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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| negatively intervene (s=-1) on the French attention module | Gemma-2-2b-it · FQuAD validation | French response ratio: 85.35% -> 1.66% | Improved |
| use SAMI module intervention instead of SAE feature steering | Gemma-2-2b-it · FQuAD validation | French response ratio: 3.98% -> 1.66% | Improved |
| amplify reasoning module with s=1.4 | Llama-3.1-8B-Instruct · GSM8K | accuracy: 84.61 -> 85.44 | Improved |
| amplify reasoning module with s=1.2 | Gemma-7B-Base · GSM8K | accuracy: 54.36 -> 56.71 | Improved |
| amplify reasoning module | Llama-3.1-8B-Instruct · MT-bench | score: -0.07 | No clear effect |
| amplify reasoning module | Llama3.1-8B-Instruct / Gemma-7B-Base · Commonsense QA | accuracy: -0.08% / +0.41% | No clear effect |
| amplify reasoning module | Llama3.1-8B-Instruct / Gemma-7B-Base · MBPP+ | accuracy: -1.8% / +1.0% | No clear effect |
| amplify reasoning module | Llama3.1-8B-Instruct / Gemma-7B-Base · HumanEval+ | accuracy: +0.6% / +0.0% | No clear effect |
| negatively intervene on safety module instead of GCG attack | Llama-2-chat-7B · HarmBench | attack success rate: 34.5 -> 71.1 | Improved |
| negatively intervene on safety module instead of Ortho | Gemma-7B-it · HarmBench | attack success rate: 73.0 -> 84.3 | Improved |
| negatively intervene on safety module instead of GCG or Ortho | Qwen-chat-7B · HarmBench | attack success rate: 79.5 -> 78.0 | Hurt |
| negatively intervene on ImageNet label recognition module | ViT-B/32 · ImageNet-1k | target label accuracy: to 0% | Improved |
Closest papers by meaning
- Rethinking Attention: Exploring Shallow Feed-Forward Neural Networks as an Alternative to Attention Layers in Transformers (Student Abstract) (AAAI 2024 · similarity 0.89)
- AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers (ICML 2024 · similarity 0.89)
- Transformers Learn Faster with Semantic Focus (NeurIPS 2025 · similarity 0.88)
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · similarity 0.87)
- No Head Left Behind – Multi-Head Alignment Distillation for Transformers (AAAI 2024 · similarity 0.87)
- Beyond Components: Singular Vector-Based Interpretability of Transformer Circuits (NeurIPS 2025 · similarity 0.87)
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.
