UIUC与哈佛大学的研究人员提出了Explorative Modeling（XM）新范式，通过在训练时生成K个候选并选择最接近真实数据的一个来解决生成模型的mode模糊问题[1]。这一方法的核心机制极为简洁，本质上是一个3到5行的for循环——每个训练步生成K个候选，随后选择离真实数据最近的一个[1]。
研究团队将探索作为生成模型的第三个缩放轴进行验证，在图像、视频、语言等多个模态上实现了单调性能提升和显著的效率收益[1]。随着数据规模增长，探索带来的增益从7%提升到36%；随着模型规模增大，增益从13%提升到23%[1]。效率方面，FLOP效率提升4.1倍、样本效率提升6.2倍、参数效率提升47%[1]。在ImageNet无引导生成任务上，该方法达到了1.43的FID指标，逼近业界最好水平[1]。
论文基于Mode Forcing理论进行设计，研究团队表示大部分结果是先被理论预测、再被实验验证[1]。第一作者Alexi Gladstone曾在2025年7月发表Energy-Based Transformers论文[1]。
Researchers from the University of Illinois Urbana-Champaign (UIUC) and Harvard University have introduced Explorative Modeling (XM), a new paradigm designed to address mode ambiguity in generative models through a surprisingly simple mechanism [1]. The core innovation is a 3-to-5 line for loop that generates K candidate outputs during each training step and selects the one closest to the true data [1]. This approach treats exploration as a third scaling axis alongside data and model size, demonstrating that generative model expressiveness can scale independently [1].
The method delivers consistent performance improvements across multiple modalities including images, video, and language [1]. Performance gains from exploration increase monotonically with scale—rising from 7% to 36% as data grows larger, and from 13% to 23% as models scale up [1]. The efficiency improvements are substantial: FLOP efficiency improves by 4.1 times, sample efficiency by 6.2 times, and parameter efficiency by 47% [1]. The approach achieved an ImageNet unconditional FID score of 1.43, approaching the best results in the field [1].
The research builds upon Mode Forcing theory, with most results first predicted by theory and subsequently validated through experiments [1]. The research team includes Alexi Gladstone, Yilun Du, and Heng Ji [1].
