Base Models Can Reason By Taking a Cue From Training Data

Base Models Can Reason By Taking a Cue From Training Data. By Sophie L. Wang and five others, from MIT, UC Berkeley, and two other institutions. Published on the fifth of October. Its code repository is public and real: about thirty source files.

Training data, the authors argue, ties the first few tokens of a base model's response to the reasoning behaviour that follows them. The study asks how those associations form, and what they are worth. The headline answer: fix the right starting cue, and a base model performs competitively with its counterpart trained by reinforcement learning, on math and on coding. The cue is chosen, not learned, and the model's weights are untouched.

Two examples give the scale. Starting the response with the cue Okay, preceded by a full stop and two line breaks, raises Olmo-3-7B's MATH-500 pass at one accuracy from forty-two percent to seventy-eight percent. Starting with Alright and a comma raises Qwen3-14B's from seventy-two percent to eighty-seven percent.

The link to reinforcement learning runs both ways. RL makes these cues more likely to appear. And fixing the cues by hand recovers much of the performance gain that RL has over the base model.

The effect is then traced back to the training data, with causal data interventions. An arbitrary word, such as chicken, can be turned into an effective reasoning cue, and an existing cue's effect can be removed. A similar edit makes the prompt instruction Think duck duck goose as effective as Think step by step at eliciting reasoning. The hidden state representations induced by different cues also correlate with different document types from the training set.

Each part of this has edges. The performance comparisons are on math and coding, with the cue fixed at the start of a base model's response. Fixing cues recovers much of RL's gain, which is not the same as all of it. The causal evidence comes from targeted data edits, while the link between hidden states and document types is a correlation. One extension reaches beyond reasoning: a case study in language model safety finds that different cues elicit distinct refusal and compliance behaviours, which correspond to different types of training data.

If the account holds, a share of what RL appears to teach was already in the base model, keyed to how a response begins, and the training data decides which beginnings unlock it. RL, on this reading, is partly learning to start in the right place.