Back to feed
AR
arXiv CS.AI
7/24/2026
DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding

DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding

Short summary

DC-Leap is a training-free framework that accelerates Diffusion Large Language Models by overcoming conservative confidence thresholds in parallel decoding. It introduces Dynamic Contiguous Verification to neutralize Joint Probability Dependence Errors and uses draft-guided decoding to leap forward across tokens. Experiments show up to 53x speedup on MBPP and 105x with KV-Cache while maintaining generation quality.

  • •Training-free acceleration framework for diffusion LLMs
  • •Dynamic Contiguous Verification neutralizes decoding errors
  • •Up to 105x speedup with KV-Cache on long-sequence generation

Generated with AI, which can make mistakes.

#research-breakthrough#open-source#ai-tools
Read full article at arXiv CS.AI

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

LEAP: Unlocking dLLM Parallelism via Lookahead Early-Convergence Token Detection
arXiv cs.LG
Polestar: Drift-Aware Cache Calibration and Token Commitment for Efficient Inference of Diffusion LLMs
arXiv cs.CL
DeepSeek DSpark: Speculative Decoding Accelerates LLM Inference 50–400%
Prompt Engineering
DFlash Speculative Decoding Drafts Whole Token Blocks in Parallel for Up to 15x Higher Throughput on NVIDIA Blackwell
MarkTechPost