Meta Flow Maps enable scalable reward alignment

Controlling generative models—whether via inference-time steering or fine-tuning—is expensive. Control relies on estimating the value function—typically necessitating costly trajectory simulations. To eliminate this bottleneck, we introduce *Meta Flow Maps (MFMs)*, stochastic extensions of consistency models and flow maps. MFMs are trained to perform \textbf{one-step posterior sampling}, generating arbitrarily many i.i.d. draws of clean data $x_1$ from any noisy state $x_t$. Crucially, these samples are differentiable in the conditioning state $x_t$, unlocking efficient estimation of the value function gradient. We leverage this capability to enable both **inference-time steering** without inner rollouts, and unbiased, off-policy **fine-tuning** to general rewards. Among our fine-tuning and steering experiments on ImageNet, we highlight that our single-particle steered-MFM sampler outperforms a Best-of-1000 baseline across multiple rewards at a fraction of the compute.

Generative AI systems can create images, text, or other data, but guiding them toward specific goals—such as better image quality, stronger prompt match, or other user preferences—can be very expensive. Current methods often need to repeatedly simulate many possible futures during generation, which slows them down and makes it costly to adapt models to new objectives. We introduce Meta Flow Maps, a method that learns to quickly predict many possible final outputs from an intermediate, noisy stage of generation. This lets the model estimate which direction will lead to higher-reward outputs without repeatedly running long simulations. As a result, the same trained model can be adapted at generation time or used to fine-tune a model for new rewards more efficiently. In experiments on simple distributions, MNIST, and ImageNet, Meta Flow Maps produced better reward-guided samples with far less computation than strong baselines. On ImageNet, even a low-cost version outperformed selecting the best image from 1,000 generated candidates, showing that this approach can make reward alignment for generative models much more scalable.