MACReD: A Multi-Agent Collaborative Reasoning Framework for Reaction Diagram Parsing

MACReD, a multi-agent AI framework, advances chemical reaction diagram parsing from scientific literature by achieving 75.2% F1 score on the RxnScribe benchmark—a 6.1 percentage point improvement over existing baselines. The system combines specialized agents for molecular recognition, arrow detection, and text extraction within a unified vision-language model architecture to handle complex spatial layouts in chemistry research documents.

MACReD represents a meaningful advancement in specialized multimodal AI systems designed to tackle domain-specific visual understanding challenges. The framework addresses a genuine pain point in computational chemistry: automated extraction of reaction information from academic papers, which traditionally requires manual curation. By decomposing the problem into specialized agents that coordinate through hierarchical planning and multigraph fusion mechanisms, the researchers demonstrate how structured reasoning can overcome limitations of general-purpose vision-language models on complex, heterogeneous visual layouts.

This work reflects a broader trend in AI development toward agent-based architectures and domain-specialized models rather than scaling monolithic foundation models. Chemical reaction parsing has practical implications for drug discovery, materials science, and literature mining in computational chemistry workflows. Organizations building chemistry-focused software platforms or conducting large-scale literature analysis could benefit from such improvements in automated diagram comprehension.

The 6.1 percentage point improvement in F1 scores indicates meaningful progress, though the 75.2% hard-match score suggests the system still misses roughly one-quarter of complex diagrams under strict evaluation criteria. This positions MACReD as a valuable tool for semi-automated workflows rather than fully autonomous parsing. The framework's ability to handle multi-step and tree-structured reactions demonstrates generalization beyond simple two-reactant scenarios.

Future developments should focus on closing the gap between hard and soft matching criteria and extending the approach to emerging chemistry diagram formats. Integration with academic databases and chemistry software platforms could unlock significant efficiency gains in research workflows.

• →MACReD achieves 75.2% F1 score on hard-match chemical diagram parsing, outperforming RxnScribe baseline by 6.1 percentage points.
• →Multi-agent architecture with specialized perception, planning, and reasoning layers enables structured handling of complex visual layouts.
• →Multigraph fusion mechanism enforces chemically consistent global reasoning across heterogeneous visual and textual cues.
• →Framework demonstrates strong generalization to multi-step and tree-structured chemical reactions in academic literature.
• →Results highlight effectiveness of decomposed agent-based approaches for domain-specific multimodal understanding tasks.