We explain how discrete phase transitions govern the emergence of structure in diffusion models through the idea of cross-fluctuations from statistical physics.

We propose a new guidance method for flow based models for better sampling and generation quality.

We introduce Resample-Previous-Tokens (RPT), a sampling method that allows models to revisit and replace previously generated tokens, leading to ~10% relative improvements in reasoning and coding after a short fine-tuning.

We present an algorithm that learns the rate matrix of a continuous-time Markov chian to satisfy the Kolmogorov equation for sampling from unnormalised discrete distributions.

we train transferable normalizing flows to sample from peptide Boltzmann distributions up to 8 residues

We demonstrate the stability of Langevin diffusion and use it to derive the first proof of convergence for Proximal Stochastic Gradient Langevin Algorithm in a non-convex setting.

We propose an adaptive multi-token unmasking sampler for masked language diffusion models, gaining speed up of 2-3x on code generation and math reasoning benchmarks without loss in accuracy.

We propose a novel guidance mechanism that significantly improves image quality, consistency and diversity.

We introduce a technique for coupling probability distributions when several samples are available from one of the distributions, and give applications to multi-draft speculative decoding and distributed lossy compression with side information.

TPP-SD is a fast and distributionally consistent sampling framework for Transformer-based temporal point processes, achieving 2–6× speedup over autoregressive methods by adapting speculative decoding with a lightweight draft model.

We develop trust region methods for stochastic optimal control to improve sampling from unnormalized densities, transition path sampling, and diffusion model finetuning.

We develop a novel speculative decoding framework for protein generation, using structure aware guidance from k-mers to generate proteins with higher likelihood and structural confidence.