Nathaniel S. Woodward

arXiv2026

Fine-Tuning Small Reasoning Models for Quantum Field Theory

Despite the growing application of Large Language Models (LLMs) to theoretical physics, there is little academic exploration into how domain-specific physics reasoning ability develops while training these models. To investigate this, we perform the first academic fine-tuning study of small (7B-parameter) reasoning models dedicated specifically to theoretical physics. Because open-source verifiable training data required to train such capabilities is scarce, we developed a robust data generation pipeline that can both create synthetic problems and make existing human-authored problems suitable for model training. Selecting Quantum Field Theory (QFT) as our primary domain, we generated over 2,500 synthetic problems alongside a curated collection of human-adapted problems sourced from arXiv and standard pedagogical resources. We conduct both Reinforcement Learning (RL) and Supervised Fine-Tuning (SFT) experiments, benchmarking performance gains as well as generalization to other physics domains. We perform an extensive analysis of model chains-of-though before and after fine-tuning, to understand how reasoning errors evolve during RL and SFT. Finally, we publicly release our data pipeline, verifiable QFT training data, and $\sim$200M tokens of QFT reasoning traces.

cs.LG↗

arXiv2024

Product Manifold Machine Learning for Physics

Physical data are representations of the fundamental laws governing the Universe, hiding complex compositional structures often well captured by hierarchical graphs. Hyperbolic spaces are endowed with a non-Euclidean geometry that naturally embeds those structures. To leverage the benefits of non-Euclidean geometries in representing natural data we develop machine learning on $\mathcal P \mathcal M$ spaces, Cartesian products of constant curvature Riemannian manifolds. As a use case we consider the classification of "jets", sprays of hadrons and other subatomic particles produced by the hadronization of quarks and gluons in collider experiments. We compare the performance of $\mathcal P \mathcal M$-MLP and $\mathcal P \mathcal M$-Transformer models across several possible representations. Our experiments show that $\mathcal P \mathcal M$ representations generally perform equal or better to fully Euclidean models of similar size, with the most significant gains found for highly hierarchical jets and small models. We discover significant correlation between the degree of hierarchical structure at a per-jet level and classification performance with the $\mathcal P \mathcal M$-Transformer in top tagging benchmarks. This is a promising result highlighting a potential direction for further improving machine learning model performance through tailoring geometric representation at a per-sample level in hierarchical datasets. These results reinforce the view of geometric representation as a key parameter in maximizing both performance and efficiency of machine learning on natural data.

hep-ph↗