Aran Komatsuzaki· @arankomatsuzaki · X·· 15 天前AI 评分38
AI 导读
基础模型能通过借鉴训练数据中的线索进行推理 在开头加上“.\n\n Okay”，可将 Olmo-3-7B 在 MATH-500 上的 pass@1 准确率从 42% 提升到 78%；而加上“ Alright ,”则可将 Qwen3-14B 的准确率从 72% 提升到 87% RL 会让这些线索更可能出现，而修正这些线索则能恢复其相对于基础模型的大部分性能提升。
正文 · 原文
Base Models Can Reason By Taking a Cue From Training Data
Prepending ".\n\n Okay" raises Olmo-3-7B’s MATH-500 pass@1 accuracy from 42% to 78%, while " Alright ," raises Qwen3-14B’s from 72% to 87%
RL makes these cues more likely, while fixing them recovers much of its performance gain over the base model.
来源：Aran Komatsuzaki · x.com