2606.19993

Total: 1

#1 Activation- and Influence-Aware Ranks (AIR): Function-Preserving SVD Compression for LLMs [PDF] [Copy] [Kimi] [REL]

Authors: Nico Harder, Daniel Becking, Karsten Mueller, Wojciech Samek

We present Activation- and Influence-Aware Ranks (AIR), an SVD-based LLM compression framework that guides each weight matrix's low-rank approximation with a backward-signal influence metric. Starting from the activation-aware optimum of SVD-LLM(W), AIR runs a single closed-form alternating least squares (ALS) sweep that integrates influence element-wise under a monotone-descent guarantee. AIR is layer-local and composes orthogonally with end-to-end methods: alone it exceeds ACIP, and AIR+LoRA outperforms it further. AIR improves perplexity over SVD-LLM(W) by >18% at <=60% parameter retention, matches its quality with ~90% less calibration data, and turns parameter savings into FLOP, peak-memory, and per-token latency gains.

Subject: Machine Learning

Publish: 2026-06-18 09:31:31 UTC