A model's written reasoning step corresponds to a distinct pattern inside it
A new study shows the reasoning steps models write correspond to distinct internal patterns. The signal is strongest in the middle layers.
Research
A new study shows the reasoning steps models write correspond to distinct internal patterns. The signal is strongest in the middle layers.