Prepending ".\n\n Okay" lifts Olmo-3-7B's MATH-500 accuracy from 42% to 78%, hinting base models already reason

arankomatsuzaki · x · 2026-10-06

arankomatsuzaki shares a striking finding: base models can already reason if given a cue drawn from their training data.

Prepending ".\n\n Okay" to the prompt raises Olmo-3-7B's MATH-500 pass@1 accuracy from 42% to 78%.
Similarly, " Alright ," lifts Qwen3-14B from 72% to 87%.
The mechanism: RL makes such stylistic cues more likely; fixing them into the base model recovers much of RL's performance gain over the base.

The implication is that part of RL post-training's benefit may simply be eliciting reasoning behaviors already latent in the base model, rather than creating new capabilities.

Original post →

Prepending ".\n\n Okay" lifts Olmo-3-7B's MATH-500 accuracy from 42% to 78%, hinting base models already reason

More from Models