Diffusion Language Models Change Speed Across Tasks and Languages
Diffusion language models can write several tokens in one forward pass, and they decide during generation how many passes to spend.
ProofPaper ↗
Key points
- To understand the underlying factors of this cost variation, we analyze two diffusion models across 10 tasks.
- We find that the number of forward passes per token depends heavily on the task.
- In multilingual generation, we find that highly tokenized languages also require more passes per token, adding to the cost they already pay with more tokens.
- Taken together, our results show that diffusion's computational advantage depends on how much of the output is already determined, while its decoding effort offers a practical signal for making inference more efficient and reliable.
Sources (1)
- [1]Diffusion Language Models Change Speed Across Tasks and LanguagesarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 05:47 AM
Diffusion language models can write several tokens in one forward pass, and they decide during generation how many passes to spend.
To understand the underlying factors of this cost variation, we analyze two diffusion models across 10 tasks.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 19, 2026SWIFT: Self-guided Weak-to-strong dIFfusion models via Target extrapolation
- Oct 19, 2026RoAR-VAE: Improving Latent Diffusability through Robust Prediction
- Oct 18, 2026AffordanceWAM: Affordance-Aware Joint World-Action Modeling for Robot Manipulation
- Oct 16, 2026[Research] 2,788 tokens/s on a single RTX 5090 with a diffusion LLM — Introducing ORL (fully open source)
- Oct 16, 2026axolotl-ai-cloud/axolotl v0.21.0
- Oct 16, 2026ScoutRoute: Efficient Mixture-of-Experts Routing Using Internal Language Model Signals