The team evaluated the performance and scalability of seven different parsers on the ALCF’s Polaris supercomputer. Image: Carlo Siebenschuh, Argonne National Laboratory and University of Chicago
Scientific progress increasingly depends on large-scale analysis of research papers, but most academic writing is organized in PDF format, which is difficult to process reliably. Even small parsing errors, such as changing “hyperthyroidism” to “hypothyroidism” or “pH” to “Ph,” can completely alter scientific meaning and compromise downstream research. To address this problem, researchers from Argonne National Laboratory and the University of Chicago are developing AdaParse, an adaptive framework that uses machine learning to select the best parsing strategy for each PDF. The team is leveraging ALCF computing resources to advance the development of AdaParse and apply it to parse large numbers of scientific papers.
Challenge
Parsing PDFs presents a fundamental obstacle for building multimodal science models and AI tools. PDFs are optimized for visual appearance rather than readability, meaning that text may be scrambled, chemical formulas corrupted, or figures and tables misinterpreted. Traditional parsing tools face a trade-off between speed and accuracy. Fast parsers often introduce serious errors, while high-quality approaches are slow and computationally expensive, limiting their use in large-scale processing efforts. Overcoming these constraints requires new methods that are simultaneously accurate, efficient, and scalable to millions of documents across diverse domains.
Approach
The team developed AdaParse (Adaptive Parallel PDF Parsing and Resource Scaling Engine) with an adaptive design that dynamically adjusts parsing strategies based on the input data. AdaParse uses a three-stage adaptive process. First, it quickly extracts text to assess document characteristics. Then, a machine learning model predicts which parsing method will yield the highest-quality results. Finally, the selected parser is applied across large collections of documents while distributing workloads across multiple nodes to optimize computational efficiency. The researchers incorporated human-in-the-loop principles and human expert feedback through direct preference optimization, aligning the system’s choices with what scientists prefer. The team evaluated the framework using ALCF’s Polaris supercomputer, enabling tests with large-scale, production-like log datasets and stress-testing performance on distributed-memory architectures representative of exascale systems.
Results
In benchmarking tests with 25,000 scientific documents spanning eight research domains and six major publishers, parses produced by AdaParse were preferred by scientists more often than any single parser, while running up to 17 times faster than state-of-the-art high-quality approaches. The adaptive selection process reduced errors like corrupted formulas or scrambled characters while maintaining high throughput. By scaling efficiently across distributed-memory architectures, AdaParse demonstrated the ability to process millions of documents at rates suitable for training advanced AI systems on scientific text.
Impact
AdaParse speeds up the creation of AI systems trained on scientific literature, leading to better AI research assistants, improved scientific discovery tools, and more accessible scientific knowledge. It has the potential to expand access to large-scale scientific datasets needed for training advanced AI models.
Publications
Siebenschuh, C., K. Hippe, O. Gokdemir, A. Brace, A. M. Khan, K. Hossain, Y. Babuji, N. Chia, V. Vishwanath, A. Ramanathan, R. L. Stevens, I. Foster, and R. Underwood. “AdaParse: An Adaptive Parallel PDF Parsing and Resource Scaling Engine,” Proceedings of the 8th MLSys Conference (May 2025) https://doi.org/10.48550/arXiv.2505.01435
