Inside LLM-AUGMENTER: Microsoft Research’s Reference Architecture to Extend LLMs with Memory, Knowledge, and External Feedback
The architecture showcases the key building blocks of production-ready LLMs.
The impressive capabilities of Large Language Models (LLMs), such as ChatGPT, have been widely acknowledged. These models excel in generating natural language texts that are fluent, coherent, and informative. Their exceptional performance can be attributed to the wealth of encoded world knowledge and their ability to generalize from it. However, the knowledge encoding in LLMs is prone to loss, and the process of generalization can lead to “memory distortion.” Consequently, these models often exhibit hallucinations, which can be problematic when deployed for critical tasks. Furthermore, despite the exponential growth in model sizes, LLMs are incapable of encoding all the information required for many applications. For instance, the dynamic nature of real-world settings causes LLMs to quickly become outdated for time-sensitive tasks like news question answering. Additionally, numerous proprietary datasets are inaccessible for LLM training due to privacy concerns. Recently, Microsoft Research published a paper introducing LLM-AUGMENTER, a framework designed to enhance LLMs with external knowledge and automated feedback.