceptual generation from unstructured text. Both approaches fail to simultaneously achieve structural fidelity, aesthetic appeal, and content completeness—three critical dimensions for figures put forward in peer-reviewed publications.

FigureBench: Foundation for Benchmarking Scientific Illustration Generation

A primary contribution of the paper is FigureBench, the first large-scale benchmark explicitly designed for Long-context Scientific Illustration Design. The dataset comprises 3,300 expert-annotated, high-complexity text-figure pairs drawn from research papers, surveys, technical blogs, and textbooks. Notably, the dataset emphasizes conceptual illustrations over data visualizations and ensures exhaustive grounding of all visual elements in the source text. The construction pipeline utilizes expert raters, achieving high inter-rater reliability (Cohen's κ = 0.91), and is further refined via fine-tuned vision-LLMs to filter and expand the development set.

Statistical analysis highlights the significant challenge imposed by FigureBench: average text length in the "Paper" category is over 12,000 tokens, with high text density (avg. 41.2%), and complex structural characteristics (avg. 6.2 colors, 5.3 components, 6.4 shapes per illustration). This positions FigureBench as a robust evaluation bed for models required to operate at the intersection of logical reasoning, design, and multimodal generation.

AutoFigure: Agentic, Decoupled Reasoned Rendering

The core technical contribution is the AutoFigure framework, which employs an agentic, decoupled pipeline based on a Reasoned Rendering paradigm. The process is decomposed into distinct stages optimized for semantic reasoning, layout optimization, and high-fidelity rendering:

Stage I: Concept Extraction and Symbolic Layout Generation

• A LLM agent parses long-form scientific text and distills a methodological summary alongside explicit entities and relations, serialized as a structured, machine-readable graph (e.g., SVG/HTML).
• This layout encodes the 2D geometry and topology required for downstream schematic generation.

Stage II: Critique-and-Refine Self-Refinement Loop

• An iterative loop simulates a dialogue between an AI designer and an AI critic, refining layout drafts for improved logical coherence, balance, and absence of visual artifacts.
• Critique-derived feedback steers subsequent generations, with quantitative scoring guiding convergence toward optimal layout hypotheses.

Stage III: Style-Guided Rendering and Text Post-processing

• Layout and style descriptors are passed to a multimodal generative model, which synthesizes the final image, conditioned to maintain structural constraints.
• An "erase-and-correct" module remedies the typical failures of T2I models in rendering crisp vector text: text is removed using an eraser, OCR extraction is conducted, candidate strings are verified and corrected against the symbolic layout, and vector-quality overlays are composed onto the image.

This agentic, compositional pipeline is architected for modular extension, allowing direct substitution of reasoning, rendering, and verification backbones.

Evaluation: Quantitative, Qualitative, Human Expert Assessment

Automated and Human Evaluations

AutoFigure is evaluated via a comprehensive protocol that combines "VLM-as-a-Judge" referenced and blind pairwise scoring, supplemented by rigorous human expert review. The VLM-based evaluation decomposes figure quality into visual design (aesthetic, expressiveness, polish), communication effectiveness (clarity, logical flow), and content fidelity (accuracy, completeness, appropriateness); both absolute scores and win-rates in head-to-head comparisons are reported for each dimension and class of document.

Key Results:

• Consistent outperformance of all baselines (T2I models, SVG/HTML code generation, agentic planners) in both overall scores and win rates across blog, survey, textbook, and—most notably—paper categories.
• AutoFigure achieves an exceptional 97.5% win rate in the textbook category and 66.7% of domain experts are willing to use its figures in camera-ready papers, indicating acceptance against real-world publication standards.
• The ablation studies demonstrate that both the iterative critique-and-refine loop and the "erase-and-correct" rendering add substantial value, with multi-iteration search boosting overall performance from 6.28 to 7.14, and post-processing yielding up to +0.10 in crucial visual quality metrics.

Qualitative analysis identifies that T2I baselines typically fail on logical structure, while code-generation methods produce visually dry, fragmented output. AutoFigure uniquely integrates hierarchical semantic decomposition, element grouping, and role-specific styling, successfully synthesizing novel visual grammars in research diagrams with diverse conventions.

Open-Source Backbone Evaluation

AutoFigure's modularity is reinforced by competitive performance with open-source large-scale VLMs; Qwen3-VL-235B achieves an overall score of 7.08, rivaling GPT-5 and surpassing several proprietary alternatives, validating reproducibility and accessibility to the broader research community.

Human-LLM Correlation

The evaluation paradigm is further validated through statistical analysis of human-LLM agreement, achieving high Pearson (r = 0.659) and Spearman (ρ = 0.593) correlations on quality and ranking tasks, indicating alignment between human and VLM judgments across all tested dimensions.

Limitations and Theoretical Implications

Despite demonstrable strengths, AutoFigure still exhibits limitations:

• Text rendering at fine granularity remains a weakness under dense layouts or small font sizes, with rare character-level errors persisting despite the "erase-and-correct" module.
• Boundary adherence in semantic-to-visual mapping: The system can drift beyond strict literal content when grappling with abstract, underspecified source text, sometimes sacrificing precise ontological distinctions for visual clarity.
• Hierarchical complexity bottleneck: Research paper diagrams, with multi-level entity organization and minimal visual precedent, continue to expose the limits of current layout-planning capacity.

These observations highlight open research challenges at the intersection of knowledge-grounded generation, formal logical reasoning, and scalable vector rendering—particularly the need for domain-verifiers and knowledge-augmented constraint enforcement.

Broader Impact and Future Development

AutoFigure represents a significant advancement toward fully automated scientific communication agents, directly addressing the present-day bottleneck impeding autonomous AI-driven research systems: the inability to visually articulate findings at human publication standards. The implications span both practical workflows—democratizing access to publication-grade figures for researchers without design expertise—and theoretical AI capabilities—moving toward agents capable of both discovering and communicating complex phenomena in a self-contained manner.

Possible future directions include:

• Extension to other domains (life sciences, economics) requiring incorporation of domain-specific visual conventions.
• Interactive and dynamic schematic generation (e.g., animated, exploratory, or temporally contextualized figures).
• Incorporation of verification-oriented modules (retrieval-augmented, rule-based domain checkers) to close the fidelity gap in complex, layered diagrams.
• Enhanced multimodal interpretability tools for AI-generated scientific outputs within AI scientist pipelines.

Conclusion

By coupling FigureBench, a demanding benchmark that enforces high standards on both reasoning and design, with the AutoFigure Reasoned Rendering agentic pipeline, this work demonstrates tangible progress in automating a critical aspect of scientific knowledge dissemination. The framework resolves core trade-offs in existing methods, enabling both semantic alignment and aesthetic achievement, and is validated through rigorous multi-modal, multi-perspective evaluation. While open problems remain in handling the full complexity of scientific visual language and ultra-fine text rendering, these contributions lay a strong, extensible foundation for the next generation of AI-driven scientific research and communication systems (2602.03828).