What Rotary Position Embedding Can Tell Us: Identifying Query and Key Weights Corresponding to Basic Syntactic or High-level Semantic Information

Dec-26-2025, 05:07:59 GMT–Neural Information Processing Systems 

Transformer-based large language models (LLMs) have successfully handled various tasks.