A 2024 end-to-end OCSR system addressing both technical and data challenges, introducing MolParser-7M (7M+ image-text pairs) and MolDet (YOLO-based detector) for extracting and recognizing molecular structures from real-world documents with diverse quality and styles.
ICML 2025 paper introducing SpaceFormer, a Transformer architecture that challenges the atom-centric paradigm by modeling the continuous 3D space surrounding molecules using adaptive multi-resolution grids, achieving state-of-the-art performance on quantum property prediction benchmarks.
We explore LLM applications for page stream segmentation in insurance document processing, demonstrating that parameter-efficient fine-tuning achieves strong accuracy but revealing significant calibration challenges that limit deployment confidence.
We create TabMe++, an enhanced page stream segmentation benchmark with commercial-grade OCR, and show that parameter-efficiently fine-tuned decoder-based LLMs like Mistral-7B achieve 80% straight-through processing rates, dramatically outperforming encoder-based models.
A provocative 2024 Nature Machine Intelligence paper challenging the assumption that invalid SMILES are failures, showing empirically that the ability to generate invalid outputs actually improves chemical language model performance by enabling quality filtering and providing richer training signals.
What happens when you achieve 99.8% accuracy on sarcasm detection? You might have accidentally built a domain classifier. A cautionary ML tale about dataset bias.
Bachelor’s thesis introducing PyConversations, an open-source library that normalizes over 308 million posts from Twitter, Reddit, Facebook, and 4chan into a unified data model for cross-platform social media research.
Research project that investigated how different NLP models perform on social media data, finding that domain-specific approaches often outperform large pre-trained models. Includes PyConversations, a Python module for analyzing conversations across social media platforms.
We demonstrate that universal adversarial triggers can control both the topic and stance of GPT-2’s generated text, revealing security vulnerabilities in deployed language models and proposing constructive applications for bias auditing.