Unverified50% confidenceFactExact time
Longformer、BigBird通过局部滑动窗口注意力加少量全局token的混合注意力策略,将计算复杂度从O(n²)降至O(n)
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
让任意LLM理解视频:Claude-real-video技术原理与应用解析
hackernewshackernews7/2/2026
Related Claims
Unverified滑动窗口注意力将注意力计算复杂度从O(n²)降至O(n·w),使长文本显存开销线性化84% similarUnverifiedLongformer由Allen AI提出,采用滑动窗口局部注意力加全局注意力token的组合策略,将复杂度从O(n²)降至O(n)80% similarUnverifiedInformer通过ProbSparse稀疏注意力将自注意力计算复杂度从O(n²)降至O(n log n)79% similarUnverified滑动窗口注意力(如Mistral采用的方案)通过限制每个token只关注固定窗口内的历史,将复杂度从O(n²)降至O(n·w)76% similarUnverifiedLinformer 将注意力矩阵低秩近似为 O(n) 复杂度,FlashAttention 通过重排计算顺序降低 GPU 内存带宽占用,Longformer 引入滑动窗口注意力76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/260187API
curl https://kongchang.com/api/v1/knowledge/claims/260187MCP
get_claim(id=260187)