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 fol…
机构:MIT
来源:arXiv 2610.06851 | AI4Papers 论文推荐平台