AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

From Concepts to Components: Concept-Agnostic Attention Module Discovery in Transformers

Jingtong Su, Julia Kempe, Karen Ullrich

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Transformer expressivity and learnability · 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
negatively intervene (s=-1) on the French attention moduleGemma-2-2b-it · FQuAD validationFrench response ratio: 85.35% -> 1.66%Improved
use SAMI module intervention instead of SAE feature steeringGemma-2-2b-it · FQuAD validationFrench response ratio: 3.98% -> 1.66%Improved
amplify reasoning module with s=1.4Llama-3.1-8B-Instruct · GSM8Kaccuracy: 84.61 -> 85.44Improved
amplify reasoning module with s=1.2Gemma-7B-Base · GSM8Kaccuracy: 54.36 -> 56.71Improved
amplify reasoning moduleLlama-3.1-8B-Instruct · MT-benchscore: -0.07No clear effect
amplify reasoning moduleLlama3.1-8B-Instruct / Gemma-7B-Base · Commonsense QAaccuracy: -0.08% / +0.41%No clear effect
amplify reasoning moduleLlama3.1-8B-Instruct / Gemma-7B-Base · MBPP+accuracy: -1.8% / +1.0%No clear effect
amplify reasoning moduleLlama3.1-8B-Instruct / Gemma-7B-Base · HumanEval+accuracy: +0.6% / +0.0%No clear effect
negatively intervene on safety module instead of GCG attackLlama-2-chat-7B · HarmBenchattack success rate: 34.5 -> 71.1Improved
negatively intervene on safety module instead of OrthoGemma-7B-it · HarmBenchattack success rate: 73.0 -> 84.3Improved
negatively intervene on safety module instead of GCG or OrthoQwen-chat-7B · HarmBenchattack success rate: 79.5 -> 78.0Hurt
negatively intervene on ImageNet label recognition moduleViT-B/32 · ImageNet-1ktarget label accuracy: to 0%Improved

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