SAC Flow was jointly proposed in October 2025 by a research team from Tsinghua University, Carnegie Mellon University, and other universities and institutions. The relevant research results were published in the paper "Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment".

SAC Flow is an efficient and high-performance offline reinforcement learning algorithm for flow-based policy samples. It addresses the gradient instability problem when training flow-based policies by treating the flow-based model as a sequence model and reparameterizing its velocity network as a GRU or Transformer. Researchers evaluated the performance of SAC Flow in both de novo training and offline-to-online training settings, demonstrating fast convergence and achieving state-of-the-art performance on multiple motion and manipulation tasks.