Research map / Architectures and efficiency / Attention and transformers
Learning Transformer Programs
Dan Friedman, Alexander Wettig, Danqi Chen
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 |
|---|---|---|---|
| Train a Transformer constrained to map deterministically to a discrete program | Transformer on CoNLL-2003 NER · CoNLL-2003 NER test | F1: 70.5 -> 76.1 | Improved |
| Train a Transformer constrained to map deterministically to a discrete program | Transformer on CoNLL-2003 NER · CoNLL-2003 NER test | F1: 66.2 -> 76.1 | Improved |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Hist | token-level accuracy: 100.0 -> 83.20 | Hurt |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · 2-Hist | token-level accuracy: 98.40 -> 55.89 | Hurt |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Most Freq | token-level accuracy: 75.69 -> 64.13 | Hurt |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Sort | token-level accuracy: 99.83 -> 99.99 | No clear effect |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Dyck-1 | token-level accuracy: 99.30 -> 99.43 | No clear effect |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Dyck-2 | token-level accuracy: 99.09 -> 99.17 | No clear effect |
| Remove numerical variables and use only categorical variables (ablation) | Transformer Program · Reverse | token-level accuracy: 99.79 -> 99.75 | No clear effect |
| Increase maximum sequence length and vocabulary size | Transformer Program · RASP tasks | accuracy | Hurt |
Closest papers by meaning
- Algorithmic Capabilities of Random Transformers (NeurIPS 2024 · similarity 0.89)
- Do Efficient Transformers Really Save Computation? (ICML 2024 · similarity 0.88)
- Towards Understanding Transformers in Learning Random Walks (NeurIPS 2025 · similarity 0.87)
- Transformers are almost optimal metalearners for linear classification (NeurIPS 2025 · similarity 0.87)
- TRANSFORMER EXPLAINER: Interactive Learning of Text-Generative Models (AAAI 2025 · similarity 0.87)
- Trainable Transformer in Transformer (ICML 2024 · 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.
