SAFE-SVD：面向物理基础模型的敏感感知保真度强制奇异值分解
文章背景与核心概要
随着科学智能（AI for Science）领域的迅猛发展，物理基础模型（PFMs）受到了广泛关注。然而，传统的模型压缩方法在处理物理场景时往往失效或导致严重的精度下降，因为物理数据具有本质上的函数属性，其中的偏导数编码了对传统压缩技术高度敏感的时空动力学特征。
为了克服这一难题，本文推出了 SAFE-SVD（Sensitivity-Aware Fidelity-Enforcing SVD）框架。该方法在压缩过程中显式地对输出函数空间中的损失感知层敏感度进行建模，从而确保严格维持物理保真度和准确性。实验评估表明，SAFE-SVD 在各种模型和数据集上能够实现显著更高的压缩比（在某些情况下达到数个数量级），同时保持出色的模型精度。该方法为构建高效、可部署且可持续的科学基础模型开辟了新的子领域。
SAFE-SVD: Sensitivity-Aware Fidelity-Enforcing SVD for Physics Foundation Models
arXiv ID: arXiv:2605.17985 [cs.LG]
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
Authors: Chengjie Hong, Feixiang He, Yiheng Zeng, Lulu Kang, He Wang
Submission History: * [v1] Mon, 18 May 2026 * [v2] Thu, 13 Aug 2026 (current version)
📌 Summary
SAFE-SVD introduces a novel model compression framework specifically designed for Physics Foundation Models (PFMs)—a rapidly growing trend within the AI for Science domain.
While conventional compression methods prioritize reducing memory footprints and accelerating inference, they frequently fail or cause severe accuracy degradation in physical contexts. This occurs because physics data is inherently functional, where partial derivatives encode spatiotemporal dynamics that exhibit high sensitivity to traditional compression techniques.
To overcome this hurdle, SAFE-SVD (Sensitivity-Aware Fidelity-Enforcing SVD) explicitly models loss-aware layer sensitivity in the output function space during compression. This ensures that physical fidelity and accuracy are rigorously maintained. Empirical evaluations demonstrate that SAFE-SVD achieves substantially higher compression ratios—in some cases by orders of magnitude—while preserving exceptional model accuracy across diverse models and datasets. This methodology paves the way for a new subfield focused on efficient, deployable, and sustainable scientific foundation models.