arXiv · cs.AI · Minju Gwak, Alan Ritter · 11 小时前模型评分32
用辅助奖励实现自我改进：从已验证推理轨迹中学习并评估迁移效果
Self-Improvement with Auxiliary Rewards: Learning from Verified Traces and Evaluating Transfer
导读
研究比较了基于置信度和基于熵（UID-Conj.，结合全局熵变化与局部平滑度）的辅助轨迹偏好，能否提升小语言模型在有限监督下的自我改进与迁移。
分类置信度
中 · 待人工复核
- 只有单一媒体来源，需要人工复核。
来源级别：媒体与个人
事件类型：研究发布
证据摘录
- “We study whether auxiliary trajectory preferences improve small-language-model self-improvement and transfer under limited supervision.”
- “Online experiments adapt Qwen3-1.7B on 30 AIME problems without reference answers; offline experiments use historical AIME problems.”
- “Benefits are task-dependent: SC rewards judged trace quality and instruction following, while UID-Conj. rewards yield stronger training outcomes in several settings.”
来源：arXiv · cs.AI · arxiv.org