Skip to main content
Back to timeline
arXivSource publication:

One diffusion sampler trained across a temperature ladder: Sequential Score Sampler posts the lowest mean in nine of ten target-metric comparisons

Synopsis

Repeated sampling from a distribution known only up to its normalizing constant stalls when a local Markov chain stays inside one basin. Sequential Score Sampler takes temperature as an input and trains a single shared score-based diffusion proposal, refreshing its training data through weighted replay buffers and path-space importance weights along a temperature ladder. Across five targets, corrected samples give the lowest distance and energy errors on both Gaussian mixtures; on LJ-13 the final buffer leads all three metrics, while on Alanine Dipeptide raw samples give the lowest backbone-angle error and the buffer the lowest energy errors.

Source-provided article image: Amortizing Tempered Sampling with Corrected Self-Training
arXiv

Interpretation

Repeated sampling no longer restarts a long chain each time; the computation goes into a reusable proposal model, and after training samples can be drawn directly from the model, from its importance-corrected proposals, or from the final-temperature buffer. Annealed importance sampling and sequential Monte Carlo cross multimodality through intermediate distributions but still need additional sampling computation for more accurate estimates, while local Markov chain Monte Carlo may explore one basin thoroughly and rarely move between basins. On GMM50, repeated buffer refreshing lowers the mean distance discrepancy of raw samples from 0.398 with fixed SMC training data to 0.166; across the ten target-metric comparisons in the table, an SSS sample set attains the lowest displayed mean in nine, including one tie, and the second-lowest in the remaining case.

The scheme is corrected self-training along a temperature ladder: the model generates proposals at each intermediate temperature, resamples them by importance weights, applies Markov updates that preserve that temperature's target, then reweights the result into the next temperature's buffer. Unlike progressive methods that generate training data as temperatures are visited, one temperature-conditioned model is trained jointly across all levels, and buffers retain weighted samples from recent iterations for reuse over several optimization steps. On DW-4, removing the path-space importance correction leaves both modes of the pairwise-distance distribution visible but distorts their relative masses, and the final buffer gains a long tail of high-energy configurations; removing the HMC updates leaves resampled batches of a few repeated configurations and makes training slower and less stable.

Perspective

The experiments cover two Gaussian mixtures, the four-particle double-well DW-4, the thirteen-particle Lennard-Jones cluster LJ-13, and alanine dipeptide at 300 K, compared against TA-BG, PITA and AS, with a matched-energy-budget SMC comparison; these settings define the demonstrated scope.

The method depends on the training buffers covering the target distribution: if a mode is missing from the buffers, resampling cannot recover it and short Markov chains may fail to discover it, and the next buffer inherits the gap that repeated training may reinforce. Even when all modes are represented, concentrated importance weights can give low effective sample size and many duplicates after resampling, and self-normalized importance estimates carry finite-sample bias. Initialization and repeated buffer refreshes cost computation, so the value of amortization depends on later reuse of the sampler.