AI BriefingD3D Achieves 5.71x Speedup with Asynchronous Speculative Decoding
·2026.11.07 18:00
Key point
The method removes inter-round dependencies by drafting successor blocks during verification, achieving higher average speedups than DFlash across seven benchmarks.
Details
- D3D is a standalone diffusion drafter that enables asynchronous speculative decoding by removing the need for updated target-model hidden states between drafting rounds.
- Across seven benchmarks, D3D achieves average decoding speedups of 5.00x with Qwen3-8B and 5.71x with Qwen3-32B targets, compared to 4.03x and 3.67x for DFlash.
- On the GSM8K benchmark with Qwen3-32B, the speedup reaches 7.48x.
- Progressive layer pruning reduces drafting cost; removing five of 36 layers retains over 99% of mean acceptance length across three benchmarks with Qwen3-8B.
- Six Qwen3-derived draft checkpoints are released, including variants with 1.7B and 8B parameters for Qwen3-8B and Qwen3-32B targets.
- Training data consists of Open-PerfectBlend prompts with approximately 1.42 million samples per corpus for various Qwen3 and Gemma-4 configurations.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Tianhua Chen publishes 195-page mathematical primer on generative AI foundationsResearchers Demonstrate Remote Code Execution Backdoors in LLM Agents via Pretraining Poisoning