A review of large language models and autonomous agents in chemistry

Analysis

Why This Paper Matters

This review arrives at a critical juncture where large language models are transitioning from pure language tasks to scientific discovery. By systematically covering LLM applications in chemistry—molecule design, property prediction, and synthesis optimization—it provides a much-needed map for researchers navigating this fast-moving field. The inclusion of autonomous agents, which combine LLMs with tools like web scraping and lab automation, extends the relevance beyond chemistry to any domain where AI can interact with physical or digital environments.

The paper's emphasis on challenges such as data quality, model interpretability, and the absence of standard benchmarks is particularly valuable. These are the bottlenecks that currently limit real-world adoption. The authors also point to future directions like multi-modal agents, which could integrate text, images, and experimental data, and enhanced collaboration between agents and experimental methods. This forward-looking perspective helps practitioners prioritize research investments.

Technical Contributions

• Comprehensive taxonomy: Categorizes LLM uses in chemistry into molecule design, property prediction, and synthesis optimization, with clear examples.
• Agent framework analysis: Reviews LLM-based autonomous agents, detailing their tool sets (e.g., paper scraping, lab APIs) and decision-making loops.
• Cross-domain extension: Broadens the agent review to scientific domains beyond chemistry, offering transferable insights.
• Living repository: Provides a GitHub repository (https://github.com/ur-whitelab/LLMs-in-science) to track the latest studies, addressing the field's rapid evolution.

Results

The paper does not present new experimental results but synthesizes findings from 212 cited works. It reports that LLMs have shown strong performance in property prediction and synthesis planning, though quantitative benchmarks are still emerging. The review notes that autonomous agents have successfully automated paper scraping and lab interfacing, but standardized evaluation metrics are lacking. The repository is actively maintained, reflecting the field's dynamism.

Significance

For AI practitioners, this review serves as both a primer and a roadmap. It clarifies where LLMs and agents currently excel (e.g., property prediction, synthesis planning) and where they fall short (e.g., interpretability, data integration). The identification of standard benchmarks as a key need is a call to action for the community. By linking to a living repository, the paper ensures its relevance over time. This work will likely influence how researchers design and evaluate LLM-based tools for scientific discovery, accelerating the adoption of AI in chemistry and beyond.