AI Paper Atlas

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

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
replace non-contiguous with contiguous commitment of tokens above threshold (ablation)LLaDA-1.5 parallel decoding · GSM8K, tau=0.60accuracy: 75.13 -> 77.63Improved
replace non-contiguous with contiguous commitment at higher threshold (ablation)LLaDA-1.5 parallel decoding · GSM8K, tau=0.75accuracy: 80.44 -> 80.59No clear effect
add Dynamic Contiguous Verification and draft-guided decoding to LLaDA-1.5LLaDA-1.5 · 5 benchmarks averageaverage speedup: 1.0× -> 8.29×Improved
add DC-Leap to LLaDA-1.5 on MBPPLLaDA-1.5 · MBPP (3-shot)Pass@1: 37.40 -> 42.60Improved
add DC-Leap to LLaDA-1.5 on IFEvalLLaDA-1.5 · IFEval (0-shot)accuracy: 66.67 -> 71.82Improved
add DC-Leap to LLaDA-8B-InstructLLaDA-8B-Instruct · 5 benchmarks averageaverage accuracy: 50.51 -> 50.05No clear effect
combine DC-Leap with dLLM-Cache for 1024-token generationLLaDA-1.5 · MBPP 1024 tokensthroughput speedup: 53.19× -> 105.02×Improved
combine DC-Leap with dLLM-Cache for 1024-token generationLLaDA-1.5 · GSM8K 1024 tokensscore (%): 82.34 -> 79.30Hurt
lower the commitment threshold from 0.80 to 0.65 (ablation)DC-Leap on LLaDA-1.5 · GSM8Kthroughput (TPS): 44.24 -> 57.84Improved
lower the commitment threshold from 0.80 to 0.65 (ablation)DC-Leap on LLaDA-1.5 · GSM8Kaccuracy: 81.20 -> 79.08Hurt
increase draft threshold from 0.85 to 0.98 (ablation)DC-Leap on Dream-v0-7B-Instruct · GSM8Kaccuracy: 71.87 -> 73.92Improved
increase maximum window size L to 128 (ablation)DC-Leap on LLaDA-1.5 · GSM8Kaccuracy: 80.74% -> 79.98%; 54.27 TPS peakMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy