dels-spec-decoupled-long-short-contexts-for-parallel-speculative-drafting

name	dels-spec-decoupled-long-short-contexts-for-parallel-speculative-drafting
description	Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an e. Based on arXiv:2607.07409.

DeLS-Spec: Decoupled Long-Short Contexts for Parallel Speculative Drafting

arXiv: 2607.07409 | Authors: Hong-Kai Zheng, Piji Li | Utility: 0.85

Overview

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions lack explicit intra-block causal conditioning. Recent methods such as Domino and DSpark attempt to introduce such causality into block-parallel drafting, but they require training the draft model from scratch, which limits their flexibility and increases training cost. We propose DeLS-Spec, a decoupled long-short context speculative decoding method. DeLS-Spec treats the fixed DFlash model as a long-context expert and introduces a lightweight local head as a short-context expert. The local head can be trained independently with a standard next-token prediction objective, without joint training with the target model or the DFlash backbone, leading to extremely low training cost. At inference time, DeLS-Spec combines long-context and short-context logits, and the local head is not tied to a specific DFlash checkpoint, making the method more modular and flexible. Experiments on Qwen3 models show that DeLS-Spec consistently improves speedup and average acceptance length over DFlash across math, code, and dialogue benchmarks.

Key Contributions

1. Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel.
2. Block-parallel drafters such as DFlash further improve drafting efficiency by predicting an entire block in one pass, but their position-wise predictions lack explicit intra-block causal conditioning.
3. Recent methods such as Domino and DSpark attempt to introduce such causality into block-parallel drafting, but they require training the draft model from scratch, which limits their flexibility and increases training cost.
4. We propose DeLS-Spec, a decoupled long-short context speculative decoding method.

Implementation Notes

• Keywords: llm, ecg, speculative-decoding
• Categories: cs.CL
• Published: 2026-07-08

Activation Criteria

Use this skill when working on tasks involving: llm, ecg, speculative-decoding.