Research map / Large language models / Language, safety and interpretability
Controlled Text Generation with Natural Language Instructions
Wangchunshu Zhou, Yuchen Eleanor Jiang, Ethan Wilcox, Ryan Cotterell, Mrinmaya Sachan
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 |
|---|---|---|---|
| Use natural-language constraint verbalization instead of omitting it (ablation) | InstructCTG · unseen semantic and style constraints | constraint satisfaction rate: 67.7 -> 83.4; 65.1 -> 87.3 | Improved |
| Use natural-language constraint verbalization (ablation) | InstructCTG · Lexical constraint | constraint satisfaction rate: 95.5 -> 97.5 | Improved |
| Include in-context demonstrations (ablation) | InstructCTG · unseen semantic and style constraints | constraint satisfaction rate: 79.8 -> 83.4; 83.1 -> 87.3 | Improved |
| Include in-context demonstrations (ablation) | InstructCTG · unseen constraints; length held out during training | constraint satisfaction rate: 82% -> 88% | Improved |
| Train with multiple constraints at the same time (ablation) | InstructCTG · seen constraints | constraint satisfaction rate: 97.5 -> 97.2; 88.3 -> 87.0; 88.2 -> 86.8; 90.6 -> 89.2 | Improved |
| Use multiple verbalization templates (ablation) | InstructCTG · Semantic constraint | constraint satisfaction rate: 84.7 -> 88.2 | Improved |
| Use multiple verbalization templates (ablation) | InstructCTG · Style constraint | constraint satisfaction rate: 90.2 -> 90.6 | Improved |
| Use multiple verbalization templates (ablation) | InstructCTG · Length constraint | constraint satisfaction rate: 100.0 -> 100.0 | No clear effect |
| Use InstructCTG instead of constraint-specific fine-tuning | T0-11B · Semantic constraint | constraint satisfaction rate: 85.1 -> 88.2 | Improved |
| Use InstructCTG instead of constraint-specific fine-tuning | T0-11B · Style constraint | fluency (perplexity): 17.4 -> 13.1 | Improved |
| Use InstructCTG rather than Diffusion-LM | T0-11B · Syntax constraint | constraint satisfaction rate: 88.6 -> 88.3 | No clear effect |
| Use InstructCTG rather than NeuroLogic decoding | T0-11B · Lexical constraint | fluency (perplexity): 21.8 -> 12.6 | Improved |
Closest papers by meaning
- A Distributional Approach to Controlled Text Generation (ICLR 2021 · similarity 0.87)
- InsNet: An Efficient, Flexible, and Performant Insertion-based Text Generation Model (NeurIPS 2022 · similarity 0.87)
- A Contrastive Framework for Neural Text Generation (NeurIPS 2022 · similarity 0.86)
- COLLIE: Systematic Construction of Constrained Text Generation Tasks (ICLR 2024 · similarity 0.85)
- Guaranteed Generation from Large Language Models (ICLR 2025 · similarity 0.85)
- Making Text Embedders Few-Shot Learners (ICLR 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.
