sampling-efficiency

FlowBudget 提出一种免训练的采样预算估计方法，通过仿真自由的插值探针估计求解器相对的采样复杂度，从而在不进行 FID-vs-NFE 扫描的情况下确定流匹配模型所需的函数求值次数（NFE）。

论文提出 Rank-Aware Speculative Sampling（RASS），一种基于 rank-aware list coupling 的扩散草稿树验证规则，通过对候选按 proposal-target 均值位移排序并优化秩权重以最小化总变差，在保持精确采样的同时加速扩散生成，在 CIFAR-10 上相比 D-GRS 节省约 20% 目标模型评估。

A Tsinghua University paper reveals that reinforcement learning in large language models does not enhance reasoning but improves sampling efficiency, narrowing the model's reasoning pathways as shown in benchmarks.

OpenAI presents sCM (simplified continuous-time consistency models), a new approach that scales consistency models to 1.5B parameters and achieves ~50x speedup over diffusion models by generating high-quality samples in just 2 steps. The method demonstrates comparable sample quality to state-of-the-art diffusion models while using less than 10% of the effective sampling compute.