PatSnap/Hiro-Chemical-Insights

Hiro-Chemical-Insights: Multimodal Chemical Structure-Text Coreference

Hiro-Chemical-Insights is a specialized 8 billion parameter multimodal model developed by PatSnap, focusing on chemical structure-text coreference in intellectual property documents. This model is designed to process patent page images containing highlighted chemical structures.

Key Capabilities

• Identifies Reference Names: Given a boxed chemical structure in a patent image, the model accurately predicts its associated textual reference name(s).
• Classifies Structure Types: It categorizes chemical structures into types such as specific compound, substituent, Markush structure, or Markush structure & substituent.
• High Accuracy: The model demonstrates superior performance on its specific task, achieving an 'All Pass@1' score of 91.92 and 'All Pass@all' of 88.38. This significantly outperforms general multimodal LLMs like Gemini-2.5-Pro, which scored 73.23 and 63.13 respectively on the same metrics.
• Rule-guided Reinforcement Learning: The model's development is associated with the ACL 2026 paper "Multimodal Chemical Structure-Text Coreference in Intellectual Property via Rule-guided Reinforcement Learning", indicating a sophisticated training methodology.

Good For

• Automated Patent Analysis: Ideal for researchers and developers working on automated analysis of chemical intellectual property, particularly for extracting and classifying chemical information from patent documents.
• Chemical Information Extraction: Useful for tasks requiring precise identification and categorization of chemical structures and their textual mentions within complex documents.

Limitations

This model is intended for research in chemical structure-text coreference within patent documents and should not be used as a standalone source for chemical, medical, legal, or regulatory decisions.