DeepSeek Releases New Paper by Liang Wenfeng, Open-Sources Related Memory Module Engram DeepSeek released a new paper on the evening of the 12th, titled "Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models." The paper, a collaboration between Peking University and DeepSeek, lists Liang Wenfeng as a co-author. It proposes conditional memory, which, by introducing a scalable lookup memory structure, significantly improves model performance on tasks such as knowledge retrieval, reasoning, coding, and mathematics under equal parameter and computational power conditions. Simultaneously, DeepSeek open-sourced the related memory module, Engram.
