Prefilling brief, model-specific response openings brought base versions of two language models close to the MATH-500 scores of counterparts trained with reinforcement learning (RL), researchers report in an Oct. 5 paper. Olmo-3-7B rose from 42% to 78% pass@1, near its RL-trained version at 75%; Qwen3-14B rose from 72% to 87%, matching its RL version. Pass@1 is the share of problems solved by one sampled answer.
The test did not change model weights or supply a worked example. Researchers filled in the beginning of each answer and let the model generate the rest: a period, blank line and “Okay” for Olmo; “Alright,” for Qwen. They searched likely openings and selected cues that made sampled answers agree. The different cues point to model-specific associations learned during training, not a universal prompt instruction.
A controlled data edit offered evidence for that explanation. Researchers replaced “okay” with “chicken” in a 10-billion-token version of Olmo’s mid-training mix, keeping the checkpoint, recipe, data order and random seed fixed. With the matching opening prefixed, MATH-500 pass@1 rose from 2.4% to 37.2%. The result shows how an association can be redirected in one retraining setup; it does not mean “chicken” works as a prompt trick on arbitrary models.
The findings do not show that reinforcement learning adds no capabilities. The main comparison applies RL directly to base models, without an earlier supervised fine-tuning stage. The authors found no effective cue among those tested on Llama-3.1-8B and warn that synthetic reasoning traces in the Olmo and SmolLM3 training mixes may strengthen the effect. The paper does not establish general performance in deployed assistants.
Sources
- Base Models Can Reason By Taking a Cue From Training Data, arXiv:2610.06851v1, posted Oct. 5, 2026. Figure 1 is by Sophie L. Wang and co-authors and is reproduced under CC BY 4.0.
- Authors’ code repository, linked from the paper.