Score-Repellent Monte Carlo: Toward Efficient Non-Markovian Sampler with Constant Memory in General State Spaces
Researchers have introduced Score-Repellent Monte Carlo (SRMC), a novel framework designed to enhance the efficiency of non-Markovian sampling in general state spaces. SRMC summarizes trajectory history using a running average of score evaluations, enabling it to discourage redundant revisits and reduce Monte Carlo variance. This approach maintains constant memory usage and offers improved estimator variance and mode coverage compared to existing methods, as demonstrated in experiments on continuous targets and discrete energy-based models. AI
IMPACT Introduces a new sampling technique that could improve efficiency and reduce variance in AI model training and inference.