Neuro-symbolic Diffusion Models

Many neurosymbolic systems treat symbols independently, which makes it hard to capture global consistency. NeSyDMs instead use a discrete diffusion process to model the full joint distribution over symbol sets.

This lets the model express dependencies between symbols while remaining compatible with existing neural predictors. The result is improved accuracy and better-calibrated confidence on multi-label symbolic tasks.

NeSyDMs show that diffusion processes can be a strong fit for structured reasoning problems where relationships between symbols matter as much as the symbols themselves.

Paper: https://arxiv.org/abs/2505.13138