PaperFit: Vision-in-the-Loop Typesetting Optimization for Scientific Documents

Researchers introduce PaperFit, a vision-in-the-loop AI agent that automates the typesetting optimization of LaTeX scientific documents by iteratively rendering pages, diagnosing visual defects, and applying constrained repairs. The work formalizes Visual Typesetting Optimization (VTO) as a critical missing stage in document automation, addressing the gap between compilable but visually flawed PDFs and publication-ready outputs through a new benchmark of 200 papers.

PaperFit addresses a genuine friction point in academic publishing: the tedious compile-inspect-edit cycles required to transform technically correct LaTeX documents into visually polished submissions. While rule-based tools operate blindly on source code and text-only LLMs cannot predict two-dimensional layout consequences, this work demonstrates that closing the visual feedback loop—rendering pages, diagnosing defects, then applying fixes—substantially improves outcomes.

The research emerges from broader advances in multimodal AI systems capable of understanding both text and rendered visuals. As large language models increasingly handle document generation tasks, the inability to verify visual results has become a bottleneck. PaperFit's taxonomy of five typesetting defect categories and comprehensive benchmark across 200 papers with 13 defect types provide empirical grounding often missing from applied AI research.

For academic publishers and research institutions, this work signals that automation can extend beyond simple text generation into complex visual-spatial optimization. The success rate improvements over baselines suggest practical deployment potential, potentially reducing author submission cycles and accelerating peer review timelines. For AI researchers, PaperFit exemplifies how constrained problem domains with clear success metrics and verifiable outputs can showcase vision-language model capabilities more convincingly than open-ended tasks.

The practical impact remains somewhat niche—limited to LaTeX document workflows in academic contexts—but the underlying methodology of vision-in-the-loop optimization applies broadly to document automation, design systems, and any domain requiring visual verification of programmatic changes.

• →PaperFit uses vision-in-the-loop feedback to optimize LaTeX typesetting, outperforming text-only baselines by significant margins.
• →Visual Typesetting Optimization (VTO) is formalized as a five-category defect taxonomy addressing floats, equations, tables, widows, and page balance.
• →The PaperFit-Bench benchmark includes 200 papers across 10 venue templates and 13 defect types, enabling rigorous evaluation.
• →Text-only LLMs cannot predict two-dimensional layout consequences, making visual verification essential for publication-ready documents.
• →This work identifies a critical missing stage in document automation pipelines between compilation and submission-ready outputs.