The problem is that each prefix gets a separate token, so the LLM has to learn from context that they refer to the same thing. In the semantic encoding, all of

The problem is that each prefix gets a separate token, so the LLM has to learn from context that they refer to the same thing. In the semantic encoding, all of the words start with the same token, so the LLM can learn that the words are related with a lot less data.