Georgia Tech’s IPO-Mine Organizes IPO Filings; Models Can Disagree With Experts on Charts

Georgia Tech’s Financial Services Innovation Lab has developed IPO-Mine, an open-source toolkit and dataset for long, multimodal IPO documents. It organizes filings into structured sections and extracts charts and infographics, addressing disclosures that can run hundreds of pages and combine legal prose, financial data, tables, and visuals.

IPO filings are disclosures submitted to the U.S. Securities and Exchange Commission before a company enters public markets. Some can exceed hundreds of thousands of words, and their mix of text, tables, charts, and infographics makes whole-document review difficult for both people and AI systems.

IPO-Mine’s mechanism is organizational rather than a claim to automatic financial judgment. It separates a filing into sections, standardizes inconsistent section formats, and preserves extracted visuals for joint text-and-image analysis. The researchers say that structure can support comparisons across companies and industries instead of leaving each filing as a single, unwieldy document.

Our results show that even strong multimodal models can disagree with expert human judgments on financial charts, especially when they are misleading.

Vidhyakshaya Kannan, Georgia Tech Financial Services Innovation Lab intern and study co-author

The accompanying study, IPO-Mine: A Toolkit and Dataset for Section-Structured Analysis of Long, Multimodal IPO Documents, is posted as an arXiv preprint. Its core finding is a split in the documents themselves: written IPO sections are becoming more standardized, while charts and infographics are becoming more complex and varied.

That divergence changes where automated analysis gets difficult. Co-author Vidhyakshaya Kannan says boilerplate language increasingly resembles language in other filings, while visuals carry more of a company’s distinctive presentation to investors. As a result, visual interpretation becomes a growing constraint: systems can organize the material at scale, but the figures that differentiate a filing may be the least reliable part to delegate without expert review.

Three obstacles remain after extraction

• Longer documents can reduce model performance across an entire filing.
• AI tools can struggle to interpret visual data correctly, even when the charts have been extracted.
• Inconsistent combinations of text, tables, and visuals complicate reliable multimodal processing.

The authors position IPO-Mine for researchers, regulators, and investors who want to study IPO disclosures more systematically, including patterns in how companies communicate risk, performance, and strategy before listing. They also describe a use case at broader scale: examining disclosure practices across decades, industries, and thousands of companies.

The preprint’s practical boundary is clear. IPO-Mine can make a large, mixed-format corpus easier to inspect and compare, but it does not remove the need for human judgment when a chart’s framing or meaning affects the conclusion. The tool’s value is therefore in making systematic review more feasible while keeping the visual layer visible to the people responsible for interpreting it.