Reward‑guided discrete diffusion enables test‑time alignment with downstream objectives, yet applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require extra reward evaluations and, for gradient‑based proposals, value‑gradient computations. Sparse guidance reduces this cost by guiding only a subset of timesteps, but commonly used schedules are fixed heuristics whose effectiveness varies widely across tasks.
We introduce VISTA, a training‑free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as an approximation of the ideal full‑guidance probability path and derives an estimable upper bound on the $$\ell_1$$ path deviation. By collecting value statistics from a small number of full‑guidance warm‑up generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warm‑up generations can be retained as model outputs, and after schedule selection VISTA has the same per‑generation guidance complexity as other sparse schedules with the same budget.
Across text, biological sequence, and image generation tasks, VISTA discovers task‑dependent schedules and outperforms hand‑designed sparse schedules in most settings, while approaching or exceeding full‑guidance performance on several metrics. For toxic‑text generation, VISTA attains a 0.99 unique‑toxic ratio (full guidance 1.00) and reduces post‑selection per‑generation latency from 204.23 s to 38.16 s, a 5.35× speedup.
Review