MACReD: Multi-Agent Framework for Chemical Reaction Diagram Parsing
A novel multi-agent hierarchical system named MACReD has been introduced for the interpretation of chemical reaction diagrams found in scientific texts. This system orchestrates specialized agents that focus on molecular perception, understanding arrows, extracting text, and reconstructing reactions, all within a cohesive VLM-guided framework. It tackles issues like diverse layouts, overlapping visual elements, and the merging of recognition with reasoning. Utilizing adaptable fine-grained detection, the planning and perception layers manage visual intricacies, while the reasoning layer applies a multigraph fusion approach to combine various cues and ensure globally consistent chemical reasoning. Tests on the RxnScribe benchmark validate the framework's efficiency. This research is available on arXiv under ID 2605.28077.
Key facts
- MACReD is a hierarchical multi-agent framework for parsing chemical reaction diagrams.
- It coordinates specialized agents for molecular perception, arrow understanding, text extraction, and reaction reconstruction.
- The framework uses a VLM-guided architecture.
- Planning and perception layers use flexible fine-grained detection.
- Reasoning layer uses a multigraph fusion mechanism.
- Experiments were conducted on the RxnScribe benchmark.
- The paper is available on arXiv with ID 2605.28077.
- The framework aims to integrate recognition and reasoning for complex diagrams.
Entities
Institutions
- arXiv