Safe Particle Selection for Diffusion Steering

We consider the problem particle selection for diffusion steering, where each current particle is branched, candidate children are evaluated by stochastic rollouts, and only a subset is propagated to the next denoising step. Standard Tweedie plug-in selection is computationally attractive, but for nonlinear rewards it can underestimate high-value ambiguous particles and prune them too early. We propose \textbf{Safe Particle Selection (SPS)}, which uses the Tweedie plug-in only as a proxy while making pruning decisions through anytime-valid confidence sequences built from e-processes. This yields finite-sample no-false-prune guarantees under adaptive rollout allocation, and the guarantee composes across the diffusion trajectory. Preliminary synthetic experiments show that our method better tracks the oracle, reduces cumulative regret, and retains more genuinely high-value particles than plug-in selection.