Research map / Image generation / Fast diffusion sampling
ReFusion: A Diffusion Large Language Model with Parallel Autoregressive Decoding
Jia-Nan Li, Jian Guan, Wei Wu, Chongxuan Li
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 ReFusion instead of Dream-7B-Instruct | MDM · HumanEval | pass@1: 56.71 -> 78.66 | Improved |
| use ReFusion instead of Qwen3-8B | ARM · GSM8K | accuracy: 81.96 -> 84.91 | Improved |
| use ReFusion instead of Qwen3-8B | ARM · MBPP | pass@1: 63.80 -> 68.20 | Improved |
| use ReFusion instead of Qwen3-8B | ARM · MATH | accuracy: 83.28 -> 54.22 | Hurt |
| train ReFusion rather than LLaDA on 120K subset | Qwen3-8B init · HumanEval | pass@1: 42.07 -> 70.12 | Improved |
| use ReFusion rather than BD3-LMs on 120K subset | Qwen3-8B init · HumanEval | pass@1: 59.15 -> 70.12 | Improved |
| ReFusion on Dream's Qwen2.5-7B backbone vs Dream-7B-Instruct | Qwen2.5-7B · HumanEval | pass@1: 56.71 -> 68.90 | Improved |
| ReFusion variant vs Dream on knowledge tasks | Qwen2.5-7B · MMLU-Pro | accuracy: 40.05 -> 35.25 | Hurt |
| add KV cache reuse by concatenation instead of re-computation (ablation) | ReFusion · GSM8K | accuracy: 84.38 -> 84.91 | No clear effect |
| add KV cache reuse by concatenation instead of re-computation (ablation) | ReFusion · HumanEval | pass@1: 77.44 -> 78.66 | Improved |
| use first-token probability instead of mean slot probability for certainty (ablation) | ReFusion · GSM8K | accuracy: 82.71 -> 84.91 | No clear effect |
| use first-token probability instead of mean slot entropy for certainty (ablation) | ReFusion · MMLU-Pro | accuracy: 41.95 -> 45.94 | Improved |
Closest papers by meaning
- Accelerating Diffusion LLMs via Adaptive Parallel Decoding (NeurIPS 2025 · similarity 0.92)
- Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion (NeurIPS 2026 · similarity 0.92)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · similarity 0.92)
- Mean-Field Parallel Decoding for Discrete Diffusion Language Models (NeurIPS 2026 · similarity 0.91)
- Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning (NeurIPS 2025 · similarity 0.91)
- Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs (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.
