研究库 论文知识库

内容库 / 论文

arXiv:2609.25963 · 工程化

GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression

GeoPair: Geometry-Preserving Cross-Layer Factorization for Training-Free Transformer Compression

类型：arxiv
标识：2609.25963
链接：https://arxiv.org/abs/2609.25963
主分类：engineering
形态：method
TLDR：Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better pre
副分类：llm-infra
待LLM分类：否
来源文件：
/inbox/tom/_candidates/2026-09-24-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

攻略 Marigold V2 · 扩散 Transformer 单目深度估计新标杆 · 干货攻略

笔记精读与批判 · Marigold V2:重访用于单目深度估计的扩散 Transformer

仓库 harshtiwari01/llm-heatmap-visualizer

仓库 fz-zsl/CAPruner

笔记 3DZip · 3D VLM 空间感知特征多样性引导 token 压缩 · 轻量精读与批判

笔记 DataSpace · 短审稿（2026-08-10 09:50 · flyP 精读棒）

arXiv

2609.25963

TLDR

Transformer architectures exhibit cross-layer redundancies, yet post-training compression pipelines typically optimize layers in isolation or rely on heuristic grouping strategies that disregard layer-specific activation geometries. We introduce a principled, training-free framework that sequentially optimizes cross-layer weight pairings and shared-dictionary factorizations. Rather than forcing weights of adjacent layers to share a basis or heuristically merging activation statistics, our approach identifies structurally compatible projections and learns a shared representation that better pre

arXiv ↗ 中文全文