arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2602.07928cs.LGcs.AI
收藏

流匹配的动能视角

A Kinetic Energy Perspective of Flow Matching

KTH Royal Institute of Technology(皇家理工学院);Nordita, Nordic Institute for Theoretical Physics(北欧理论物理研究所);Hasso Plattner Institute, University of Potsdam(波茨坦大学哈索 Plattner 研究院);Trinity College Dublin(都柏林三一学院);Hangzhou Institute of Technology, Xidian University(西安电子科技大学杭州研究院);The Hong Kong University of Science(香港科学大学);Mohamed bin Zayed University of Artificial Intelligence(莫莫丁·本·扎耶德人工智能大学)

Ziyun Li, Huancheng Hu, Soon Hoe Lim, Xuyu Li, Fei Gao, Enmao Diao, Zezhen Ding, Michalis Vazirgiannis, Henrik Bostrom

AI总结:

本文引入动能路径能量(KPE)作为流匹配生成模型的诊断工具,发现其与语义保真度和数据稀疏性相关,并基于此提出无训练的动能轨迹塑形(KTS)策略以改善生成质量。

摘要:

基于流的生成模型可以通过物理视角来审视:采样通过积分学习到的速度场将粒子从噪声传输到数据,每个样本对应一条具有自身动力学努力的轨迹。受经典力学启发,我们引入了动能路径能量(KPE),这是一种类似作用量的每样本诊断指标,用于测量沿常微分方程(ODE)轨迹累积的动能努力。实验上,KPE表现出两种稳健的对应关系:{i} 较高的KPE预测更强的语义保真度;{ii} 高KPE轨迹落在稀疏表示区域。我们进一步提供了将轨迹能量与数据稀疏性联系起来的理论保证。矛盾的是,这种相关性是非单调的。在足够高的能量下,生成可能退化为记忆。利用经验流匹配的闭式公式,我们表明极端能量驱动轨迹接近训练样本的副本。这产生了金发姑娘原则,并激发了动能轨迹塑形(KTS),一种无训练的两阶段推理策略,该策略增强早期运动并强制执行后期软着陆,从而减少记忆并提高基准任务上的生成质量。

英文摘要:

Flow-based generative models can be viewed through a physics lens: sampling transports a particle from noise to data by integrating a learned velocity field, and each sample corresponds to a trajectory with its own dynamical effort. Motivated by classical mechanics, we introduce Kinetic Path Energy (KPE), an action-like, per-sample diagnostic that measures the accumulated kinetic effort along an ordinary differential equation (ODE) trajectory. Empirically, KPE exhibits two robust correspondences: {i} higher KPE predicts stronger semantic fidelity; {ii} high-KPE trajectories land in sparse representation regions. We further provide theoretical guarantees linking trajectory energy to data sparsity. Paradoxically, this correlation is non-monotonic. At sufficiently high energy, generation can degenerate into memorization. Leveraging the closed-form formula of empirical flow matching, we show that extreme energies drive trajectories toward near-copies of training examples. This yields a Goldilocks principle and motivates Kinetic Trajectory Shaping (KTS), a training-free two-phase inference strategy that boosts early motion and enforces a late-time soft landing, reducing memorization and improving generation quality across benchmark tasks.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑