Saved in:
Bibliographic Details
Main Authors: Lu, Ying, Zhou, Peng-Fei, Fang, Qi-Xuan, Zhang, Pan, Ran, Shi-Ju, Su, Gang
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2605.25344
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913161167241216
author Lu, Ying
Zhou, Peng-Fei
Fang, Qi-Xuan
Zhang, Pan
Ran, Shi-Ju
Su, Gang
author_facet Lu, Ying
Zhou, Peng-Fei
Fang, Qi-Xuan
Zhang, Pan
Ran, Shi-Ju
Su, Gang
contents Large language models (LLMs) are dominated by dense linear transformations, whose storage, memory and computational overheads hinder efficient adaptation and deployment while masking the functional impacts of structural simplification. Here we present Tensor Mixture (MixT), a general tensor-structured compression scheme that replaces targeted dense linear layers with natively executable mixtures of tensor operators. Operating directly on generic linear projections instead of model-specific components, MixT is potentially applicable across Transformer-based LLMs and other dense neural mappings. We evaluate MixT on Qwen3-8B and LLaMA2-7B under a unified recovery protocol, identifying a broad compressible regime in which MMLU accuracy is largely preserved before an abrupt transition at model-specific boundaries. This transition coincides with coordinated shifts in output entropy, prediction entropy and inter-layer geometry. At the LLaMA2-7B transition boundary, MixT reduces full-model parameters by 47.5\%, inference FLOPs by 37.1\%, training FLOPs by 52.1\% and peak inference memory by 60.4\%, demonstrating its practical potential for lower-cost LLM compression.
format Preprint
id arxiv_https___arxiv_org_abs_2605_25344
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A general tensor-structured compression scheme for efficient large language models
Lu, Ying
Zhou, Peng-Fei
Fang, Qi-Xuan
Zhang, Pan
Ran, Shi-Ju
Su, Gang
Computation and Language
Artificial Intelligence
Machine Learning
Quantum Physics
Large language models (LLMs) are dominated by dense linear transformations, whose storage, memory and computational overheads hinder efficient adaptation and deployment while masking the functional impacts of structural simplification. Here we present Tensor Mixture (MixT), a general tensor-structured compression scheme that replaces targeted dense linear layers with natively executable mixtures of tensor operators. Operating directly on generic linear projections instead of model-specific components, MixT is potentially applicable across Transformer-based LLMs and other dense neural mappings. We evaluate MixT on Qwen3-8B and LLaMA2-7B under a unified recovery protocol, identifying a broad compressible regime in which MMLU accuracy is largely preserved before an abrupt transition at model-specific boundaries. This transition coincides with coordinated shifts in output entropy, prediction entropy and inter-layer geometry. At the LLaMA2-7B transition boundary, MixT reduces full-model parameters by 47.5\%, inference FLOPs by 37.1\%, training FLOPs by 52.1\% and peak inference memory by 60.4\%, demonstrating its practical potential for lower-cost LLM compression.
title A general tensor-structured compression scheme for efficient large language models
topic Computation and Language
Artificial Intelligence
Machine Learning
Quantum Physics
url https://arxiv.org/abs/2605.25344