Learning When to Stop: Prefix-Optimal Dynamic Diffusion Policies for Continuous Control
POGP cuts diffusion-policy denoising work by about 2.7x while keeping near-full control performance.
The paper introduces Prefix-Optimal Generative Policies, which train a value function at intermediate denoising steps. That value function both improves intermediate actions during training and decides at test time when more denoising is unlikely to help. In four MuJoCo environments against 12 baselines, the method also reports about a 3.5% final-performance gain over state-of-the-art dynamic diffusion baselines. ArXiv · AI/CL/LG's note
The paper introduces Prefix-Optimal Generative Policies, which train a value function at intermediate denoising steps. That value function both improves intermediate actions during training and decides at test time when more denoising is unlikely to help. In four MuJoCo environments against 12 baselines, the method also reports about a 3.5% final-performance gain over state-of-the-art dynamic diffusion baselines. ArXiv · AI/CL/LG's note
score 5