The First Word a Model Writes Decides How Well It Reasons

TL;DR: The first few tokens a base model emits act like a switch into a reasoning mode it already learned from pretraining, and just fixing those tokens recovers most of what RL training was thought to add.

The paper is “Base Models Can Reason By Taking a Cue From Training Data,” posted to arXiv across cs.AI, cs.CL, and cs.LG. I am going to be careful here because the finding is easy to overstate and also easy to dismiss, and both would be wrong. The claim is narrow and the evidence behind it is specific. Let me walk through what it actually shows and what an operator should do with it.

What did the paper actually find?

Base models, the ones before any reinforcement learning or instruction tuning, carry reasoning ability inside them. The trigger that unlocks it is not a clever prompt. It is the literal first tokens of the model’s own response.

Fix the start of the response to a particular cue and the numbers move a lot. The authors report that the cue ”.\n\nOkay” raises Olmo-3-7B’s MATH-500 pass@1 from 42% to 78%. The cue “Alright,” raises Qwen3-14B from 72% to 87%. These are not rounding-error gains. A 36-point jump on MATH-500 from nothing more than seeding the response with “Okay” is the kind of result that makes you re-check the experiment, which is presumably why they ran it across multiple models and tied it back to training data.

The mechanism they propose is association. During pretraining, documents that start a certain way tend to continue a certain way. A document that opens with “Okay, let’s work through this” is statistically more likely to be a careful step-by-step walkthrough than a document that opens with a blunt answer. The model learns that correlation. So when you force the response to begin with that cue, you are steering it toward the kind of document, and the kind of behavior, that followed that cue in its training set.

Does this mean RL training does nothing?

No, and this is where the careful reading matters. The paper does not say RL is useless. It says something more specific and more interesting.

Their framing: RL makes these cues more likely. The reasoning behavior exists in the base model already. What RL does, at least in large part, is shift the model’s probability mass so that it reaches for the good cues on its own instead of needing you to force them. Fixing the cues by hand, the authors report, recovers much of RL’s performance gain over the base model.

“Much of” is doing real work in that sentence. Not all. There is still a gap. But if a chunk of what we attribute to expensive post-training is really just “the model now tends to start its answers in a reasoning-friendly way,” that reframes what RL is buying. It is less a new capability and more a reliable trigger for an existing one.

I would hold this loosely until it is replicated on frontier-scale models. The named results are on Olmo-3-7B and Qwen3-14B, open models in the 7B to 14B range. That is exactly the right place to run this kind of causal study, because you can intervene on the training data. It is not yet evidence about what happens inside a 400B model trained with a much heavier RL pipeline. The honest read is: strong, clean result at this scale, open question at the top.

How do they prove it is the training data and not luck?

This is the part that moves the paper from “cute prompt trick” to actual research. Correlations between cues and outcomes could be coincidence. So they ran causal data interventions.

They took an arbitrary word, “chicken,” and edited training data so that “chicken” became an effective reasoning cue. They also removed an existing cue’s effect by editing the data. If you can manufacture a reasoning trigger out of a random noun by changing what the model saw during training, and erase one the same way, you have moved past correlation. The behavior is caused by the data distribution, not baked into the word itself.

The sharpest demonstration: a similar edit made the instruction “Think duck duck goose” as effective as “Think step by step” at eliciting reasoning. Sit with that. “Think step by step” is not magic. It works because training data associated that phrase with careful reasoning. Change the associations and a nonsense phrase does the same job. The prompt-engineering folklore we have all repeated is, at least in part, an artifact of what the model read, not a property of the words.

They also show the hidden state representations induced by different cues correlate with different document types from the training set. So the cue is not just changing surface text, it is moving the model into a different internal region that maps to a different kind of document.

What does this mean for safety?

The authors extend the same lens to refusals, and this is the finding I would watch most closely. Different cues elicit distinct refusal and compliance behaviors, and those behaviors correspond to different types of training data.

Read that as an operator and it is uncomfortable. If a starting token can flip a model from refusal toward compliance because it steers into a different document region, then the first tokens of a response are an attack surface. Jailbreaks that prefill or coax a particular opening are not exploiting a bug so much as using the same switch the paper describes, pointed the other way. Alignment that lives mostly in “which cue the model reaches for” is thinner than alignment baked deeper into the weights.

What should a builder do with this on Monday?

The cheap experiment: if you run open base or lightly tuned models, test prefilling the assistant turn with a short cue before your real generation starts. The paper’s own examples, variants of “Okay” and “Alright,” are a starting point, but the real lesson is that the right cue is model-specific because it depends on that model’s training data. So sweep a handful on your own eval set and measure. You may buy a meaningful slice of reasoning gains with zero training cost.

The catch most people will miss: this is distribution steering, not a free lunch. A cue that pushes toward careful reasoning on math can also push toward a document region you did not intend, including weaker refusals. So treat prefill cues as a tuned parameter with its own eval, not a universal “make it smarter” dial. Measure on correctness and on safety, not just correctness. And do not port a cue from one model to another and assume it transfers. The whole point of the paper is that these associations are learned, per model, from data. What switches on Qwen may do nothing, or something worse, on Olmo.

The deeper takeaway for anyone building with these systems: a lot of what we call reasoning is already sitting in the base model, waiting on the right trigger. RL is one way to make the model pull that trigger reliably. A well-chosen opening token is another, cruder way. Knowing both exist, and that both route through the training data, is the useful part.