Research map / Large language models / Language, safety and interpretability
WeDLM: Reconciling Diffusion Language Models with Standard Causal Attention for Fast Inference
Aiwei Liu, Minghua He, Shaoxun Zeng, Sijun Zhang, Linhao Zhang, Chuhan Wu, Wei Jia, Yuan Liu, Zhou Xiao, Jie Zhou
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 |
|---|---|---|---|
| replace block-wise decoding with streaming parallel decoding (ablation) | WeDLM-7B-Instruct · MATH, tau=0.9 | tokens/s: 221 -> 423 tokens/s (1.9x) | Improved |
| increase distance penalty lambda from 0.01 to 0.05 (ablation) | WeDLM-7B-Instruct · MATH | accuracy: +2.6 points with 3% speed reduction | Improved |
| raise entropy threshold tau above 0.5 (ablation) | WeDLM-7B-Instruct · MATH | accuracy | Hurt |
| use causal intra-block attention instead of bidirectional intra-block attention (ablation) | WeDLM · average over benchmarks (AR-initialized) | average score | Improved |
| use diffusion-style training instead of continued AR pretraining (7B) (ablation) | WeDLM-7B · average over benchmarks | average score: +3.5 points over AR continue | Improved |
| apply WeDLM adaptation to small models (0.6B, 1.5B) (ablation) | WeDLM · average over benchmarks | average score | Hurt |
| vary training block size B among 4, 8, 32 (ablation) | WeDLM-7B · average over benchmarks | average score: 71.23 / 70.45 / 70.84 | No clear effect |
| convert Qwen3-8B-Instruct into WeDLM-8B-Instruct | WeDLM-8B · average over 9 benchmarks | average score: 75.12 -> 77.53 | Improved |
| convert Qwen3-8B-Instruct into WeDLM-8B-Instruct | WeDLM-8B · MATH (4-shot) | accuracy: 69.60 -> 64.80 | Hurt |
| convert Qwen3-8B-Instruct into WeDLM-8B-Instruct | WeDLM-8B · HumanEval (4-shot) | accuracy: 71.95 -> 80.49 | Improved |
| convert base models to WeDLM | WeDLM-7B/8B base · MBPP | accuracy: 3-4 points drop | Hurt |
| raise tau from 0.2 to 0.9 in streaming decoding (ablation) | WeDLM-8B-Instruct · GSM8K | accuracy / speedup: 92.3% at 1.97x -> above 79% at 3.2x | Mixed |
Later papers that compared with this method
1 later accepted paper reports WeDLM-8B as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- DiffuReason: Enhancing Reasoning Ability for Diffusion Language Models via Monte Carlo Tree Search (ICML 2026 · better on 4 of 4 shared results)
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
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.94)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.92)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.91)
- Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion (NeurIPS 2026 · similarity 0.91)
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.
