Research map / Image generation / Autoregressive and latent generation
Diffusion-LM Improves Controllable Text Generation
Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, Tatsunori Hashimoto
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 |
|---|---|---|---|
| Diffusion-LM with classifier guidance vs FUDGE | controllable generation · E2E Syntax Tree control | control success: 17.9 -> 86.0 | Improved |
| Diffusion-LM vs fine-tuning oracle FT-search | controllable generation · E2E Semantic Content control | control success: 89.9 vs 81.2 | Hurt |
| compose controls via summed classifier gradients vs product of experts of fine-tuned models | Diffusion-LM · Semantic Content + Syntax Tree | syntax ctrl: 29.2 -> 74.8 | Improved |
| Diffusion-LM infilling without specialized training vs AR trained for infilling | infilling · ROCStories infilling | BLEU-4: 6.7 vs 7.1 | No clear effect |
| diffusion instead of autoregressive LM | small Transformer LM · ROCStories | NLL: 3.05 -> 3.88 | Hurt |
| learned end-to-end embeddings instead of random (ablation) | Diffusion-LM · ROCStories, E2E | lm-score | Improved |
| predict x0 instead of noise epsilon (ablation) | Diffusion-LM · E2E, ROCStories | lm-score | Improved |
| sqrt noise schedule instead of cosine/linear (ablation) | Diffusion-LM · E2E, ROCStories | lm-score | Improved |
| Transformer instead of U-Net denoiser (ablation) | Diffusion-LM · E2E, ROCStories | lm-score | Improved |
| add fluency regularization term in guided decoding | Diffusion-LM · controllable generation | fluency | Improved |
Closest papers by meaning
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.91)
- Energy-Based Diffusion Language Models for Text Generation (ICLR 2025 · similarity 0.90)
- TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings (AAAI 2025 · similarity 0.90)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.90)
- Continuous Latent Diffusion Language Model (NeurIPS 2026 · similarity 0.90)
- How to Train Your Latent Diffusion Language Model Jointly With the Latent Space (NeurIPS 2026 · similarity 0.90)
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.
