Leveraging Ising-sampling at test-time compute to increase reasoning in LLMs

Many recent developments in the field of LLMs have attempted to leverage test-time-compute to better tackle reasoning tasks. We introduce an optimization-based approach to leveraging test-time-compute, which creates a prompt by using the Ising variables from the ground state of an all-to-all connected Ising problem. This is done by mapping reasons to Ising variables and constructing a Hamiltonian using 1-point and 2-point correlations in the LLM. We hypothesize that known properties of the Gibbs distribution generated by combinatorial samplers are beneficial in obtaining sets of reasons which are conducive to achieving better prompt engineering outcomes. Preliminary results suggest that using combinatorial samplers in AI pipelines is a promising area for future research in the application of quantum and combinatorial optimization techniques.