Research map / Architectures and efficiency / Attention and transformers
DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding
Yanhua Jiao, Tianyi Wu, Xiaoxi Sun, Yulin Li, Hui-Ling Zhen, Libo Qin, Baotian Hu, Zhuotao Tian, Min Zhang
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 non-contiguous with contiguous commitment of tokens above threshold (ablation) | LLaDA-1.5 parallel decoding · GSM8K, tau=0.60 | accuracy: 75.13 -> 77.63 | Improved |
| replace non-contiguous with contiguous commitment at higher threshold (ablation) | LLaDA-1.5 parallel decoding · GSM8K, tau=0.75 | accuracy: 80.44 -> 80.59 | No clear effect |
| add Dynamic Contiguous Verification and draft-guided decoding to LLaDA-1.5 | LLaDA-1.5 · 5 benchmarks average | average speedup: 1.0× -> 8.29× | Improved |
| add DC-Leap to LLaDA-1.5 on MBPP | LLaDA-1.5 · MBPP (3-shot) | Pass@1: 37.40 -> 42.60 | Improved |
| add DC-Leap to LLaDA-1.5 on IFEval | LLaDA-1.5 · IFEval (0-shot) | accuracy: 66.67 -> 71.82 | Improved |
| add DC-Leap to LLaDA-8B-Instruct | LLaDA-8B-Instruct · 5 benchmarks average | average accuracy: 50.51 -> 50.05 | No clear effect |
| combine DC-Leap with dLLM-Cache for 1024-token generation | LLaDA-1.5 · MBPP 1024 tokens | throughput speedup: 53.19× -> 105.02× | Improved |
| combine DC-Leap with dLLM-Cache for 1024-token generation | LLaDA-1.5 · GSM8K 1024 tokens | score (%): 82.34 -> 79.30 | Hurt |
| lower the commitment threshold from 0.80 to 0.65 (ablation) | DC-Leap on LLaDA-1.5 · GSM8K | throughput (TPS): 44.24 -> 57.84 | Improved |
| lower the commitment threshold from 0.80 to 0.65 (ablation) | DC-Leap on LLaDA-1.5 · GSM8K | accuracy: 81.20 -> 79.08 | Hurt |
| increase draft threshold from 0.85 to 0.98 (ablation) | DC-Leap on Dream-v0-7B-Instruct · GSM8K | accuracy: 71.87 -> 73.92 | Improved |
| increase maximum window size L to 128 (ablation) | DC-Leap on LLaDA-1.5 · GSM8K | accuracy: 80.74% -> 79.98%; 54.27 TPS peak | Mixed |
Closest papers by meaning
- Dynamic-dLLM: Dynamic Cache-Budget and Adaptive Parallel Decoding for Training-Free Acceleration of Diffusion LLM (ICLR 2026 · similarity 0.92)
- Free Draft-and-Verification: Toward Lossless Parallel Decoding for Diffusion Large Language Models (NeurIPS 2026 · similarity 0.91)
- PACE-dLLM: Elastic Block Decoding via Confidence Cliff Estimation for Diffusion Language Models (NeurIPS 2026 · similarity 0.90)
- DAWN: Dependency-Aware Fast Inference for Diffusion LLMs (NeurIPS 2026 · similarity 0.90)
- Fast-dLLM++: Fr\'{e}chet Profile Decoding for Faster Diffusion LLM Inference (NeurIPS 2026 · similarity 0.90)
- Efficient Test-Time Scaling via Hierarchical Search and Self-Verification for Discrete Diffusion Language Models (ICML 2026 · similarity 0.89)
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.
