CDM speeds up reward-guided sampling in discrete diffusion by 50x with under 5% overhead

CatAstro_Piyush · x · 2026-08-28

A paper from KAIST, University of Michigan and NVIDIA proposes Contrastive Distribution Matching (CDM). Reward-tilted sampling in discrete diffusion typically relies on Twisted SMC, but estimating the optimal twist function requires costly Monte Carlo approximations, creating a severe inference bottleneck.

CDM learns a parameterized twist function via positive and negative samples, reducing per-step twist evaluation to a constant-time operation with less than 5% additional compute versus a single forward pass, and is up to 50x faster than Twisted SMC. Training exploits closed-form forward kernels of discrete diffusion for the gradient estimator.

Experiments show CDM consistently outperforms baselines under matched wall-clock time across toxic text generation, regulatory DNA sequence design, protein designability, and diffusion LLM alignment. Paper and code are available.

Original post →

More from Research

Research channel →