GLASS Flows and Diamond Maps: Efficient Alignment via Stochastic Flows and Flow Maps

A major bottleneck in post-training and inference-time alignment of generative models is computational efficiency. In this presentation, I will detail recent advances in the sampling and training of flow and diffusion models that unlock significant efficiency gains for alignment: (1) GLASS Flows: First, I will introduce GLASS Flows, a novel sampling paradigm for diffusion models that serves as a powerful alternative to existing SDE and ODE sampling. GLASS Flows combine the best of both worlds: they are stochastic, enabling robust exploration of the search space, yet fundamentally grounded in ODEs to maintain low discretization error and high efficiency. By effectively rendering standard SDE/DDPM sampling obsolete, GLASS Flows unlock zero-cost performance gains for search, Sequential Monte Carlo, and RL fine-tuning. (2) Diamond Maps: Second, I will present Diamond Maps, demonstrating how flow maps and consistency models can be leveraged to make reward guidance both statistically accurate and highly efficient. I will outline two distinct approaches to achieve this: one utilizing GLASS Flows distillation, and another introducing a lightweight inference-time modification applicable to existing large-scale distilled models.