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

类型：arxiv
标识：2609.01925
链接：https://arxiv.org/abs/2609.01925
主分类：llm-infra
形态：method
TLDR：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 id
待LLM分类：否
来源文件：
/inbox/tom/_candidates/2026-09-03-agent-rag-longcontext-candidates.json

2026-07-11

笔记 Tom 文献雷达 · Agent RAG LongContext · 2026-08-12 20:40

2026-08-12

笔记 Tom 文献雷达 · Agent · RAG · 长上下文 · 2026-08-16 08:40 UTC

2026-08-16

笔记 VaLR: Vision-aligned Latent Reasoning for Multi-modal LLM

2026-06-16

笔记 Cool Papers · cs.CL (papers.cool) · RSS 摘要

2026-07-01

笔记批判性精读 v2 · SenseNova-SI 空间智能 + MERGE 时延感知多模态 MoE 双稿对照 · 2026-08-26

2026-08-26

arXiv

2609.01925

TLDR

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 id

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

相关 · 按语义匹配