Research map / Large language models / Language, safety and interpretability
Copy That! Editing Sequences by Copying Spans
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq editor · WikiAtomicEdits | Accuracy: 67.8% -> 78.1% | Helped |
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq editor · GitHubEdits | Accuracy: 64.4% -> 67.4% | Helped |
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq editor · C# Fixers | Accuracy: 18.8% -> 24.2% | Helped |
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq code repair · BFPsmall | Accuracy: 14.8% -> 17.7% | Helped |
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq code repair · BFPmedium | Accuracy: 7.0% -> 8.0% | Helped |
| add copying of input spans rather than limiting copying to single tokens | Seq2Seq grammar correction · FCE test | F0.5: 0.1853 -> 0.2134 | Helped |
| add marginalization over all correct action sequences (ablation) | Seq2Seq+CopySpan code repair · BFPsmall | Accuracy: 16.9% -> 17.7% | Helped |
| add marginalization over all correct action sequences (ablation) | Seq2Seq+CopySpan code repair · BFPmedium | Accuracy: 2.5% -> 8.0% | Helped |
| allow the model to choose among correct span lengths instead of always selecting the longest (ablation) | Seq2Seq+CopySpan code repair · BFPsmall | Accuracy: 14.2% -> 17.7% | Helped |
| allow the model to choose among correct span lengths instead of always selecting the longest (ablation) | Seq2Seq+CopySpan code repair · BFPmedium | Accuracy: 7.2% -> 8.0% | Helped |
| merge equivalent beam rays during search rather than only after search (ablation) | Seq2Seq+CopySpan code repair · BFPsmall | Accuracy@20: 41.6% -> 45.0% | Helped |
| merge equivalent beam rays during search rather than only after search (ablation) | Seq2Seq+CopySpan code repair · BFPmedium | Accuracy@20: 18.0% -> 25.4% | Helped |
Closest papers by meaning
- Repeat After Me: Transformers are Better than State Space Models at Copying (ICML 2024 · similarity 0.86)
- The Missing Alignment Link of In-context Learning on Sequences (ICML 2025 · similarity 0.86)
- Editing Memories Through Few Targeted Neurons (AAAI 2025 · similarity 0.85)
- Rethinking Residual Distribution in Locate-then-Edit Model Editing (NeurIPS 2025 · similarity 0.85)
- Frontier Language Models Struggle to Copy: Text Can Be Better Viewed in 2D (NeurIPS 2026 · similarity 0.85)
- Spectral Editing of Activations for Large Language Model Alignment (NeurIPS 2024 · 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.