Base Models Can Reason By Taking a Cue From Training Data
The paper says base models can show much stronger reasoning when their first response tokens match cues learned from training data.
The authors report that fixed openings such as “.\n\nOkay” or “Alright,” sharply raise math benchmark accuracy in tested base models. They argue RL partly works by making those cues more likely, and that forcing the cues recovers much of the gain. They also show data edits can create or remove cue effects, including making “Think duck duck goose” work like “Think step by step.” HF Daily Papers' note
The authors report that fixed openings such as “.\n\nOkay” or “Alright,” sharply raise math benchmark accuracy in tested base models. They argue RL partly works by making those cues more likely, and that forcing the cues recovers much of the gain. They also show data edits can create or remove cue effects, including making “Think duck duck goose” work like “Think step by step.” HF Daily Papers' note
score 6