# ChemCrow: Augmenting large-language models with chemistry tools

**Authors:** Andres M Bran, Sam Cox, Andrew D White, Philippe Schwaller  
**arXiv:** [2304.05376](https://arxiv.org/abs/2304.05376)  
**Date:** 2023-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.05376v1/

## Key points

- Chemistry-related problems are challenging for large-language models (LLMs) due to limited access to external knowledge sources.
- ChemCrow is an LLM chemistry agent designed to tackle tasks across organic synthesis, drug discovery, and materials design.
- ChemCrow integrates 13 expert-designed tools to augment LLM performance in chemistry and unlock new capabilities.
- Evaluation of ChemCrow's effectiveness includes both LLM and expert human assessments that demonstrate its ability to automate a diverse set of chemical tasks.
- GPT-4 as an evaluator cannot distinguish between clearly wrong GPT-4 completions and GPT-4 + ChemCrow performance, raising concerns about potential misuse of such tools like ChemCrow and highlighting the importance of responsible use.
- Responsible use of such tools can aid expert chemists while lowering barriers for non-experts and foster scientific advancement by bridging the gap between experimental and computational chemistry.
- With further development and responsible use, ChemCrow has the potential to significantly impact scientific research in various domains related to chemistry.

## Summary

Chemistry-related problems have been a challenge for large-language models (LLMs) due to their limited access to external knowledge sources. However, a recent study by Andres M Bran, Sam Cox, Andrew D White, and Philippe Schwaller introduces ChemCrow, an LLM chemistry agent designed to tackle tasks across organic synthesis, drug discovery, and materials design. By integrating 13 expert-designed tools, ChemCrow augments the LLM performance in chemistry and unlocks new capabilities. The evaluation of ChemCrow's effectiveness includes both LLM and expert human assessments that demonstrate its ability to automate a diverse set of chemical tasks. Interestingly, the study finds that GPT-4 as an evaluator cannot distinguish between clearly wrong GPT-4 completions and GPT-4 + ChemCrow performance. This raises concerns about the potential misuse of such tools like ChemCrow and highlights the importance of responsible use. The authors discuss the potential harms of misusing these tools but also emphasize their responsible use can aid expert chemists while lowering barriers for non-experts. Furthermore, they suggest that such tools can foster scientific advancement by bridging the gap between experimental and computational chemistry. Overall, this study presents an innovative solution to address challenges faced by LLMs in chemistry-related problems. With further development and responsible use, ChemCrow has the potential to significantly impact scientific research in various domains related to chemistry.