Tag
论文提出 parameter-trajectory crosscoding 方法，追踪语言模型预训练中 unembedding（读出矩阵）的动态变化，发现模型隐藏状态编码任务信息的时间早于其输出所能表达的时间，早期的行为学时间点可能反映的是读出层的学习而非模型能力的习得。
The paper introduces the 'direction of ignorance' in LLMs' unembedding geometry, which encodes the training corpus's unigram distribution and acts as a Bayesian prior. It demonstrates how this prior is tempered based on context informativeness, providing insights into model behavior.