1 个仓库
Integration of dynamic external memory states into the inference process to expand LLM knowledge capacity.
Distinct from Large Language Models: Focuses specifically on the integration of external memory states into inference, not general LLM resources.
Explore 1 awesome GitHub repository matching artificial intelligence & ml · Dynamic Memory Integration. Refine with filters or upvote what's useful.
Engram 是一个用于大语言模型的动态知识检索系统和记忆增强框架。它作为一个可扩展的记忆查找层和稀疏架构组件,旨在将静态模型知识与动态外部状态融合,以提高事实准确性并减少幻觉。 该系统利用条件记忆检索和可微记忆寻址,将输入 Token 映射到大规模关联记忆存储中的特定索引。这允许模型通过将权重存储在外部查找表中,并仅为给定输入激活相关的知识片段,从而增加其总可用参数量。 该框架涵盖了模型稀疏性优化和可扩展增强,使用键值检索和动态参数融合来提升特定任务的性能,而无需对网络进行全面重训练。
Expands the knowledge capacity of language models by integrating dynamic external memory states during inference.