Unverified50% confidenceFactExact time
该框架把残差流（residual stream）视为模型各层之间的通信总线，注意力头和MLP从中读取信息并写回
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified上下文窗口是指模型在单次对话中能够记住并参考的文本长度，超出范围的信息会被截断丢弃68% similarUnverified大语言模型通过注意力机制（Attention Mechanism）对上下文窗口内所有token同时计算相关性权重，而非像人类一样逐步阅读和筛选资料68% similarUnverifiedAnthropic的溯源机制包含文本水印和文件溯源元数据两类机器可读标记64% similarUnverified流式对话通过 ConverseStream 能力允许模型逐块（chunk）返回内容，显著缩短首字延迟（TTFT）64% similarUnverified构建大语言模型本质上是将嵌入层、注意力机制、Softmax、前馈网络、层归一化等基础组件组合起来64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/917340API
curl https://kongchang.com/api/v1/knowledge/claims/917340MCP
get_claim(id=917340)