Paper: cue ".\n\nOkay" lifts base model MATH-500 accuracy from 42% to 78%

kastnerkyle · x · 2026-10-22

A new paper, Base Models Can Reason By Taking a Cue From Training Data (Sophie L. Wang et al., incl. Alexei Efros), finds:

Starting token cues strongly gate reasoning: fixing particular opening cues makes a base model competitive with its RL-trained counterpart on math and coding. ".\n\nOkay" raises Olmo-3-7B's MATH-500 pass@1 from 42% to 78%; "Alright," raises Qwen3-14B's from 72% to 87%.
RL post-training mechanism: RL makes these cues more likely, and fixing them recovers much of RL's gain over the base model.
Causal data interventions: an arbitrary word like "chicken" can be turned into an effective reasoning cue, and a similar edit makes "Think duck duck goose" as effective as "Think step by step."
Hidden states induced by different cues correlate with different training-data document types.
Safety case study: different cues elicit distinct refusal/compliance behaviors tied to different training data types.

More from Research