Definition
Plain language
A technique for sampling something that unfolds step by step by running a population of candidates forward and repeatedly culling the weak ones and cloning the strong ones.
As stated in the literature
A particle-filtering family of methods that approximates a target distribution over sequences by propagating weighted particles and periodically resampling toward high-weight ones; used in Agentic Monte Carlo to sample from the reward-tilted posterior over agent trajectories.
Also called: particle filter, particle filters, SMC
Why it matters: It makes it possible to estimate things that unfold step by step, where computing the answer directly would be intractable.
For example, to track a moving target you might run hundreds of guessed paths forward and, at each step, throw out the ones that fit the data poorly and duplicate the ones that fit well.
Heard on the show
“The tool they reach for is Sequential Monte Carlo.”Episode 119 — Beating Reinforcement Learning Without Ever Touching the Model's Weights