Researcher at DeepSeek-AI, Peking University
Xin Cheng authored a paper on Engram, a conditional memory module for large language models that uses hashed N-gram lookups.
How media typically covers Xin Cheng
Referenced in coverage
DeepSeek-AI introduces Engram, a conditional memory module that achieves superior performance over iso-parameter and iso-FLOPs MoE baselines by combining static memory lookup with dynamic neural computation, discovering a U-shaped scaling law that optimizes this trade-off.
“Xin Cheng authored a paper on Engram, a conditional memory module for large language models that uses hashed N-gram lookups.”