Research map / Architectures and efficiency / Attention and transformers
How Many Different Outputs Can a Transformer Generate?
Maxime Meyer, Mario Michelessa, Caroline Chaux, Vincent Y. F. Tan
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 |
|---|---|---|---|
| replace ball support with ellipsoid support (ablation) | slope upper bound · PG19 cramming, Pythia-160M | bound/empirical slope ratio: 9.24 -> 7.92 | Improved |
| replace ball support with cone support (ablation) | slope upper bound · PG19 cramming, Qwen-2.5-1.5B | bound/empirical slope ratio: 20.4 -> 20.34 | Improved |
| add non-uniform cell volumes to the ellipsoid bound (ablation) | slope upper bound · PG19 cramming, Pythia-1B | bound/empirical slope ratio: 6.12 -> 4.56 | Improved |
| add non-uniform cell volumes to the ellipsoid bound (ablation) | slope upper bound · PG19 cramming, Qwen-2.5-1.5B | bound/empirical slope ratio: 15.30 -> 10.82 | Improved |
| use variable epsilon instead of uniform precision (ellipsoid) (ablation) | slope upper bound · PG19 cramming, Gemma-3-270M | bound/empirical slope ratio: 11.12 -> 13.42 | Hurt |
| increase soft prompt length m | frozen transformer cramming · PG19 and random targets | n_50 vs m linear fit R^2: R^2 0.999 and 0.995 | Improved |
| use natural text (PG19) instead of random target strings | frozen transformer cramming · cramming accessibility | accessible length at same m | Improved |
Closest papers by meaning
- Arithmetic Transformers Can Length-Generalize in Both Operand Length and Count (ICLR 2025 · similarity 0.87)
- Two Heads are Better than One: Simulating Large Transformers with Small Ones (NeurIPS 2025 · similarity 0.87)
- Approximation Rate of the Transformer Architecture for Sequence Modeling (NeurIPS 2024 · similarity 0.86)
- A Formal Framework for Understanding Length Generalization in Transformers (ICLR 2025 · similarity 0.86)
- Understanding the Expressive Power and Mechanisms of Transformer for Sequence Modeling (NeurIPS 2024 · similarity 0.86)
- Continuity and Isolation Lead to Doubts or Dilemmas in Large Language Models (NeurIPS 2025 · similarity 0.85)
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.
