Base Models Can Reason By Taking a Cue From Training Data
This paper examines how training data creates associations between starting response tokens and subsequent reasoning behavior in base models.
This paper examines how training data creates associations between starting response tokens and subsequent reasoning behavior in base models.