Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion

Authors: Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim

Subjects: Artificial Intelligence (cs.AI); Machine Learning (cs.LG)

arXiv: arXiv:2610.20669 [cs.AI] — Submitted 14 Oct 2026

DOI: https://doi.org/10.48550/arXiv.2610.20669

Abstract

Reward-guided discrete diffusion enables test-time alignment with downstream objectives, but applying Sequential Monte Carlo (SMC) guidance at every denoising step is computationally expensive. Active guidance steps require additional 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 can vary substantially across tasks.

We propose VISTA, a training-free method that automatically selects guided timesteps under a fixed budget. VISTA formulates sparse guidance as approximation of the ideal full-guidance probability path and derives an estimable upper bound on the \(\ell_1\) path deviation. Using value statistics collected from a small number of full-guidance warmup generations, VISTA optimizes the resulting empirical surrogate exactly via dynamic programming. The warmup 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.

Results

Across text, biological sequence, and image generation tasks, VISTA discovers task-dependent schedules and outperforms hand-designed sparse schedules in most tested settings, while approaching or exceeding full-guidance performance on several task metrics. In toxic-text generation, for example, VISTA attains a 0.99 unique-toxic ratio compared with 1.00 for full guidance, while reducing post-selection per-generation latency from 204.23 s to 38.16 s — a 5.35× speedup.

Key Contributions

Training-free schedule selection: VISTA automatically identifies which denoising timesteps should receive SMC guidance, removing reliance on hand-tuned heuristics.
Principled objective: Sparse guidance is framed as an approximation of the ideal full-guidance probability path, with a tractable upper bound on \(\ell_1\) deviation guiding schedule optimization.
Exact optimization via dynamic programming: Given value statistics from a small number of full-guidance warmup generations, the surrogate objective is solved exactly rather than approximated greedily.
No added inference overhead: Once the schedule is selected, per-generation guidance complexity matches other sparse schedules under the same budget, and warmup generations are retained as valid model outputs.
Broad empirical gains: Consistent improvements over hand-designed sparse schedules across text, biological sequence, and image domains.

Why It Matters (2026 Context)

As discrete diffusion models mature in 2026, reward-guided generation has become a standard mechanism for test-time alignment in domains ranging from protein and DNA sequence design to controllable text and image synthesis. However, the compute cost of Sequential Monte Carlo guidance at every denoising step remains a practical bottleneck for deployment, particularly for gradient-based proposals that require value-gradient evaluations. VISTA addresses this bottleneck directly: it keeps the benefits of full-guidance alignment while cutting the number of actively guided steps, yielding order-of-magnitude latency improvements (e.g., 204.23 s → 38.16 s) without sacrificing alignment quality on the tasks tested. Because VISTA is training-free and budget-aware, it integrates cleanly with existing discrete diffusion pipelines and offers a general recipe for efficient test-time alignment as model scales continue to grow.

Citation

@article{kim2026vista,
  title={Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion},
  author={Kim, Jaehyeon and Lee, Ziseok and Choi, Siyoung and Kim, Kyungsu},
  journal={arXiv preprint arXiv:2610.20669},
  year={2026}
}

via ArXiv AI

Value-Informed Schedule Trimming for Accelerated Sequential