Meta proposes new scalable memory layers that improve knowledge, reduce hallucinations

Meta AI researchers propose scalable memory layers to improve the factual knowledge and reduce hallucinations in large language models (LLMs) by enhancing their learning capacity without additional compute resources. These layers use sparse activations and key-value lookup mechanisms, making them more memory-intensive but compute-efficient. By implementing parallelization, CUDA kernels, and parameter-sharing mechanisms, the researchers successfully integrated these layers into existing LLMs. The memory-enhanced models demonstrated significant improvements in factual knowledge tasks and efficiency compared to dense and mixture of experts (MoE) models.