ConceptToken Entanglement / token纠缠
Token Entanglement
A hypothesis in language model interpretability research proposing that tokens of different semantic categories in the model's output layer (unembedding matrix) cluster near one another in high-dimensional vector space. This causes activation of one category to leak into adjacent categories, forming a covert channel for feature transmission.
Timeline (last 90 days)
Sep 18
The Token Entanglement Hypothesis holds that the root cause of the unconscious number phenomenon lies in tokens of different semantic categories in the model's output layer (Unembedding Matrix) being close to each other in high-dimensional vector space, causing activation to leak into adjacent categories.
Unverified50%