The 8088 The 8088 ← All news
arXiv cs.LG AI Research 11h ago

Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces

★★★★★ significance 2/5

The paper introduces Score-Repellent Monte Carlo (SRMC), a new framework for efficient non-Markovian sampling in general state spaces. It uses a running average of score evaluations to reduce variance and improve sampling efficiency without the high memory costs of traditional history-dependent methods.

Why it matters Optimizing non-Markovian sampling efficiency offers a pathway to more stable and memory-efficient generative modeling in complex state spaces.
Read the original at arXiv cs.LG

Tags

#monte carlo #sampling #mcmc #stochastic approximation #variance reduction

Related coverage