AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

STEM: SCALING TRANSFORMERS WITH EMBEDDING MODULES

Ranajoy Sadhukhan, Sheng Cao, Harry Dong, Changsheng Zhao, Attiano Purpura-Pontoniere, Yuandong Tian, Zechun Liu, Beidi Chen

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Block-sparse attention kernels · 13 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
replace 1/3 of FFN up-projections with STEM embeddingsMobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 49.72 -> 50.90Improved
STEM-1/3 vs Hash-MoE at same total paramsMobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 50.58 -> 50.90Improved
replace 1/2 instead of 1/3 of FFN layers (ablation)STEM MobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 50.90 -> 54.20Improved
replace all FFN layers instead of 1/2 (ablation)STEM MobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 54.20 -> 53.43Mixed
replace up-projection instead of gate-projection (ablation)STEM MobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 49.10 -> 50.90Improved
keep up-projection with additive STEM embedding (STEM-dagger) (ablation)STEM MobileLLM-350M · Zero-shot commonsense avg, 350Mavg accuracy: 50.90 -> 50.60No clear effect
apply STEM-1/3 at 1B scaleLlama3.2-1B · Zero-shot commonsense avgavg accuracy: 55.82 -> 56.63Improved
midtrain 1B STEMLlama3.2-1B · GSM8Kaccuracy: 44.2 -> 46.4Improved
midtrain 1B STEMLlama3.2-1B · MMLUaccuracy: 29.92 -> 32.38Improved
apply STEM on ARC-Challenge with 1/2 replacement (ablation)MobileLLM-350M · ARC-Challengeaccuracy: 30.55 -> 40.00Improved
apply STEM-1/3 on SIQAMobileLLM-350M · SIQAaccuracy: 41.10 -> 39.76Hurt
apply STEM at 1BLlama3.2-1B · BoolQaccuracy: 64.21 -> 61.66Hurt

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