WTF Algorithm Accelerates RL Finetuning 50 to 100x via Simulation Free RL

The Wasserstein-Tilted Flow Maps (WTF) algorithm enables simulation-free reinforcement learning for flow-based generative models by using a single function evaluation instead of backpropagation through rollouts. The method replaces KL regularization with a transport-based alternative and is 50 to 100 times faster than current state of the art finetuning approaches. The research, a spotlight selection for the NeurIPS 2026 conference, addresses the computational expense of adjoint methods and the difficulty of sampling reward-tilted distributions by redefining the problem using Wasserstein distance.