CET: Spectral Geometry for Mixed-Precision Quantization
Year: 2026
Type: conference-paper
Abstract: As deep models continue to scale, post-training quantization has become an important tool for efficient deployment. Compared with uniform quantization, mixed-precision quantization offers greater compression potential by assigning different bit-widths according to the quantization tolerance of different layers. However, existing methods mainly rely on layer-wise sensitivity ranking or discrete search for bit allocation, while making limited use of the directional differences of quantization perturbations in the local loss landscape. To address this limitation, we propose Compression Error Topography (CET), which models mixed-precision quantization as structured perturbations in parameter space and relates quantization perturbations to loss variation through a local second-order approximation. CET further leverages the spectral structure of the Hessian to characterize curvature variations across perturbation directions, identify low-curvature geometric subspaces that are more compression-friendly, and use them to guide layer-wise mixed-precision allocation. Extensive experiments across vision, language, and generative models demonstrate that CET achieves competitive compression–accuracy trade-offs, with particularly strong performance in mixed-precision and ultra-low-bit regimes.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225319871