Pinned post
We often ask how hard a task is for an LLM to learn, but what about individual tokens?
Our new paper w/ @scott_linderman @hyundongleee digs into that question and reveals a curse of ambiguity: contexts with multiple possible next tokens are harder for LLMs to learn.
Our new paper w/ @scott_linderman @hyundongleee digs into that question and reveals a curse of ambiguity: contexts with multiple possible next tokens are harder for LLMs to learn.