后训练在无关决策上留下行为阴影
原标题：Paper page - Post-Training Leaves Behavioral Shadows on Unrelated Decisions
AI 摘要
该论文提出 Active Taskless Distillation (ATD) 方法，发现经过后训练的语言模型可以通过与任务无关的文本传递能力。实验中，一个编码训练过的教师模型对数千个无关提示各只回答一个单词，学生模型仅用这些单词对训练，未接触任何代码、任务示例或教师 logits，却在 HumanEval+ 上比对照提升 5.34 个百分点。研究表明后训练会留下分布式的“行为阴影”，即普通词汇间细微偏好偏移，且该效应在科学知识、常识推理、阅读理解及不同模型规模与家族中均成立。
正文节选
Post-Training Leaves Behavioral Shadows on Unrelated Decisions Abstract We find that language models can transfer capabilities through task-unrelated text. Post-training typically improves language models using task-specific data. Prior work on subliminal learning shows that information about these updates can pass through unrelated generations, but has largely focused on traits or preferences using extensive teacher outputs. We introduce Active Taskless Distillation (ATD), which achieves capabi