VISTA Trims Diffusion Guidance Schedules for Faster Reward Alignment

Summary

Reward-guided discrete diffusion can align generated outputs with downstream objectives, but applying Sequential Monte Carlo guidance at every denoising step requires additional reward evaluations and, for gradient-based proposals, value-gradient calculations. The paper introduces VISTA, a training-free method that chooses which timesteps to guide under a fixed computational budget. It treats sparse guidance as an approximation of the ideal full-guidance probability path and derives an estimable upper bound on the resulting L1 path deviation. VISTA gathers value statistics from a small number of full-guidance warmup generations, then uses dynamic programming to optimize an empirical schedule objective exactly. The warmup outputs can be retained, and after schedule selection each generation has the same guidance complexity as other sparse schedules with the same budget. Across text, biological sequence, and image generation tasks, VISTA finds schedules that vary by task and generally outperforms hand-designed sparse schedules. On several metrics it approaches or exceeds full guidance. In toxic-text generation, its unique-toxic ratio is 0.99 versus 1.00 for full guidance, while post-selection per-generation latency falls from 204.23 seconds to 38.16 seconds, a reported 5.35x speedup.