D-DOIT guides frozen discrete diffusion models to black-box rewards without training
新智元 · wechat · 2026-11-15
A NeurIPS 2026 paper introduces D-DOIT, a training-free sampler that adapts masked discrete diffusion models to arbitrary black-box rewards via the Doob h-transform, requiring no gradients or parameter updates. On HepG2 enhancer design it reaches 6.19 predicted activity, beating the fine-tuned DRAKES (5.61); on protein inverse folding it hits a 68.1% joint stability-structure success rate vs 63.6% for SMC and 62.9% for TDS. Key tricks: a prediction-head completion estimator (one forward pass per candidate) and finite-candidate Doob weighting, plus optional late-stage best-of-K branching to hedge irreversible unmasking decisions. Paper: arxiv.org/abs/2610.04938.
More from Research
- Pinterest swaps clustering pipeline for a fine-tuned 4B LLM, lifts email CTR 1.1% — nilukush · 2026-11-15
- Curated 10-resource path to learn LLM distributed training from scratch — East-Muffin-6472 · 2026-11-15
- Broad/MGH preprint models inherited variation effects on disease trajectories using longitudinal EHRs — anshulkundaje · 2026-11-15
- Llama 405B Base Model Natively Mimics o1's Chain-of-Thought from 128 Pages of Context — Sauers_ · 2026-11-15
- Magic claims >10x more efficient pretraining, matching DeepSeek V4 Base for ~$0.5M — oran_ge · 2026-11-15
- OpenDS-World benchmark: best of 12 frontier models solves only 39% of real-web deep research tasks — jiqizhixin · 2026-11-15