Diamond Maps: Efficient Reward Alignment via Stochastic Flow Maps

In three linesDiamond Maps are stochastic flow map models enabling efficient reward alignment at inference time. They amortize multiple simulation steps into a single-step sampler while preserving stochasticity required for optimal alignment. Learned via distillation from GLASS Flows, they outperform existing methods in performance and scalability.

Summary generated by Claude — human-verified