Combinatorial Reasoning: Using physics-inspired methods to improve reasoning on generative language models

Recent advances in large language models (LLMs) have shown strong performance in text generation but limited reliability in explainable reasoning. We explore a framework termed Combinatorial Reasoning (CR), which integrates physics-inspired combinatorial optimization into the reasoning process of generative models. In this approach, intermediate reasoning candidates produced by an LLM are mapped onto a discrete optimization problem, equivalent to finding the ground state of an Ising Hamiltonian, and solved using methods motivated by statistical physics. The resulting optimized subset of reasoning paths is then used to guide the model's final output. Building on the preliminary results reported in [1], we discuss extensions to the CR pipeline that improve accuracy and interpretability, and analyze how such hybrid architectures may bridge linguistic and statistical representations of reasoning. We further outline the prospects of executing CR on specialized hardware, such as annealers and Ising machines, to accelerate discrete reasoning tasks for generative AI. These results suggest that optimization principles common in physical systems can provide a useful abstraction for understanding and improving reasoning in modern language models.