Research map / Image generation / Fast diffusion sampling
Scaling Behavior of Discrete Diffusion Language Models
Dimitri von Rütte, Janis Fluri, Omead Pooladzandi, Bernhard Schölkopf, Thomas Hofmann, Antonio Orvieto
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 |
|---|---|---|---|
| move from masked to uniform noise (ablation) | discrete diffusion LM · compute-optimal scaling | model-size exponent alpha_M: 0.566 -> 0.589 | Improved |
| move from masked to uniform noise (ablation) | discrete diffusion LM · compute-bound loss scaling | loss exponent alpha_L: -0.0496 -> -0.0522 | Improved |
| increase uniform noise share (ablation) | discrete diffusion LM · likelihood vs compute | ELBO gap to masked: 3.2% -> 1.7% (1e18 -> 1e21 FLOPs) | Improved |
| add 20% LR cooldown (annealing) (ablation) | discrete diffusion LM · ELBO | ELBO improvement: 2.45% | Improved |
| add confidence-based adaptive sampling instead of ancestral sampling (ablation) | Masking 3B · GSM8k | accuracy (%): 1.06 -> 1.67 | Improved |
| add confidence-based adaptive sampling instead of ancestral sampling (ablation) | Uniform 3B · GSM8k | accuracy (%): 1.44 -> 2.05 | Improved |
| use uniform instead of masked 3B model | 3B DLM · downstream multiple-choice benchmarks | accuracy (%): ARC-E 49.9 -> 50.6; BoolQ 60.9 -> 56.4 | Mixed |
| scale uniform diffusion from 3B to 10B | uniform DLM · ARC-Easy | accuracy (%): 50.6 -> 61.8 | Improved |
| fix optimal batch size by training tokens | discrete diffusion LM · optimal batch size | power-law exponent: 0.8225 | Improved |
Closest papers by meaning
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.92)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.91)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · similarity 0.90)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.90)
- Simple and Effective Masked Diffusion Language Models (NeurIPS 2024 · similarity 0.90)
- Looped Diffusion Language Models (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.
