Sequential Monte Carlo
An inference method based on particle filtering that approximates a posterior distribution by maintaining a weighted set of particles and iteratively resampling them. In diffusion language model decoding it is used to guide the search toward high-reward candidate sequences.
Timeline (last 90 days)
The paper uses sequential Monte Carlo (SMC) to correct the error accumulation caused by particles deviating from the original unconditional marginal distribution after guidance, so that the sampler converges to the true target density as the number of particles goes to infinity.
Sequential Monte Carlo is a class of particle methods that progressively approximate a target distribution through importance sampling and resampling, and is widely used in nonlinear non-Gaussian state-space models.
All Facts (2)
The paper uses sequential Monte Carlo (SMC) to correct the error accumulation caused by particles deviating from the original unconditional marginal distribution after guidance, so that the sampler converges to the true target density as the number of particles goes to infinity.
50%UnverifiedSequential Monte Carlo is a class of particle methods that progressively approximate a target distribution through importance sampling and resampling, and is widely used in nonlinear non-Gaussian state-space models.
50%