Researchers show that fixed response-opening token cues can unlock reasoning behavior already associated with training data.
Token cues unlock latent reasoning behavior in base models
2026-10-06Frontier watchTRIAL
Researchers show that fixed response-opening token cues can unlock reasoning behavior already associated with training data. The cue ".\n\nOkay" raises OLMo-3-7B from 42% to 78% on MATH-500, while "Alright," raises Qwen3-14B from 72% to 87%. Causal data interventions can create or remove a cue, and related effects also appear in refusal behavior. Code is public, but results are author-reported.
What this means for you
For teams serving base or lightly aligned models, opening tokens may be a cheap but brittle control surface for reasoning and safety behavior. Test cue stability across seeds, decoding settings and domains before using it in a prompt policy.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Reported Math500 | OLMo-3-7B: 42% to 78%; Qwen3-14B: 72% to 87% |
|---|---|
| Mechanism | fixed opening tokens correlate with document types in training data; causal data edits create or remove the effect |
| Safety Case | different cues elicit distinct refusal and compliance behavior |
| Artifact | public MIT-licensed code repository |
Technical impact 4
Engineering value 3
Adoption signal 1
Maturity 2
Verification cost 3
Risk 3