AI / ML research · 2026
Inference-Time Reward Alignment for Diffusion & Flow Matching
Research on sparse vs. dense inference-time reward alignment; proposed ODE-DITA — improved PickScore by +0.46 with best-of-K selection.
- PyTorch
- Diffusion Models
- Flow Matching
- VLMs
- SMC
+0.46PickScore (sparse best-of-K)
The problem
Align diffusion / flow-matching generations to a reward at inference time without reward hacking or credit-assignment failure.
My role
Researcher (MS project).
What I built
Studied sparse vs. dense inference-time alignment and proposed ODE-Dense Inference-Time Alignment (ODE-DITA): deterministic n-step probability-flow ODE rollouts, Sequential Monte Carlo resampling, dense gain weighting, and a VLM semantic corrector in a hybrid variant.
Architecture & stack
- Deterministic n-step probability-flow ODE rollouts
- Sequential Monte Carlo resampling + dense gain weighting
- VLM semantic corrector (hybrid variant)
- Evaluated across 50 prompts × 3 seeds