AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Learning Transformer Programs

Dan Friedman, Alexander Wettig, Danqi Chen

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Transformer expressivity and learnability · 13 citations (OpenAlex)

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 changedWhereMeasuredEffect
Train a Transformer constrained to map deterministically to a discrete programTransformer on CoNLL-2003 NER · CoNLL-2003 NER testF1: 70.5 -> 76.1Improved
Train a Transformer constrained to map deterministically to a discrete programTransformer on CoNLL-2003 NER · CoNLL-2003 NER testF1: 66.2 -> 76.1Improved
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Histtoken-level accuracy: 100.0 -> 83.20Hurt
Remove numerical variables and use only categorical variables (ablation)Transformer Program · 2-Histtoken-level accuracy: 98.40 -> 55.89Hurt
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Most Freqtoken-level accuracy: 75.69 -> 64.13Hurt
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Sorttoken-level accuracy: 99.83 -> 99.99No clear effect
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Dyck-1token-level accuracy: 99.30 -> 99.43No clear effect
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Dyck-2token-level accuracy: 99.09 -> 99.17No clear effect
Remove numerical variables and use only categorical variables (ablation)Transformer Program · Reversetoken-level accuracy: 99.79 -> 99.75No clear effect
Increase maximum sequence length and vocabulary sizeTransformer Program · RASP tasksaccuracyHurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy