A method for estimating the Metropolis-Hastings acceptance function directly from a learned score function, so accept/reject correction can be applied to score-based samplers that lack an explicit energy function. It introduces the Score Balance Matching objective, derived from the detailed balance condition, which learns the log-acceptance ratio from samples and the estimated score alone. This lets Metropolis-adjusted sampling be layered onto unadjusted Langevin sampling of diffusion models to improve sample quality and mode coverage.
Introduces Score Balance Matching (SBM), a loss function that lets the Metropolis-Hastings acceptance probability be estimated from a learned score function alone, without access to an unnormalized t…