Researchers propose VISTA to accelerate reward-guided discrete diffusion with task-specific guidance schedules

Researchers from an unspecified institution propose Value-Informed Schedule Trimming (VISTA), a training-free method to accelerate reward-guided discrete diffusion by automatically selecting guided timesteps under a fixed computational budget.

The method formulates sparse guidance as an approximation of the ideal full-guidance probability path and derives an estimable upper bound on ℓ₁ path deviation to guide schedule selection.

VISTA uses value statistics collected from a small number of full-guidance warmup generations to optimize the guidance schedule via dynamic programming, ensuring task-dependent schedules are discovered without additional training.

After schedule selection, VISTA maintains the same per-generation guidance complexity as other sparse schedules with the same budget, but with improved performance across tested tasks.

In experiments spanning text, biological sequence, and image generation, VISTA outperformed hand-designed sparse schedules in most settings and approached or exceeded full-guidance performance on several task metrics.

For toxic-text generation specifically, VISTA achieved a unique-toxic ratio of 0.99 compared to 1.00 for full guidance, while reducing post-selection per-generation latency from 204.23 seconds to 38.16 seconds—a 5.35× speedup.