Title: Scalable Inference-Time Annealing with Surrogate Likelihood Estimators

Abstract:

Efficiently sampling the Boltzmann distribution of molecules remains a persistent hurdle in the fields of biophysics and computational chemistry. To overcome the constraints of traditional sampling methods, researchers have turned to generative modeling, aiming to bypass the heavy computational burden associated with simulations. One particularly promising approach involves iteratively fine-tuning diffusion models across a temperature ladder, a process that relies on generating training data through importance sampling during inference-time annealing. However, these techniques face a significant bottleneck: they necessitate calculating a divergence over the score field to determine importance weights, a requirement that becomes computationally intractable as system size increases.

In this work, we introduce Scalable Inference-Time Annealing (SITA). This method re-trains flow-based models to produce samples at increasingly lower temperatures, leveraging an energy-based model to enable rapid surrogate likelihood calculations. By sidestepping the need for expensive divergence terms, SITA achieves state-of-the-art results on both Alanine Dipeptide and Alanine Tripeptide benchmarks. The source code for this approach is accessible at https://github.com/countrsignal/sita.git.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC