We propose a sequential Monte Carlo estimator for expected information gain that reduces computation in Bayesian optimal experiment design using backward tempering.

This paper presents a lightweight MCMC-based approach for constrained decoding in LLMs

We benchmark feel-good thompson sampling for contextual bandits with MCMC methods and show that they perform well in the linear setting but do not perform well in neural bandit tasks.

We parallelize MCMC over the sequence, yielding more than order of magnitude wall-clock speedup on sampling.