Research map / Label-efficient and robust learning / Recognition and long-tail
A Distributional Approach to Controlled Text Generation
Muhammad Khalifa, Hady Elsahar, Marc Dymetman
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 GDC instead of PPLM for positive sentiment control | GPT-2 · Positive sentiment | sentiment expectation: 0.52 -> 0.56 | Improved |
| Use GDC instead of PPLM for positive sentiment control | GPT-2 · Positive sentiment | perplexity: 29.26$\pm$22.07 -> 13.53$\pm$3.18 | Improved |
| Use GDC instead of PPLM for negative sentiment control | GPT-2 · Negative sentiment | sentiment expectation: 0.14 -> 0.51 | Improved |
| Use GDC instead of PPLM for negative sentiment control | GPT-2 · Negative sentiment | perplexity: 27.72$\pm$23.95 -> 13.59$\pm$3.84 | Improved |
| Add a female-rate distributional constraint targeting 50% | GPT-2bio · Biography generation | female biography rate: 07.4 -> 36.7 | Improved |
| Add multiple profession-rate distributional constraints | GPT-2bio · Biography generation | Science rate: 01.5 -> 20.1 | Improved |
| Add multiple profession-rate distributional constraints | GPT-2bio · Biography generation | Art rate: 10.9 -> 31.6 | Improved |
| Add multiple profession-rate distributional constraints | GPT-2bio · Biography generation | Business rate: 10.9 -> 10.2 | Improved |
| Add multiple profession-rate distributional constraints | GPT-2bio · Biography generation | Sports rate: 19.5 -> 11.9 | Improved |
| Add adaptive proposal updates to DPG (ablation) | DPG for rare-word constraints · Vampire, Paris, and US word constraints | convergence speed | Improved |
| Use TVD rather than KL divergence to decide proposal updates (ablation) | DPG-KLD · Vampire, Paris, and US word constraints | convergence speed | No clear effect |
| Remove adaptive proposal updates (ablation) | DPG · Vampire, Paris, and US word constraints | convergence speed | Hurt |
Closest papers by meaning
- Guaranteed Generation from Large Language Models (ICLR 2025 · similarity 0.91)
- Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo (ICLR 2025 · similarity 0.90)
- Controlled Text Generation via Language Model Arithmetic (ICLR 2024 · similarity 0.88)
- Controllable Generation via Locally Constrained Resampling (ICLR 2025 · similarity 0.87)
- Controlled LLM Decoding via Discrete Auto-regressive Biasing (ICLR 2025 · similarity 0.87)
- Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · similarity 0.87)
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.
