The Golden Ticket

Diffusion-based robot policies generate actions by denoising a random Gaussian sample. At inference time, a fresh noise vector is drawn, and the learned denoising network transforms it into a motor command. The randomness is considered essential — it provides the diversity that makes generative models generative.

Patil, Biza, and colleagues (arXiv:2603.15757) freeze the noise to a single vector and find that performance improves. On 38 of 43 manipulation tasks, a fixed “golden ticket” noise vector outperforms random sampling, with success rate gains up to 60%. No retraining. The policy weights are untouched. The only change is replacing stochastic noise with a constant.

The mechanism: the learned denoising trajectory is the policy’s actual competence. The noise was supposed to provide useful variety, but it mostly introduced failure modes — suboptimal regions of the noise space that map to poor actions through the denoising network. Collapsing the noise distribution to a single well-chosen point eliminates these failure modes while preserving the denoising network’s ability to generate high-quality actions.

Different golden tickets are optimal for different tasks, and multi-task deployment creates a Pareto frontier: you can trade off performance between tasks by selecting different frozen vectors. The noise space, which was treated as undifferentiated randomness, turns out to be a structured landscape where some points are universally good and others are task-specifically optimal.

The deeper implication: in generative models applied to control, the stochasticity that defines the model class may be counterproductive. The model learned to denoise from noise, but the best denoising happens from a specific noise, not from arbitrary noise. The randomness was never the engine of capability — it was a search strategy that included bad starting points. A single good starting point beats a distribution of mixed ones.