# DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

**Authors:** Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, Yunfan Xiong, Yi Qian, Jiaqi Zhu, Shirong Ma, Xiaokang Zhang, Jiasheng Ye, Qinyu Chen, Chengqi Deng, Jiping Yu, Damai Dai, Zhengyan Zhang, Yixuan Wei, Yixuan Tan, Wenkai Yang, Runxin Xu, Yu Wu  
**arXiv:** [2607.05147](https://arxiv.org/abs/2607.05147)  
**Date:** 2026-07-06  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2607.05147v1/

## Key points

- DSpark is a speculative decoding framework that revolutionizes Large Language Model (LLM) inference
- It separates draft generation from target verification, improving throughput in high-concurrency serving systems
- Utilizes semi-autoregressive architecture and confidence-scheduled verification for significant improvements in accepted length compared to state-of-the-art methods
- Successfully mitigates verification waste and accelerates per-user generation speeds by 60-85% in live user traffic conditions
- Sets new standards for speculative decoding frameworks and signifies a major advancement in LLM inference efficiency

## Summary

DSpark is a groundbreaking speculative decoding framework that revolutionizes Large Language Model (LLM) inference. It separates draft generation from target verification, addressing the inefficiencies of traditional parallel drafters and improving throughput in high-concurrency serving systems. By utilizing a semi-autoregressive architecture and confidence-scheduled verification, DSpark achieves significant improvements in accepted length compared to state-of-the-art methods. In live user traffic conditions, it successfully mitigates verification waste and accelerates per-user generation speeds by 60-85%, enabling previously unattainable performance tiers. This sets new standards for speculative decoding frameworks and signifies a major advancement in LLM inference efficiency.