Research map / Architectures and efficiency / Attention and transformers
Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed
Yonggan Fu, Lexington Whalen, Zhifan Ye, Xin Dong, Shizhe Diao, Jingyu Liu, Chengyue Wu, Hao Zhang, Enze Xie, Song Han, Maksim Khadkevich, Jan Kautz and 2 more
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 block-wise attention instead of bidirectional (no clean context) (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 19.29 -> 28.23 | Improved |
| add clean context to block-wise attention (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 28.23 -> 37.69 | Improved |
| remove token shift (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 37.69 -> 38.41 | Improved |
| use full best scheme instead of bidirectional attention (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 19.29 -> 38.41 | Improved |
| use LoRA rank 64 instead of full-model training (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 38.41 -> 30.78 | Hurt |
| increase LoRA rank from 16 to 64 (ablation) | AR-to-dLM conversion · Qwen2.5 1.5B, 50B tokens, 6 generation tasks | average accuracy: 27.93 -> 30.78 | Improved |
| add position-dependent masking (ablation) | AR-to-dLM conversion · Qwen3 4B, AR-to-dLM components (6 math/coding tasks) | average accuracy: 60.27 -> 62.02 | Improved |
| scale training to 300B tokens (ablation) | AR-to-dLM conversion · Qwen3 4B, AR-to-dLM components (6 math/coding tasks) | average accuracy: 62.02 -> 64.05 | Improved |
| add block-wise attention with clean context to Dream's bidirectional setting (ablation) | AR-to-dLM conversion · Qwen3 4B, AR-to-dLM components (6 math/coding tasks) | average accuracy: 44.59 -> 59.01 | Improved |
| lower initial learning rate from 1e-4 to 1e-5 (ablation) | AR-to-dLM conversion · Qwen3 4B, 25B tokens, initial learning rate sweep | average accuracy: 50.95 -> 60.63 | Improved |
| lower initial learning rate from 1e-5 to 1e-6 (ablation) | AR-to-dLM conversion · Qwen3 4B, 25B tokens, initial learning rate sweep | average accuracy: 60.63 -> 54.56 | Hurt |
| use Efficient-DLM 1.5B instead of Qwen2.5 1.5B | Efficient-DLM · 12-task average (coding, math, MMLU, commonsense) | average accuracy: 54.47 -> 52.04 | Mixed |
Closest papers by meaning
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.93)
- Scaling Diffusion Language Models via Adaptation from Autoregressive Models (ICLR 2025 · similarity 0.91)
- Diffusion Beats Autoregressive in Data-Constrained Settings (NeurIPS 2025 · similarity 0.90)
- Looped Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.90)
- Scaling up Masked Diffusion Models on Text (ICLR 2025 · 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.
