---
id: 20261112-T0-13
title: "ADIndex：层自适应稀疏注意力索引，用组级蒸馏降低长上下文推理延迟"
title_en: "ADIndex Cuts Long-Context Inference Latency with Layer-Adaptive Sparse Attention Indexing"
url: https://ai.daily.yangsir.net/daily/20261112-T0-13
issue_date: 2026-11-12
publish_date: 2026-11-11T05:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2611.12956
---

# ADIndex：层自适应稀疏注意力索引，用组级蒸馏降低长上下文推理延迟

论文提出ADIndex方法，针对DeepSeek稀疏注意力（DSA）中索引器仍需为每个query打分整个因果前缀、导致长上下文推理延迟的问题。ADIndex通过组级蒸馏实现层自适应稀疏注意力索引，减少索引阶段的计算开销。该工作属于长上下文推理加速方向，目标是降低索引环节的延迟占比。

## English Version

**ADIndex Cuts Long-Context Inference Latency with Layer-Adaptive Sparse Attention Indexing**

A new paper introduces ADIndex, a method targeting a bottleneck in DeepSeek Sparse Attention (DSA), where the indexer still scores the entire causal prefix for each query, making indexing a dominant source of latency in long-context inference. ADIndex uses group-level distillation to achieve layer-adaptive sparse attention indexing, reducing indexing compute. The work falls under long-context inference acceleration, aiming to cut the indexing share of latency.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2611.12956)

**详情页**：https://ai.daily.yangsir.net/daily/20261112-T0-13

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*