AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Demystifying When Pruning Works via Representation Hierarchies

Shwai He, Guoheng Sun, Haichao Zhang, Yun Fu, Ang Li

ICML 2026 · Architectures and efficiency › Pruning and efficient tuning · approach: Structured network pruning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
drop 8 attention layersMistral-7B-Instruct · multiple-choice averageaccuracy (%): 69.3 -> 69.8No clear effect
drop 8 MLP layersMistral-7B-Instruct · multiple-choice averageaccuracy (%): 69.3 -> 64.3Hurt
drop 8 attention layersMistral-7B-Instruct · generation averagescore: 22.3 -> 13.2Hurt
drop 8 MLP layersMistral-7B-Instruct · generation averagescore: 22.3 -> 0.8Hurt
drop 8 attention layersMistral-7B-Instruct · GSM8Kaccuracy (%): 48.4 -> 36.2Hurt
drop 8 attention layersE5-Mistral · BEIR retrieval averagenDCG@10: 58.9 -> 53.4Hurt
drop 8 MLP layersE5-Mistral · BEIR retrieval averagenDCG@10: 58.9 -> 56.8Hurt
drop 8 MLP layersE5-Mistral · TREC-COVIDnDCG@10: 85.9 -> 79.6Hurt
apply intra-layer pruning (Wanda 50%, 4:8, 2:4)Qwen2.5-7B, Llama3-8B, Mistral-7B · GSM8K vs HellaSwagperformance (%)Mixed
measure deviation in logit space versus embedding space after layer dropQwen-2.5-7B-Instruct · cosine similarity1 - CosineSimNo clear effect
map logits to probabilities with softmaxQwen-2.5-7B-Instruct · probability space deviationcosine similarityHurt
evaluate over generation steps rather than first stepQwen-2.5-7B-Instruct Drop-8A · representation similaritycosine similarityHurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy