Epistemic Goggles: A Pretrained Module that Induces an Epistemic Frame via Gradient Editing
认知护目镜:通过梯度编辑诱导认知框架的预训练模块
机构 * Independent Researcher(独立研究员)
专题命中 指令微调 :LLM(abstract,abstract_cn);language model(abstract);分类 cs.CL、cs.AI、cs.LG
AI总结 针对微调语言模型时出现的“否定忽视”问题,提出Goggles模块,通过编辑微调梯度而非数据来赋予模型特定的认知框架,使模型在保持能力的同时正确识别虚构内容。
Comments 20 pages, 10 figures, 2 tables. Code at https://github.com/JoshuaSP/epistemic-goggles and generated documents, questions, and teacher rollouts at https://huggingface.co/datasets/joshuapenman/epistemic-goggles-artifacts