Stop the Sampler! Classifier-Based Adaptive Stopping for Sampling Kernels
2026-06-15 • Machine Learning
Machine Learning
AI summaryⓘ
The authors address the problem of sampling from complicated probability distributions, which is important in statistics and machine learning. Traditional methods that explore these distributions can be slow and require fixed settings. They introduce a new method that learns when to stop exploring in a smarter way, using neural networks to recognize when a good solution is reached. This approach helps the sampler find important regions faster and cover more possibilities effectively. Their experiments show it works better than usual methods in several test cases.
Bayesian inferenceMarkov chain Monte Carlotrajectory lengthgenerative flow networksneural classifiersdetailed balancesamplingmode coveragemixing
Authors
Kirill Korolev, Nikita Morozov, Stepan Pavlenko, Esmeralda S. Whitammer, Sergey Samsonov
Abstract
Sampling from complex, unnormalized probability densities is a fundamental challenge in Bayesian inference and probabilistic modeling. While Markov chain Monte Carlo (MCMC) methods provide asymptotic guarantees, they often suffer from slow mixing and high computational costs due to fixed or manually tuned trajectory lengths. In this work, we propose a novel framework that treats trajectory termination as a learnable component of the sampling dynamics. By framing MCMC within the theory of non-acyclic generative flow networks (GFlowNets), we train state-dependent neural classifiers to decide when a trajectory has reached a high-density region and should terminate. We theoretically establish the connection between optimal classifiers and the target density via detailed balance conditions and introduce a multilevel training scheme to facilitate exploration in complex geometries. Experimental results across various benchmark densities demonstrate that our approach significantly reduces average trajectory lengths while improving mode coverage and mixing compared to standard MCMC baselines.