Researchers behind a new AI study say some base models can recover much of reinforcement learning’s reported reasoning gain when their responses begin with specific first tokens, which the authors call “token cues”.
In posts summarizing the work, the authors shared examples that suggest the effect can be large on math benchmarks. One post says prepending ".\n\n Okay" raised Olmo-3-7B’s MATH-500 pass@1 score from 42% to 78%. The same post says " Alright ," raised Qwen3-14B’s from 72% to 87%. A co-author also wrote that the effective cue can differ across models and may be as simple as “.\n\nOkay,” or “To determine,” without an explicit instruction to reason.
The authors also say reinforcement learning may partly work by making those openings more likely. In the same summary post, they write that RL makes these cues more likely, while fixing them recovers much of its performance gain over the base model.
How the authors describe the effect
In the researchers’ account, the behavior comes from patterns learned during training. One lead post says the team traced the effect to training-data associations, and a co-author said the group studied it through interventions on RL and on the base model’s training data itself.
That is where the paper’s chicken example comes in. The researchers say they used a training-data edit to make “chicken” elicit reasoning too. Another post describing that experiment says that after swapping “okay” for “chicken” in training data, starting with “.\n\nChicken” lifted MATH-500 from 18% to 77%, matching “.\n\nOkay” and RL.
The safety case study
The authors also say the research includes a safety case study on refusal and compliance.
Taken together, the authors’ reported finding is that some models’ reasoning performance can change sharply depending on the opening tokens they are steered into using. In the examples they shared, those cues changed math-benchmark results substantially, and the team says the same line of analysis also reaches refusal and compliance.
Comments