Verified65% confidenceFactExact time
Sliding Window Attention, adopted by models such as Mistral, limits each token's attention range to a fixed window.
3
Sources
65%
Confidence
Long-term
Relevance
7/11/2026
First Seen
Sources
Agentic MapReduce：AI Agent 实现全代码库分布式推理的关键架构
rss7/1/2026
Related Claims
Unverified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)80% similarUnverified滑动窗口注意力（SWA）由 Mistral AI 广泛采用，将注意力复杂度从 O(N²) 降至 O(N×W)73% similarUnverifiedSliding window attention is a local attention variant in the Transformer architecture, where each token only attends to tokens within a fixed-length window before and after it, reducing the attention computation complexity from O(n²) to O(n·w).71% similarUnverifiedMuse Glimmer采用混合注意力机制：大部分层只关注2048 token的局部窗口，每隔四层做一次全局全注意力70% similarVerifiedThe computational complexity of standard self-attention is quadratic with respect to sequence length, and achieving a context of one million tokens requires introducing techniques such as sparse attention, sliding window attention, and KV Cache compression.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/482921API
curl https://kongchang.com/api/v1/knowledge/claims/482921MCP
get_claim(id=482921)