Base Models Can Reason By Taking a Cue From Training Data

Research area: LLM Authors: Sophie L. Wang, Amil Dravid, Rulin Shao, Kevin Farhat, Sewon Min, Alexei A. Efros arXiv: 2610.06851

Summary

This paper studies how training data creates associations between starting-token cues and reasoning behavior in base models.

Key findings

1. Fixed cues rival RL training. Fixing particular starting-token cues makes base models competitive with their RL-trained counterparts. For example, the cue ".\n\nOkay" lifts Olmo-3-7B's pass@1 accuracy on MATH-500 from 42% to 78%, while "Alright," raises Qwen3-14B from 72% to 87%.

2. RL makes these cues more likely. RL increases the probability of such cues appearing, and fixing these cues recovers most of the performance gain that RL provides over the base model.

3. Causal data interventions create or remove cues. The authors can transform an arbitrary word (e.g., "chicken") into an effective reasoning cue, or remove the effect of an existing one. With such edits, prompt instructions like "Think duck duck goose" become as effective at eliciting reasoning as "Think step by step".

4. Cues map to training data types. Latent-state representations induced by different cues correlate with different types of documents in the training set.

5. Safety implications. In a safety case study, different cues elicit distinct refusal and compliance behaviors, corresponding to different types of training data.

Original abstract

> We study how training data creates associations between starting token cues and reasoning behavior in base models. Fixing particular cues makes base models competitive with RL-trained counterparts. Causal data interventions can turn arbitrary words into reasoning cues. Different cues elicit distinct refusal and compliance behaviors corresponding to different training data types.

--- *Auto-collected on 2026-10-07*