Research map / Large language models / Language, safety and interpretability
A Retrieve-and-Edit Framework for Predicting Structured Outputs
Tatsunori B Hashimoto, Kelvin Guu, Yonatan Oren, Percy Liang
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 |
|---|---|---|---|
| Add retrieved-example editing to a Seq2Seq model | Seq2Seq · Python GitHub autocomplete | BLEU: 19.2 -> 34.7 | Improved |
| Add retrieved-example editing instead of retrieval alone | TaskRetriever · Python GitHub autocomplete | BLEU: 29.9 -> 34.7 | Improved |
| Use task-dependent output-reconstructing embeddings instead of bag-of-words embeddings (ablation) | Python autocomplete retriever · Python GitHub autocomplete | BLEU: 23.1 -> 29.9 | Improved |
| Train the encoder-decoder to reconstruct y rather than the input x (ablation) | InputRetriever · Python GitHub autocomplete | Longest completed length: 12.3 -> 13.5 | Improved |
| Train the encoder-decoder to reconstruct y rather than the input x (ablation) | InputRetriever · Python GitHub autocomplete | BLEU: 29.8 -> 29.9 | No clear effect |
| Add retrieve-and-edit to a sequence-to-sequence model | Sequence-to-sequence · Hearthstone cards | BLEU: 60.4 -> 70.0 | Improved |
| Add retrieve-and-edit to a sequence-to-sequence model | Sequence-to-sequence · Hearthstone cards | Accuracy: 1.5 -> 9.1 | Improved |
| Add identity-map data augmentation with probability 0.1 (ablation) | Retrieve+Edit editor · Python GitHub autocomplete and Hearthstone cards | Performance | No clear effect |
Closest papers by meaning
- Repoformer: Selective Retrieval for Repository-Level Code Completion (ICML 2024 · similarity 0.84)
- CoRNStack: High-Quality Contrastive Data for Better Code Retrieval and Reranking (ICLR 2025 · similarity 0.84)
- CPRet: A Dataset, Benchmark, and Model for Retrieval in Competitive Programming (NeurIPS 2025 · similarity 0.83)
- LILO: Learning Interpretable Libraries by Compressing and Documenting Code (ICLR 2024 · similarity 0.83)
- Structured Packing in LLM Training Improves Long Context Utilization (AAAI 2025 · similarity 0.83)
- Let the Code LLM Edit Itself When You Edit the Code (ICLR 2025 · similarity 0.83)
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.
