Can Theoretical Physics Research Benefit from Language Agents?

yet their application in theoretical physics research is not yet mature. This position paper argues that LLM agents can potentially help accelerate theoretical, computational, and applied physics when properly integrated with domain knowledge and toolbox. We analyze current LLM capabilities for physics—from mathematical reasoning to code generation—identifying critical gaps in physical intuition, constraint satisfaction, and reliable reasoning. We envision future physics-specialized LLMs that could handle multimodal data, propose testable hypotheses, and design experiments. Realizing this vision requires addressing fundamental challenges: ensuring physical consistency, and developing robust verification methods. We call for collaborative efforts between physics and AI communities to help advance scientific discovery in physics.

LLMs can generate textbook-style explanations through summarization, yet this apparent understanding can be superficially derived from statistical correlations rather than causal models of physical laws [18, 19]. This is evident in explanations that seem correct but contain subtle physical inaccuracies or miss crucial assumptions (e.g., in the context of perturbation theory, failing to state the conditions for its validity)

promising progress in applying formulas to well-defined problems mirroring textbook examples [27, 38, 26], but they may resort to memorized solutions rather than reasoning from first principles when confronted with novel variants of the same problems.

For instance, correctly applying vector calculus to electromagnetic fields requires not just knowing the formulas for divergence or curl, but understanding what these operations mean for fields, sources, and boundaries in a physical system. LLMs are improving but can still falter in maintaining this contextual awareness through complex derivations.

Analysis A common strategy in physics research is to gain intuition about a complex problem by analyzing simpler, solvable special cases (e.g., zero temperature limit T → 0, one-dimensional version of a 2D problem, specific symmetry points in parameter space) or by relating it to analogous systems (e.g., mapping a quantum spin system to a classical statistical mechanics model). LLMs show some ability to follow instructions to analyze special cases if explicitly prompted. For example, given a general expression for the magnetic susceptibility χ(T), an LLM might be able to evaluate its behavior as T → 0 (e.g., Curie’s law χ ∝ 1/T for paramagnets [42]) or T → ∞. However, spontaneously identifying fruitful special cases or insightful analogies that can simplify a problem or suggest a solution path is a more advanced reasoning skill that remains underdeveloped.