Research map / Image generation / Autoregressive and latent generation
Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation
Alexander Shabalin, Viacheslav Meshchaninov, Dmitry Vetrov
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 smoothing diffusion space instead of simplex space (ablation) | diffusion text model (~100M) · XSum | ROUGE-L: 23.0 -> 26.0 | Improved |
| use smoothing diffusion space instead of embedding space (ablation) | diffusion text model (~100M) · XSum | ROUGE-L: 24.6 -> 26.0 | Improved |
| use smoothing diffusion space instead of embedding space (ablation) | diffusion text model (~100M) · ROCStories | MAUVE: 41.5 -> 80.7 | Improved |
| use smoothing diffusion space instead of simplex space (ablation) | diffusion text model (~100M) · ROCStories | MAUVE: 15.2 -> 80.7 | Improved |
| use smoothing diffusion space instead of embedding space (ablation) | diffusion text model (~100M) · QQP | BLEU: 31.3 -> 30.8 | No clear effect |
| use Eq. 9 solver (ignores previous latent) instead of DDPM (ablation) | Smoothie · XSum / Quasar-T / QQP / ParaDetox | ROUGE-L / J-Score: 25.6 -> 26.0 (XSum R-L) | Improved |
| use DDIM solver instead of Eq. 9 solver (ablation) | Smoothie · Quasar-T | ROUGE-L: 36.5 -> 31.8 | Hurt |
| lower reverse-process noise delta-tilde (ablation) | Smoothie unconditional generation · ROCStories | perplexity vs diversity | Mixed |
| increase number of denoising steps beyond 200 (ablation) | Smoothie · XSum | BERTScore: 68.8 -> 68.4 (200 -> 500 steps) | Hurt |
| increase number of denoising steps (ablation) | Smoothie · Quasar-T | BERTScore: 63.1 at 25 to 500 steps | No clear effect |
| use cross-entropy loss instead of MSE objective (ablation) | Smoothie · seq2seq tasks | generation quality | Hurt |
| use embedding diffusion instead of smoothing for speed | diffusion text model · generation time | seconds: 2.897 -> 1.642 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- TEncDM — TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings (AAAI 2025) · better on 4 of 5 shared results
- AR-Diffusion — AR-Diffusion: Auto-Regressive Diffusion Model for Text Generation (NeurIPS 2023) · better on 1 of 1 shared result
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
- TEncDM: Understanding the Properties of the Diffusion Model in the Space of Language Model Encodings (AAAI 2025 · similarity 0.89)
- Diffusion Adaptive Text Embedding for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.89)
- TokenCompose: Text-to-Image Diffusion with Token-level Supervision (CVPR 2024 · similarity 0.88)
- Smooth Diffusion: Crafting Smooth Latent Spaces in Diffusion Models (CVPR 2024 · similarity 0.88)
- Compressed and Smooth Latent Space for Text Diffusion Modeling (NeurIPS 2025 · similarity 0.88)
- Continuous Latent Diffusion Language Model (NeurIPS 2026 · 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.
