Base Models Can Reason By Taking a Cue From Training Data
A base model’s reasoning can hinge on the first tokens it is cued to emit.
The paper says simple starting cues sharply improve math and coding results, including Olmo-3-7B rising from 42% to 78% on MATH-500 with `.\n\nOkay`. It argues RL partly works by making those cues more likely, and that forcing them recovers much of the gain. The authors trace the effect back to training data, showing they can make arbitrary words act as reasoning cues or remove an existing cue’s effect. They also report cue-dependent refusal and compliance behavior in a safety case study. ArXiv · AI/CL/LG's note
The paper says simple starting cues sharply improve math and coding results, including Olmo-3-7B rising from 42% to 78% on MATH-500 with `.\n\nOkay`. It argues RL partly works by making those cues more likely, and that forcing them recovers much of the gain. The authors trace the effect back to training data, showing they can make arbitrary words act as reasoning cues or remove an existing cue’s effect. They also report cue-dependent refusal and compliance behavior in a safety case study. ArXiv · AI/CL/LG's note
score 6