SimDiff: Depth pruning via similarity and difference

Yuli Chen

doi:10.1016/j.eswa.2026.133944

SimDiff: Depth pruning via similarity and difference

2026-08-10

0

PRE

AI

DOI：10.1016/j.eswa.2026.133944

Original

Original request for help

Save

Abstract

En 中文

• SimDiff prunes LLMs by jointly measuring inter-layer similarity and difference. • Two metrics, MSSD and MASD, capture distinct and complementary layer behaviors. • SimDiff preserves >91% performance at 25% pruning and speeds up inference 1.49 × .

Keywords:

Depth pruning

Model compression

Large language models

SimDiff: Depth pruning via similarity and difference

Abstract

Journal

Organization