Research map / Large language models / Language, safety and interpretability
Fast-dLLM v2: Efficient Block-Diffusion LLM
Chengyue Wu, Hao Zhang, Shuchen Xue, Shizhe Diao, Yonggan Fu, Zhijian Liu, Pavlo Molchanov, Ping Luo, Song Han, Enze Xie
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 |
|---|---|---|---|
| add padding to align packed samples to block size (ablation) | Fast-dLLM v2 1.5B · avg of 7 suites | average score: 41.3 -> 42.2 | Improved |
| add complementary masking (ablation) | Fast-dLLM v2 1.5B · avg of 7 suites | average score: 42.2 -> 45.0 | Improved |
| use block diffusion adaptation vs NTP fine-tuning on same data | Qwen2.5-1.5B-Instruct · avg of 7 suites | average score: 44.3 -> 45.0 | Improved |
| use block diffusion adaptation vs NTP fine-tuning on same data | Qwen2.5-7B-Instruct · avg of 7 suites | average score: 59.6 -> 60.3 | Improved |
| block diffusion vs NTP fine-tuning on MATH | Qwen2.5-7B-Instruct · MATH | accuracy: 72.0 -> 61.6 | Hurt |
| enable confidence-threshold 0.9 parallel decoding (ablation) | Fast-dLLM v2 7B · GSM8K | throughput (tokens/s): 39.1 -> 101.7 | Improved |
| enable confidence-threshold 0.9 parallel decoding (ablation) | Fast-dLLM v2 7B · GSM8K | accuracy | No clear effect |
| use diffusion decoding vs AR decoding on H100 at batch 64 | Fast-dLLM v2 7B vs Qwen2.5-7B-Instruct · GSM8K | throughput speedup: 1.8x | Improved |
| use sub-block size 8 (vs 2 or 32) (ablation) | Fast-dLLM v2 1.5B · HumanEval | pass@1: 38.4 -> 43.9 | Improved |
| use smaller sub-block size 2 on GSM8K (ablation) | Fast-dLLM v2 1.5B · GSM8K | accuracy: 60.2 -> 62.8 | Improved |
| change inference block size away from trained size 32 (ablation) | Fast-dLLM v2 1.5B · GSM8K | accuracy: 60.2 -> 53.2 | Hurt |
| add sub-block cache (ablation) | Fast-dLLM v2 · GSM8K | accuracy | No clear effect |
Closest papers by meaning
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- From Next-Token to Next-Block: A Principled Adaptation Path for Diffusion LLMs (NeurIPS 2026 · similarity 0.92)
- SimSD: Simple Speculative Decoding in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.92)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.91)
- dKV-Cache: The Cache for Diffusion Language Models (NeurIPS 2025 · 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.
