AI Paper Atlas

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

Sparser, Faster, Lighter Transformer Language Models

Edoardo Cetin, Stefano Peluchetti, Emilio Castillo, Akira Naruse, Mana Murakami, Llion Jones

ICML 2026 · Large language models › Language, safety and interpretability · approach: Parallel and early-exit decoding · 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
add L1 regularization (2e-5) and sparse TwELL kernelsTransformer++ 0.5B · FineWeb, 0.5Bforward execution (input tokens/ms): 410 -> 480Improved
add L1 regularization and sparse kernelsTransformer++ 1B · FineWeb, 1Bforward execution (input tokens/ms): 185 -> 219Improved
add L1 regularization and sparse kernelsTransformer++ 1.5B · FineWeb, 1.5Bforward execution (input tokens/ms): 119 -> 141Improved
add L1 regularization and sparse kernelsTransformer++ 2B · FineWeb, 2Bforward execution (input tokens/ms): 87.8 -> 106Improved
add L1 regularization and sparse kernelsTransformer++ 2B · FineWeb, 2Benergy per token (mJ): 7.85 -> 6.51Improved
add L1 regularization and sparse kernelsTransformer++ 0.5B · FineWeb, 0.5Btraining step (input tokens/ms): 97.3 -> 95.9No clear effect
add L1 regularization and sparse kernelsTransformer++ 2B · FineWeb, 2Btraining step (input tokens/ms): 22.4 -> 27.3Improved
add L1 regularization and sparse kernelsTransformer++ 1.5B · FineWeb, 1.5Bpeak memory (GB): 62.8 -> 45.1Improved
add L1 regularization and sparse kernelsTransformer++ 1B · FineWeb, 1Bmean task accuracy (%): 44.6% -> 44.7%No clear effect
add L1 regularization and sparse kernelsTransformer++ 2B · FineWeb, 2Bmean task accuracy (%): 49.1% -> 48.8%No clear effect
raise L1 coefficient up to 3e-5 (ablation)Transformer++ 1.5B · FineWeb, 1.5Bdownstream task performanceNo clear effect
raise L1 so fewer than 0.5% of neurons are active (ablation)Transformer++ 1.5B · FineWeb, 1.5Bdownstream task performanceHurt

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