Efficient Expectation Estimation in Autoregressive Models via Potentials

This method reduces variance when estimating test functionals under autoregressive language models by exploiting next-token conditional probabilities. Using real-valued additive functions called potentials, the estimator achieves lower variance at comparable computational costs to standard sampling.