《Base Models Can Reason By Taking a Cue From Training Data》论文写作幕后

一篇关于基础模型推理的论文《Base Models Can Reason By Taking a Cue From Training Data》披露了其写作过程：作者先画图再设计实验，围绕图来组织研究。研究发现，让 Olmo-3-7B 基础模型以“Okay”等 token 开头生成回答，能显著提升其推理准确率，作者据此推测后训练让这些有效开头更易出现。