AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Universal Properties of Activation Sparsity in Modern Large Language Models

Filip Szatkowski, Patryk Będkowski, Alessio Devoto, Jan Dubiński, Pasquale Minervini, Mikołaj Piórczyński, Simone Scardapane, Bartosz Wójcik

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations · 4 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
scale Gemma3 from 1B to 27BGemma3 pretrained, intermediate activations · Intermediate sparsificationcritical sparsity (%): 50.22 -> 74.12Improved
scale Gemma3 from 1B to 27BGemma3 pretrained, all inputs · All-Inputs sparsificationcritical sparsity (%): 35.15 -> 66.26Improved
scale LLaMA from 3.2-1B to 3.1-8BLLaMA pretrained · Intermediate sparsificationcritical sparsity (%): 44.44 -> 51.89Improved
scale Qwen2.5 from 0.5B to 32BQwen2.5 pretrained · All-Inputs sparsificationcritical sparsity (%): 36.66 -> 58.93Improved
scale Qwen2.5 from 3B to 7BQwen2.5 pretrained · Intermediate sparsificationcritical sparsity (%): 71.16 -> 60.98Hurt
instruction-tune Gemma3-27BGemma3-27B · Intermediate sparsificationcritical sparsity (%): 74.12 -> 84.05Improved
instruction-tune Gemma3-12BGemma3-12B · All-Inputs sparsificationcritical sparsity (%): 52.54 -> 61.45Improved
instruction-tune Gemma3-1BGemma3-1B · Intermediate sparsificationcritical sparsity (%): 50.22 -> 49.98No clear effect
instruction-tune Qwen2.5-0.5BQwen2.5-0.5B · Intermediate sparsificationcritical sparsity (%): 46.54 -> 43.92Hurt
sparsify intermediate activations instead of all FFN inputsLLaDA-8B · Average over taskscritical sparsity (%): 56.79 -> 68.13Improved
sparsify intermediate activations instead of gate in Gemma3-27BGemma3-27B pretrained · critical sparsitycritical sparsity (%): 53.03 -> 74.12Improved
use top-k or max-p instead of top-p (ablation)sparsification rule · sparsity-accuracy tradeoffcritical sparsity (%)No clear effect

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