ConceptSMC / Sequential Monte Carlo
Sequential Monte Carlo
A class of Bayesian inference methods based on particle filtering that approximate complex distributions by maintaining and iteratively resampling a set of weighted particles. The approach has been applied to LLM inference-time computation to improve sampling quality.
Timeline (last 90 days)
Oct 30
The study proposes using sequential Monte Carlo (SMC) methods to guide a small model's dialogue generation toward grammar goals set by an expert.
Unverified50%