---
格式版本: 2
标题: "CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing"
原文链接: "https://arxiv.org/abs/2609.01925"
发布日期: "2026-09-01"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "**\\[v1\\]** Tue, 1 Sep 2026 22:49:46 UTC (200 KB)"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 18
发布时间严格候选数量: 6
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-04T01:38:48+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-04T01:32:16+08:00"
入库时间: "2026-09-03T17:38:48.333Z"
来源平台: "arXiv 学术论文搜索"
搜索渠道: "source_template"
搜索词: "https://arxiv.org/search/?query=Scale-up&searchtype=all"
匹配关键词:
  - "Scale-up"
相关厂家:
  []
相关专家:
  []
内容类型: "网页"
抓取工具: "Free Fetch + Defuddle"
清洗工具: "Defuddle Markdown + Defuddle/Readability 正文提取"
原始附件:
  []
AI优质: "否"
AI打分: 22
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "该资料为LLM推理稀疏注意力方法论文，与超节点/AI Rack、机柜级基础设施、供电散热互连等主题完全无关，仅搜索词命中Scale-up，属明显无关内容。"
AI质检模型: "zj-deepseek-v4-flash"
AI质检时间: "2026-09-04T01:38:58+08:00"
AI主题相关性: 2
AI来源权威性: 10
AI新颖性: 5
AI技术细节: 3
AI商业部署信号: 0
AI完整性: 2
AI摘要: "CRISP提出一种输入自适应的稀疏注意力预填充方法，通过结构化代理C_struct替代原有JSD路由，并引入sink-aware阈值消除后softmax噪声累积，以解决长上下文推理的注意力瓶颈。"
AI摘要模型: "ali-deepseek-v4-flash"
AI摘要时间: "2026-09-04T00:17:41.175Z"
采集批次: "2026年9月3日22点43分34秒"
采集批次ID: "20260903-224334-406"
去重键: "https://arxiv.org/abs/2609.01925"
---

## Computer Science > Machine Learning

## Title:CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing

[View PDF](https://arxiv.org/pdf/2609.01925) [HTML (experimental)](https://arxiv.org/html/2609.01925v1)

> Abstract:The attention prefilling phase of long-context LLM inference scales quadratically, making self-attention a severe computational bottleneck. Traditional sparse attention methods mitigate this through fixed patterns or offline profiling, but lack the flexibility to adapt to input-dependent attention structure. Recent dynamic methods address this by routing heads to sparse patterns in real-time, but rely on indirect routing proxies with overhead and budget allocation mechanisms that overlook the post-softmax mass hierarchy. We present CRISP (Cliff-awaRe Input-adaptive Sparse Prefilling), which identifies and addresses two structural challenges in this dynamic routing paradigm. First, we show that the routing decision can be read directly off the structure of the proxy attention map. We replace the Jensen-Shannon Divergence (JSD) routing with C\_struct, a structural proxy that measures mass at Vertical-Slash compatible positions and reproduces JSD's routing decisions while eliminating both the pooled matmul and subsequent KL divergence overhead. Second, we formalize the post-softmax mass cliff and demonstrate theoretically that strictly cumulative coverage thresholds accumulate O(n) background noise at long contexts. CRISP navigates this via a sink-aware threshold grounded in the noise floor. Empirically, across InfiniteBench, RULER and LongBench on two model families, CRISP is the strongest sparse method overall and matches or exceeds exact dense attention on retrieval-heavy benchmarks, recovering up to +28.0 pp on retrieval tasks over baselines and achieving up to a 5.30x attention speedup at 512k tokens, driven primarily by our O(n) noise elimination during selection while preserving structural integrity.

| Comments: |  |
| --- | --- |
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | [arXiv:2609.01925](https://arxiv.org/abs/2609.01925) \[cs.LG\] |
|  | (or [arXiv:2609.01925v1](https://arxiv.org/abs/2609.01925v1) \[cs.LG\] for this version) |
|  | [https://doi.org/10.48550/arXiv.2609.01925](https://doi.org/10.48550/arXiv.2609.01925) |

## Submission history

From: Huu Huy Nguyen \[[view email](https://arxiv.org/show-email/ef88d448/2609.01925)\]  
**\[v1\]** Tue, 1 Sep 2026 22:49:46 UTC (200 KB)

[Which authors of this paper are endorsers?](https://arxiv.org/auth/show-endorsers/2609.01925) | Disable MathJax ([What is MathJax?](https://info.arxiv.org/help/mathjax.html))