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.
More from Research
- Google open sources SAM: Sovereign Agent Mesh framework — adnan_hashmi · 2026-08-28
- Marin 535B-A23B Training at 7%, Team to Share Design Decisions and Tradeoffs — bariskasikci · 2026-08-28
- Judea Pearl: ordinary language is pre-scientific — it helped alchemists make glass, not turn lead into gold — yudapearl · 2026-08-28
- Princeton's i1: the first fully open, reproducible text-to-image recipe, beating prior open models by 29.5 points — jiqizhixin · 2026-08-28
- Using National Curricula to Build Culturally Grounded Data for LLMs — davlanade · 2026-08-28
- 7,400 trajectories analyzed: Claude Code and Codex carry ~25k ISL vs Terminus 2's 8k — zainhas · 2026-08-28