Research map / Image generation / Fast diffusion sampling
Scaling Beyond Masked Diffusion Language Models
Subham Sekhar Sahoo, Jean-Marie Lemercier, Zhihan Yang, Justin Deschenaux, Jingyu Liu, John Thickstun, Ante Jukić
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 |
|---|---|---|---|
| train with the low-variance objective instead of the true ELBO (ablation) | MDLM · IsoFLOP scaling | compute multiplier to match AR: 16x -> 14x | Improved |
| train with the low-variance objective (ablation) | MDLM · IsoFLOP scaling | compute-optimal model size | Improved |
| use uniform-state diffusion (Duo) instead of AR | 1.7B models · GSM8K (after SFT) | accuracy: 62.9 -> 65.8 | Improved |
| use uniform-state diffusion (Duo) instead of masked diffusion | 1.7B models · GSM8K (after SFT) | accuracy: 58.8 -> 65.8 | Improved |
| use interpolating diffusion (Eso-LM) instead of AR | 1.7B models · GSM8K (after SFT) | accuracy: 62.9 -> 33.4 | Hurt |
| use Duo instead of AR | 1.7B models · ARC-e | accuracy: 72.7 -> 53.4 | Hurt |
| use Duo instead of MDLM | 1.7B models · PIQA | accuracy: 62.2 -> 62.7 | No clear effect |
| use Duo instead of MDLM | 1.7B models · BoolQ | accuracy: 62.8 -> 59.6 | Hurt |
| use Duo instead of MDLM | 1.7B models · ARC-e | accuracy: 50.5 -> 53.4 | Improved |
| use Eso-LM with KV caching instead of Duo | speed-quality frontier · throughput 400-600 | Gen PPL vs throughput | Improved |
| use few-step generation (Duo) instead of AR | speed-quality frontier · throughput 200-400 and 600+ | Gen PPL vs throughput | Improved |
| use Duo instead of AR for compute efficiency | IsoFLOP scaling · validation loss | compute multiplier vs AR: 23x | Hurt |
Closest papers by meaning
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.95)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.94)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.93)
- Breaking AR’s Sampling Bottleneck: Provable Acceleration via Diffusion Language Models (NeurIPS 2025 · similarity 0.93)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.92)
- Masked Diffusion Models are Secretly Time-Agnostic Masked Models and Exploit Inaccurate Categorical Sampling (ICLR 2025 · similarity 0.92)
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.
