Back to feed
arXiv cs.LG
arXiv cs.LG
7/13/2026
Reward Transport: Property Control in Flow Matching via Noise-Space Alignment

Reward Transport: Property Control in Flow Matching via Noise-Space Alignment

Short summary

Reward Transport introduces a method to control molecular properties in flow matching by aligning noise vectors with data points via optimal transport coupling. A single scalar coordinate at inference steers generated distributions toward target properties like logP and QED without needing oracle models or gradient guidance. Experiments on ZINC-250K and GuacaMol show monotone property control, with the knob producing opposite structural responses for different targets, ruling out generic size bias.

  • •Flow matching coupling repurposed as alignment interface for molecular property control
  • •Single scalar coordinate steers generation at inference with no extra computation
  • •Validated on ZINC-250K and GuacaMol with monotone logP and consistent QED control

Generated with AI, which can make mistakes.

#research-breakthrough#open-source
Read full article at arXiv cs.LG

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Mind the Residual Gap: Probabilistic Downscaling under Real-World Bias
arXiv cs.LG
TMPO: Trajectory Matching Policy Optimization for Diverse and Efficient Diffusion Alignment
arXiv cs.LG
KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling
arXiv CS.AI
The user wants me to rewrite a news headline for a mobile feed based on the given rules.
arXiv CS.AI