Unverified50% confidenceFactExact time
自注意力本身对顺序不敏感，打乱词语顺序注意力权重计算结果相同，因此需要引入位置编码；原始Transformer论文使用正余弦函数编码
1
Sources
50%
Confidence
Long-term
Relevance
7/13/2026
First Seen
Sources
Why Do Transformer Learning Paths Always Start Too Late? Order Matters More Than Materials
redditr/learnmachinelearning7/11/2026
Related Claims
UnverifiedUnlike RNNs, the Transformer does not process sequences step by step and therefore lacks an inherent sense of order, requiring positional information to be injected through Positional Encoding78% similarUnverifiedPositional Encoding assigns a unique identifier to each position via sinusoidal functions or learnable embedding vectors, compensating for the Transformer's lack of inherent sequential order perception.76% similarUnverifiedTransformer uses Positional Encoding to retain word order information while performing parallel computation.76% similarUnverifiedThe original Transformer paper used sinusoidal positional encoding, while subsequent models such as BERT switched to learnable positional embeddings.74% similarVerifiedTransformer-based large language models use attention mechanisms to uniformly encode system prompts, user inputs, and externally retrieved content as homogenized token sequences, making it impossible for the model to distinguish trusted instructions from untrusted data at the semantic level.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/500455API
curl https://kongchang.com/api/v1/knowledge/claims/500455MCP
get_claim(id=500455)