Research map / Image generation / Autoregressive and latent generation
Latent Diffusion for Language Generation
Justin Lovelace, Varsha Kishore, Chao Wan, Eliot Seo Shekhtman, Kilian Q Weinberger
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 |
|---|---|---|---|
| diffuse in compressed 32x64 latent space instead of full encoder features (ablation) | LD4LG (BART-base) · ROCStories unconditional | MAUVE: .605 -> .716 | Improved |
| diffuse in compressed latent space instead of full encoder features (ablation) | LD4LG (BART-base) · ROCStories unconditional | relative training speedup: 1.0x -> 3.86x | Improved |
| add self-conditioning (ablation) | LD4LG (BART-base) · ROCStories unconditional | MAUVE: .480 -> .716 | Improved |
| add self-conditioning (ablation) | LD4LG (BART-base) · ROCStories unconditional | perplexity: 79.3 -> 30.6 | Improved |
| add self-conditioning (ablation) | LD4LG (BART-base) · ROCStories unconditional | diversity: .427 -> .331 | Hurt |
| use LD4LG instead of Diffusion-LM | Diffusion models · ROCStories unconditional | MAUVE: .043 -> .716 | Improved |
| use LD4LG instead of GPT-2-Medium | Language models · AG News unconditional | memorization: .829 -> .293 | Improved |
| use LD4LG (BART) instead of fine-tuned GPT-2-Medium | Language models · ROCStories unconditional | MAUVE: .788 -> .716 | Hurt |
| use LD4LG instead of DiffuSeq | Seq2seq diffusion · XSum | ROUGE-L: 14.1 -> 31.9 | Improved |
| increase sampling steps from 250 to 1000 (ablation) | LD4LG (BART-base) · ROCStories unconditional | MAUVE: .716 -> .667 | Mixed |
| reduce latent dimensionality from 32x64 to 32x32 (ablation) | BART autoencoder · ROCStories unconditional | reconstruction BLEU: 97.6 -> 92.4 | Hurt |
| use classifier-free guidance w=2.0 (ablation) | LD4LG · XSum validation | ROUGE-L | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DiffuSeq — DiffuSeq: Sequence to Sequence Text Generation with Diffusion Models (ICLR 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- How to Train Your Latent Diffusion Language Model Jointly With the Latent Space (NeurIPS 2026 · similarity 0.93)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.92)
- Continuous Latent Diffusion Language Model (NeurIPS 2026 · similarity 0.92)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.91)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.91)
- TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings (AAAI 2025 · similarity 0.91)
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.
