Today for AI
HOT RADAR
llmHEAT 9.6°

AI CURATED HOT STORY · 11/15/2026

NeurIPS 2026 Paper Enables Training-Free Black-Box Reward Steering of Discrete Diffusion Models

1 reports archived1 independent sourcesupdated 11/15/2026, 13:52:00
Synthesis & Latest Updates
1 Sources Cross-Validated

A NeurIPS 2026 paper introduces D-DOIT, which frames reward adaptation as a distribution transport problem and derives a Doob h-transform formulation for masked discrete diffusion, enabling a frozen model to target arbitrary black-box rewards without retraining or reward gradients. On HepG2 enhancer design, high-budget D-DOIT-Q reaches a predicted activity of 6.19, surpassing the fine-tuned DRAKES (5.61), with results also shown on protein inverse folding.

LATEST/A NeurIPS 2026 paper introduces D-DOIT, which frames reward adaptation as a distribution transport problem and derives a Doob h-transform formulation for masked discrete diffusion, enabling a frozen model to target arbitrary black-box rewards without retraining or reward gradients. On HepG2 enhancer design, high-budget D-DOIT-Q reaches a predicted activity of 6.19, surpassing the fine-tuned DRAKES (5.61), with results also shown on protein inverse folding.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first
  1. 新智元 (微信公众号)T2·72 pts

    NeurIPS 2026 Paper Enables Training-Free Black-Box Reward Steering of Discrete Diffusion Models

    • Uses the Doob h-transform to recast reward adaptation as distribution transport, aligning to black-box rewards by changing only the sampling rule
    • Achieves 6.19 predicted activity on HepG2 enhancer design versus 5.61 for the fine-tuned DRAKES baseline
    • Requires no differentiable reward, suiting black-box scorers such as Enformer predictors and structure scoring