Research map / Architectures and efficiency / Attention and transformers
Birth of a Transformer: A Memory Viewpoint
Alberto Bietti, Vivien Cabannes, Diane Bouchacourt, Herve Jegou, Leon Bottou
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 |
|---|---|---|---|
| Use two transformer layers instead of one | Vanilla transformer · Synthetic bigram data, fixed triggers | accuracy: around 55% -> over 99% | Improved |
| Use two transformer layers instead of one | Vanilla transformer · Synthetic bigram data, random triggers | accuracy: around 55% -> 95% | Improved |
| Train with uniform output tokens rather than outputs sampled from global bigram distributions (ablation) | Two-layer transformer · Synthetic bigram task, out-of-distribution output tokens | in-context accuracy | Improved |
| Use fewer trigger tokens (ablation) | Two-layer transformer · Synthetic bigram task | in-context learning speed | Hurt |
| Use only rare fixed triggers (ablation) | Two-layer transformer · Synthetic bigram task | in-context learning speed | Hurt |
| Use random triggers rather than fixed triggers (ablation) | Two-layer transformer · Synthetic bigram task | in-context learning speed | Hurt |
| Freeze the output matrix W_O^2 while training key-query matrices (ablation) | Simplified two-layer transformer induction head · Synthetic trigger-output prediction | memory recall | Hurt |
| Train W_O^2 before the key-query matrices (ablation) | Simplified two-layer transformer induction head · Synthetic trigger-output prediction | in-context accuracy | Improved |
| Train W_K^2 before W_K^1 (ablation) | Simplified two-layer transformer induction head · Synthetic trigger-output prediction | memory recall | Improved |
| Train W_K^1 on early context positions rather than late positions (ablation) | Simplified two-layer transformer induction head · Synthetic trigger-output prediction | memory recall | Improved |
Closest papers by meaning
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · similarity 0.91)
- Algorithmic Capabilities of Random Transformers (NeurIPS 2024 · similarity 0.91)
- Beyond Attention: Breaking the Limits of Transformer Context Length with Recurrent Memory (AAAI 2024 · similarity 0.90)
- Understanding Transformers via N-Gram Statistics (NeurIPS 2024 · similarity 0.90)
- Extrapolation by Association: Length Generalization Transfer In Transformers (NeurIPS 2025 · similarity 0.89)
- Trainable Transformer in Transformer (ICML 2024 · similarity 0.89)
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.
