Get Started
Research questionHow can MCMC explore multimodal targets while preserving mode weights and avoiding approximation bias?Local MCMC can remain trapped in separate modes, while tempering may distort their relative probabilities. Diffusion-based intermediate paths can improve movement between modes, but estimated scores and discretization can introduce sampling bias.
Diffusion Models
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Markov Chain Monte Carlo with Diffusion PathsThe source studies diffusion-path MCMC for multimodal Bayesian posteriors using intermediate scores estimated from unnormalized targets. It supplies spectral-gap analysis, a Metropolis correction in augmented path space, error and acceptance analysis, and experiments comparing against tempering-based and unadjusted diffusion samplers.research paper · Sep 3, 2026
Related questions
How can SGMCMC hyperparameters be tuned without Metropolis-Hastings acceptance rates when computation is limited?How can unified multimodal models reduce redundant inference computation across understanding and generation without sacrificing quality?How should imbalanced multi-label data be sampled to support rare-label inference while accounting for target frequencies and label dependencies?How should multi-expert models balance diversity and prediction variance for long-tailed visual recognition?
Home
Topics
Search
Library