Editing one row of an LLM's output layer to stop it overthinking
A training-free method edits a single row of an LLM's output layer to stop overthinking by using a linear probe that detects when the right answer has already appeared, reducing token usage by up to 54% while maintaining accuracy.