SMC
A particle method that progressively approximates a target distribution through importance sampling and resampling. It is widely applied to nonlinear, non-Gaussian state-space models, and its empirical distribution converges to the true target distribution as the number of particles tends to infinity.
Timeline (last 90 days)
SMC is a class of inference algorithms based on sequential importance sampling, originally used for Bayesian filtering and state-space models, and later introduced into the field of natural language generation.
During generation, SMC maintains a set of weighted candidate sequences (particles), reweighting and resampling them at each step according to the potential functions.
SMC can dynamically correct deviations mid-generation, making it particularly suitable for hard-constraint scenarios such as "certain structures must appear".
The generated potential functions only need to pass simple tests to be plugged into the SMC pipeline, and the whole process involves no retraining of the small model or additional labeled data.
The theoretical guarantee of SMC is that, under appropriate proposal kernels and resampling strategies, the empirical distribution of particles converges to the true target distribution as the number of particles goes to infinity.
The paper's framework can embed any existing negative guidance method as a proposal kernel, providing a wrapper that gives existing methods asymptotic exactness under SMC correction.
All Facts (6)
SMC is a class of inference algorithms based on sequential importance sampling, originally used for Bayesian filtering and state-space models, and later introduced into the field of natural language generation.
50%UnverifiedDuring generation, SMC maintains a set of weighted candidate sequences (particles), reweighting and resampling them at each step according to the potential functions.
50%UnverifiedSMC can dynamically correct deviations mid-generation, making it particularly suitable for hard-constraint scenarios such as "certain structures must appear".
50%UnverifiedThe generated potential functions only need to pass simple tests to be plugged into the SMC pipeline, and the whole process involves no retraining of the small model or additional labeled data.
50%UnverifiedThe theoretical guarantee of SMC is that, under appropriate proposal kernels and resampling strategies, the empirical distribution of particles converges to the true target distribution as the number of particles goes to infinity.
50%UnverifiedThe paper's framework can embed any existing negative guidance method as a proposal kernel, providing a wrapper that gives existing methods asymptotic exactness under SMC correction.
50%