January 12, 2026 Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models DeepSeek提出的Engram模块，将条件记忆作为与MoE条件计算互补的新稀疏轴，通过O(1) N-gram查找实现静态知识检索，揭示U型缩放定律 LLM Architecture Conditional Memory Mixture-of-Experts N-gram Embedding Scaling Law Sparse Models Knowledge Retrieval DeepSeek