探索性建模：从K个候选中训练最优输出

研究人员提出了探索性建模这一新型生成模型范式，通过在训练阶段探索K个可能的输出结果，仅对其中最优匹配进行梯度更新，从而解决生成模型在多模态分布下的平均化问题1。该方法实现了数据效率提升6.2倍、计算效率提升4.1倍、参数效率提升47%1。

在ImageNet 256数据集上，探索性建模达到1.43 FID的性能指标1。随着数据规模增长，性能提升从7%增至36%；随参数规模增长，性能提升从13%增至23%1。在行为克隆任务中，探索性策略仅需1次前向传播便能匹配100步扩散策略的效果1。在世界模型任务上，该方法实现了16-256倍的前向传播次数减少1。

该研究定义了生成表达性的概念，即模型能够捕获的不同答案数量1。这一新范式使生成模型能够实现端到端的训练和推理1。

Researchers Alexi Gladstone, Yilun Du, and Heng Ji have introduced explorative modeling, a novel training paradigm for generative models that addresses the averaging problem inherent in multimodal distributions 1. The approach explores K possible outputs during training and applies gradient updates only to the best-matching candidate, enabling end-to-end training and inference 1.

The method demonstrates substantial performance improvements across multiple dimensions. On ImageNet 256, explorative modeling achieves a Fréchet Inception Distance (FID) of 1.43, corresponding to a 6.2-fold increase in sample efficiency and a 4.1-fold improvement in computational efficiency, alongside a 47% reduction in parameter requirements 1. Performance gains scale with data availability, rising from 7% improvement with smaller datasets to 36% with larger datasets, while gains from parameter scaling range from 13% to 23% 1.

The framework shows particular promise in specialized applications. In behavior cloning tasks, the explorative strategy matches the performance of a 100-step diffusion policy using only a single forward pass 1. For world models, the approach achieves forward pass reductions ranging from 16-fold to 256-fold 1. The researchers define generative expressivity as the quantity of distinct answers a model can capture, a metric central to evaluating the method's effectiveness 1.