Research map / Large language models / Language, safety and interpretability
Predicting the Emergence of Induction Heads in Language Model Pretraining
Tatsuya Aoyama, Ethan Gotlieb Wilcox, Nathan Schneider
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 |
|---|---|---|---|
| increase batch size (fixed context 1024) (ablation) | GPT2 50M · Natural data (CC100) | induction head emergence point (updates): rho=-1.0, p<0.001 | Improved |
| increase batch size (ablation) | GPT2 50M · Natural data (CC100) | final induction head strength (PS): larger batch -> lower eventual PS | Hurt |
| increase context size (fixed batch 16) (ablation) | GPT2 50M · Natural data (CC100) | induction head emergence point: rho=-1.0, p<0.001 | Improved |
| increase context size (ablation) | GPT2 50M · Natural data (CC100) | slope after emergence: rho=1.0, p<0.001 | Improved |
| increase proportion of chunks with bigram repetitions (ablation) | GPT2 50M · Natural data (CC100) | emergence point: rho=-0.64, p=0.12 | No clear effect |
| increase proportion of chunks with bigram repetitions (ablation) | GPT2 50M · Natural data (CC100) | slope: rho=0.93, p=0.002 | Improved |
| add model size as predictor of emergence updates (ablation) | log-linear law · Natural data (CC100) | emergence updates regression: -0.0027, p 0.966 | No clear effect |
| make next-token distribution independent of current token (-LD) (ablation) | synthetic Markov data · Synthetic data | induction head formation | Hurt |
| use Zipfian marginal with categoriality at near-boundary repetition (0.1-0.3) (ablation) | synthetic Markov data · Synthetic data | induction head formation | Improved |
| increase batch size beyond B=64 (context 1024) (ablation) | GPT2 50M · Natural data (CC100) | final validation loss: lower until B=64, higher beyond | Mixed |
Closest papers by meaning
- What needs to go right for an induction head? A mechanistic study of in-context learning circuits and their formation (ICML 2024 · similarity 0.90)
- The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains (NeurIPS 2024 · similarity 0.90)
- Which Attention Heads Matter for In-Context Learning? (ICML 2025 · similarity 0.89)
- KV Shifting Attention Enhances Language Modeling (ICML 2025 · similarity 0.89)
- Beyond Induction Heads: In-Context Meta Learning Induces Multi-Phase Circuit Emergence (ICML 2025 · similarity 0.88)
- The emergence of sparse attention: impact of data distribution and benefits of repetition (NeurIPS 2025 · 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.
