Research map / Architectures and efficiency / Attention and transformers
How Do Transformers Learn to Associate Tokens: Gradient Leading Terms Bring Mechanistic Interpretability
Shawn Im, Changdae Oh, Zhen Fang, Sharon Li
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 |
|---|---|---|---|
| keep output weight leading term (vs removing it) (ablation) | 3-layer transformer on TinyStories · TinyStories (lr 0.05) | loss: 8.287 -> 5.349 | Improved |
| keep value layer 1 leading term (vs removing it) (ablation) | 3-layer transformer on TinyStories · TinyStories (lr 0.05) | loss: 6.526 -> 5.349 | Improved |
| keep value layer 0 leading term (vs removing it) (ablation) | 3-layer transformer on TinyStories · TinyStories (lr 0.05) | loss: 6.192 -> 5.349 | Improved |
| keep attention layer 2 leading term (vs removing it) (ablation) | 3-layer transformer on TinyStories · TinyStories (lr 0.05) | loss: 5.361 -> 5.349 | No clear effect |
| keep attention layer 0 leading term (vs removing it) (ablation) | 3-layer transformer on TinyStories · TinyStories (lr 0.05) | loss: 5.350 -> 5.349 | No clear effect |
| train with BPE tokenization instead of word-level (ablation) | 3-layer transformer on TinyStories · TinyStories (BPE 10k, lr 0.005) | min cosine similarity: 0.998486 -> 0.997891 | No clear effect |
| use larger learning rate 0.05 instead of 0.005 (ablation) | 3-layer transformer on TinyStories · TinyStories | cosine similarity: 0.7 | Mixed |
| remove MLP output when computing embedding correlations (E_attn) (ablation) | 3-layer transformer on TinyStories · Pythia-1.4B (OpenWebText) | cosine similarity of covariances | No clear effect |
Closest papers by meaning
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.88)
- Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformers (NeurIPS 2024 · similarity 0.88)
- A distributional simplicity bias in the learning dynamics of transformers (NeurIPS 2024 · similarity 0.88)
- Extrapolation by Association: Length Generalization Transfer In Transformers (NeurIPS 2025 · similarity 0.88)
- Training Dynamics of Transformers to Recognize Word Co-occurrence via Gradient Flow Analysis (NeurIPS 2024 · similarity 0.88)
- Talking Heads: Understanding Inter-Layer Communication in Transformer Language Models (NeurIPS 2024 · similarity 0.88)
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.
