Research map / Image generation / Fast diffusion sampling
Beyond Fixed: Training-Free Variable-Length Denoising for Diffusion Large Language Models
Jinsong Li, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Jiaqi Wang, Dahua Lin
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 DAEDAL instead of best fixed length | LLaDA-Instruct-8B · MATH500 | accuracy: 39.6 -> 44.2 | Improved |
| use DAEDAL instead of best fixed length | LLaDA-Instruct-8B · GSM8K | accuracy: 83.8 -> 85.8 | Improved |
| use DAEDAL instead of best fixed length | LLaDA-1.5-8B · MATH500 | accuracy: 43.6 -> 42.4 | Hurt |
| use DAEDAL instead of best fixed length | LLaDA-1.5-8B · HumanEval | accuracy: 50.0 -> 48.8 | Hurt |
| use DAEDAL instead of best fixed length | LLaDA-1.5-8B · Average of 4 benchmarks | accuracy: 54.15 -> 54.23 | No clear effect |
| add DAEDAL at same initial length 64 | LLaDA-Instruct-8B · GSM8K | accuracy: 48.0 -> 85.8 | Improved |
| add DAEDAL for efficiency (lower total tokens) | LLaDA-Instruct-8B · GSM8K | effective token ratio: 27.7% -> 73.5% | Improved |
| add Initial Length Adjustment (Stage 1) to Stage 2 (ablation) | DAEDAL · GSM8K | accuracy: 72.3 -> 85.8 | Improved |
| add Iterative Mask Insertion (Stage 2) to Stage 1 (ablation) | DAEDAL · GSM8K | accuracy: 84.1 -> 85.8 | Improved |
| raise initial length for Stage 2 only (ablation) | DAEDAL Stage 2 · GSM8K | accuracy: 72.3 -> 84.7 | Improved |
| increase initial length from 64 to 512 (ablation) | DAEDAL · GSM8K | accuracy: 85.8 -> 85.1 | No clear effect |
| increase expansion factor from 8 to 24 (ablation) | DAEDAL · GSM8K | accuracy: 85.8 -> 86.4 | No clear effect |
Closest papers by meaning
- Denoising Time Matters: Diverse Generation in Diffusion Language Models (NeurIPS 2026 · similarity 0.92)
- Uncertainty Quantification for Large Language Diffusion Models (NeurIPS 2026 · similarity 0.91)
- Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (ICLR 2025 · similarity 0.90)
- Scaling Inference Time Compute for Diffusion Models (CVPR 2025 · similarity 0.90)
- Accelerating Diffusion Language Models via Structured Suffix Modeling (NeurIPS 2026 · similarity 0.90)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 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.
