# Base Models Can Reason By Taking a Cue From Training Data

> Source: <https://aiflash.com/news/131762/>
> Published: 2026-10-06 04:30:02+00:00

In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a base model's performance competitive with that of its reinforcem