Using large language models for research in the natural sciences (github.com)

A new guide by Danielle Loader, Jonathan Oppenheim, and Tobias J. Osborne outlines effective workflows for employing large language models (LLMs) in natural sciences research. While LLMs significantly enhance the speed of idea exploration and enable scientists to tackle grand challenges, they often generate plausible yet incorrect outputs. To mitigate this issue, the authors suggest employing adversarial verification protocols alongside structured proof formats developed by Lamport, which help expose logical dependencies and refine findings into manageable sub-claims. The significance of this guide lies in its focus on enhancing the reliability of LLM outputs without resorting to temporary fixes. By advising scientists to optimize the input to LLMs and utilizing wikis and selective notebooks, the authors stress the importance of diagnosing common issues like context rot and hyperfixation. This enables researchers to inspect model outputs critically while remaining aware of the broader risks AI technologies pose. Ultimately, the guide aims to bridge the gap between casual use of LLMs and their more effective, rigorous application in scientific research, making advanced AI tools more accessible for those with limited experience.