Reward Transport: New method steers molecular generation without extra models

Researchers have introduced Reward Transport, a technique that turns the coupling step in flow matching into a powerful control interface for molecular generation. The key insight is that the pairing of noise vectors with training data points—typically a computational convenience—can instead be optimized to align a scalar noise coordinate with a molecular property (e.g., logP or QED). By using optimal transport coupling during training, Reward Transport embeds controllability directly into the learned flow field. At inference time, simply sweeping that scalar coordinate shifts the generated distribution monotonically along the desired property, requiring no oracle reward model, gradient guidance, or additional computation.

Empirical results on ZINC-250K and GuacaMol benchmarks demonstrate the method's effectiveness: varying the scalar produces monotone control of logP and consistent control of QED across its operating range. Notably, the same knob triggers opposite structural responses—growing molecules for logP but shrinking them for QED—confirming that the effect is not due to generic size bias. The authors also show that Reward Transport is complementary to classifier-free guidance and conditional flow matching, while a negative result under epsilon-prediction diffusion clarifies its limitations. This work provides a principled, continuously adjustable distribution-level control knob for generative molecular design, with implications for drug discovery and materials science.