Scalable Inference-Time Annealing with Surrogate Likelihood Estimators

Researchers introduced SITA, a scalable inference-time annealing method for molecular sampling.

SITA retrains flow-based generative models across a temperature schedule using energy-based surrogate likelihoods.

The method avoids expensive divergence computations used by prior approaches.

It achieved state-of-the-art results on Alanine Dipeptide and Alanine Tripeptide, improving low-temperature sampling for chemistry and biophysics.