VISTA DP Schedule Trims Discrete Diffusion SMC to 5.35x Latency at 0.99 Toxic Ratio

The arXiv 2610.20669 paper introduces VISTA, a training-free method that collects value statistics from a small set of full-guidance warmup trajectories then optimizes guided timestep selection under a fixed budget. On toxic-text, biological sequence, and image tasks the method recovers task-dependent schedules that outperform fixed heuristics and match or exceed full-guidance metrics in most settings while retaining identical post-selection complexity.

Empirical results show 204.23 s reduced to 38.16 s latency with unique-toxic ratio falling only from 1.00 to 0.99; similar patterns hold across other domains where value variance concentrates early or late in the denoising path. The approach directly addresses the variance in effectiveness of prior hand-designed sparse schedules documented in earlier SMC guidance work.

Operationally VISTA enables deployment of reward-guided discrete diffusion under tighter compute envelopes without retraining, provided warmup generations can be retained as usable outputs. The method’s reliance on observable value statistics rather than model internals supports immediate integration into existing sampling pipelines that already expose reward or value functions.

Next steps include extension to continuous diffusion and online schedule adaptation once value estimators become available mid-generation.