Saved in:
Bibliographic Details
Main Authors: He, Shwai, Sun, Guoheng, Zhang, Haichao, Fu, Yun, Li, Ang
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2603.24652
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914556366815232
author He, Shwai
Sun, Guoheng
Zhang, Haichao
Fu, Yun
Li, Ang
author_facet He, Shwai
Sun, Guoheng
Zhang, Haichao
Fu, Yun
Li, Ang
contents Network pruning, which removes less important parameters or architectures, is often expected to improve efficiency while preserving performance. However, this expectation does not consistently hold across language tasks: pruned models can perform well on non-generative tasks but frequently fail in generative settings. To understand this discrepancy, we analyze network pruning from a representation-hierarchy perspective, decomposing the internal computation of language models into three sequential spaces: embedding (hidden representations), logit (pre-softmax outputs), and probability (post-softmax distributions). We find that representations in the embedding and logit spaces are largely robust to pruning-induced perturbations. However, the nonlinear transformation from logits to probabilities amplifies these deviations, which accumulate across time steps and lead to substantial degradation during generation. In contrast, the stability of the categorical-token probability subspace, together with the robustness of the embedding space, supports the effectiveness of pruning for non-generative tasks such as retrieval and multiple-choice selection. Our analysis disentangles the effects of pruning across tasks and provides practical guidance for its application. Code is available at https://github.com/CASE-Lab-UMD/Pruning-on-Representations
format Preprint
id arxiv_https___arxiv_org_abs_2603_24652
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Demystifying When Pruning Works via Representation Hierarchies
He, Shwai
Sun, Guoheng
Zhang, Haichao
Fu, Yun
Li, Ang
Computation and Language
Machine Learning
Network pruning, which removes less important parameters or architectures, is often expected to improve efficiency while preserving performance. However, this expectation does not consistently hold across language tasks: pruned models can perform well on non-generative tasks but frequently fail in generative settings. To understand this discrepancy, we analyze network pruning from a representation-hierarchy perspective, decomposing the internal computation of language models into three sequential spaces: embedding (hidden representations), logit (pre-softmax outputs), and probability (post-softmax distributions). We find that representations in the embedding and logit spaces are largely robust to pruning-induced perturbations. However, the nonlinear transformation from logits to probabilities amplifies these deviations, which accumulate across time steps and lead to substantial degradation during generation. In contrast, the stability of the categorical-token probability subspace, together with the robustness of the embedding space, supports the effectiveness of pruning for non-generative tasks such as retrieval and multiple-choice selection. Our analysis disentangles the effects of pruning across tasks and provides practical guidance for its application. Code is available at https://github.com/CASE-Lab-UMD/Pruning-on-Representations
title Demystifying When Pruning Works via Representation Hierarchies
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2603.24652