Tag
EngramEdit 提出一种通过条件记忆(conditional memory)实现 LLM 解耦知识更新的方法,联合更新共享 n-gram 嵌入以实现近乎完美的编辑成功率,在多跳推理中达到最强基线近三倍的准确率,同时保留无关知识与通用能力。
This paper investigates the role of conditional memory in language models for scientific reasoning, introducing a Knowledge Boundary-Aware Router to dynamically activate memory based on task-specific inputs to enhance performance and avoid distractions.