AI Paper Atlas

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

ICML 2023 · Large language models › Language, safety and interpretability · approach: Controlled text generation · 10 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
Use natural-language constraint verbalization instead of omitting it (ablation)InstructCTG · unseen semantic and style constraintsconstraint satisfaction rate: 67.7 -> 83.4; 65.1 -> 87.3Improved
Use natural-language constraint verbalization (ablation)InstructCTG · Lexical constraintconstraint satisfaction rate: 95.5 -> 97.5Improved
Include in-context demonstrations (ablation)InstructCTG · unseen semantic and style constraintsconstraint satisfaction rate: 79.8 -> 83.4; 83.1 -> 87.3Improved
Include in-context demonstrations (ablation)InstructCTG · unseen constraints; length held out during trainingconstraint satisfaction rate: 82% -> 88%Improved
Train with multiple constraints at the same time (ablation)InstructCTG · seen constraintsconstraint satisfaction rate: 97.5 -> 97.2; 88.3 -> 87.0; 88.2 -> 86.8; 90.6 -> 89.2Improved
Use multiple verbalization templates (ablation)InstructCTG · Semantic constraintconstraint satisfaction rate: 84.7 -> 88.2Improved
Use multiple verbalization templates (ablation)InstructCTG · Style constraintconstraint satisfaction rate: 90.2 -> 90.6Improved
Use multiple verbalization templates (ablation)InstructCTG · Length constraintconstraint satisfaction rate: 100.0 -> 100.0No clear effect
Use InstructCTG instead of constraint-specific fine-tuningT0-11B · Semantic constraintconstraint satisfaction rate: 85.1 -> 88.2Improved
Use InstructCTG instead of constraint-specific fine-tuningT0-11B · Style constraintfluency (perplexity): 17.4 -> 13.1Improved
Use InstructCTG rather than Diffusion-LMT0-11B · Syntax constraintconstraint satisfaction rate: 88.6 -> 88.3No clear effect
Use InstructCTG rather than NeuroLogic decodingT0-11B · Lexical constraintfluency (perplexity): 21.8 -> 12.6Improved

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