Paris seminar: Monte Carlo methods bring exactness guarantees to diffusion model sampling
kastnerkyle · x · 2026-11-12
The Mostly Monte Carlo seminar (Nov 13, PariSanté Campus, Paris) features two talks:
- Tyler Farghly (Inria Paris/ENS): Monte Carlo Methods for Diffusion Model Sampling — diffusion models are usually sampled fast and approximately at scale, but a growing set of applications need exact samples or accuracy guarantees; he presents two works: adding Barker accept–reject steps to Langevin correctors (computable from the score alone), and a rejection-sampling-based approach;
- Simon Coste (LPSM, Université Paris Cité): Importance Sampling — when and how to use IS (with chemistry and combinatorics examples), how to measure accuracy locally for a specific φ and determine required sample size n, and global accuracy via Wasserstein distances, even in infinite dimensions for stochastic diffusions.
A substantive preview for anyone interested in exactness guarantees in diffusion sampling.
More from Research
- Northwestern's Jessica Hullman to lecture on validating LLM simulations of human behavior — JessicaHullman · 2026-11-13
- CoRL Poster: Extending Open-Universe Assistance Games and Dialog Goal Inference to Robots — dhadfieldmenell · 2026-11-13
- 4M-parameter topic monitor reads Gemma 3 activations, labels every token in real time — CatAstro_Piyush · 2026-11-13
- Company Agent Bench: agents know ownership (73–91%) but not people (42–54%) — philhchen · 2026-11-13
- arXiv paper shows optimal liquidation strategies are Lipschitz, enabling neural-net approximation — chaumian · 2026-11-13
- AFNO: a certifiably arbitrage-free neural operator framework for European call-price surfaces — chaumian · 2026-11-13