Latent Policy Steering through One-Step Flow Policies

Latent Policy Steering through One-Step Flow Policies explores Latent Policy Steering (LPS) enables high-fidelity latent policy improvement by backpropagating original-action-space Q-gradients through a differentiable one-step MeanFlow policy to update a latent-action-space actor, achieving state-of-the-art performance in offline RL.. Commercial viability score: 7/10 in Offline Reinforcement Learning.