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

Computer Science > Artificial Intelligence arXiv:2610.20669 (cs) [Submitted on 14 Oct 2026] Title: Value-Informed Schedule Trimming for Accelerated Sequential Monte Carlo Guidance in Discrete Diffusion Authors: Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim.

计算机科学 > 人工智能 arXiv:2610.20669 (cs) [提交于 2026 年 10 月 14 日] 标题：离散扩散中用于加速序列蒙特卡洛引导的价值感知调度修剪，作者：Jaehyeon Kim, Ziseok Lee, Siyoung Choi, Kyungsu Kim。

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.

摘要：奖励引导的离散扩散实现了测试时与下游目标的对齐，但在每个去噪步骤应用序列蒙特卡洛（SMC）引导的计算成本很高。主动引导步骤需要额外的奖励评估，对于基于梯度的提议，还需要进行价值梯度计算。稀疏引导通过仅引导时间步的一个子集来降低这种成本，但常用的调度是固定的启发式方法，其有效性在不同任务间差异巨大。

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 (\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.

我们提出了 VISTA，这是一种无需训练的方法，可以在固定预算下自动选择引导时间步。VISTA 将稀疏引导表述为对理想全引导概率路径的近似，并推导出了 (\ell_1) 路径偏差的可估计上限。通过使用从少量全引导预热生成中收集的价值统计数据，VISTA 通过动态规划精确优化了所得的经验代理。

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. 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.

预热生成的结果可以作为模型输出保留，在调度选择完成后，VISTA 与具有相同预算的其他稀疏调度具有相同的单次生成引导复杂度。在文本、生物序列和图像生成任务中，VISTA 发现了依赖于任务的调度，并在大多数测试设置中优于手工设计的稀疏调度，同时在多个任务指标上接近或超过了全引导性能。

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(\times) speedup.

例如，在有毒文本生成中，VISTA 达到了 0.99 的唯一有毒比率（全引导为 1.00），同时将选择后的单次生成延迟从 204.23 秒降低至 38.16 秒，实现了 5.35 倍的加速。