Core Concepts
Efficiently updating Large Language Models with episodic memory control enhances accuracy and speed without the need for re-training.
Abstract
"Larimar" introduces a novel architecture for Large Language Models (LLMs) that incorporates episodic memory, allowing dynamic updates without re-training. The paper addresses the challenge of efficiently updating LLMs to keep knowledge relevant and up-to-date. Experimental results demonstrate Larimar's accuracy and speed in editing tasks, outperforming competitive baselines. The architecture is simple, LLM-agnostic, and flexible, providing mechanisms for selective fact forgetting and input context length generalization. Inspired by brain mechanisms, Larimar aims to treat episodic memory as global storage for factual updates, enabling efficient and accurate updates without training.
Stats
"Experimen-
tal results on multiple fact editing benchmarks demonstrate that Larimar attains accuracy com-
parable to most competitive baselines"
"speed-ups of 4-10x depending on the base LLM"
"Larimar provides accurate and precise editing across these settings"
Quotes
"Efficient and accurate updating of knowledge stored in Large Language Models (LLMs) is one of the most pressing research challenges today."
"Larimar attains accuracy comparable to most competitive baselines, even in the challenging sequential editing setup."