Feasibility of LLMs Assisting Human Reasoning in Theoretical Physics

Establish whether large language models can assist or augment human reasoning in specialized research settings in theoretical physics so as to enable the generation of new scientific knowledge.

Background

The paper investigates whether LLMs can perform research-level tasks in quantum many-body physics by focusing on Hartree-Fock mean-field calculations. The authors highlight a key uncertainty about the role of LLMs in specialized research workflows and frame their paper as an initial exploration of that question.

They design multi-step prompt templates and evaluate GPT-4 across a curated corpus of papers to assess the model’s ability to execute analytic calculations, aiming to shed light on the broader feasibility of using LLMs to augment human reasoning in theoretical physics.