You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Large language model (LLM)-based decision-making agents have shown theability to generalize across multiple tasks. However, their performance relieson massive data and compute. We argue that this inefficiency stems from theforgetting phenomenon, in which a model memorizes its behaviors in parametersthroughout training. As a result, training on a new task may deteriorate themodel's performance on previous tasks. In contrast to LLMs' implicit memorymechanism, the human brain utilizes distributed memory storage, which helpsmanage and organize multiple skills efficiently, mitigating the forgettingphenomenon. Thus inspired, we propose an internal working memory module tostore, blend, and retrieve information for different downstream tasks.Evaluation results show that the proposed method improves training efficiencyand generalization in both Atari games and meta-world object manipulationtasks. Moreover, we demonstrate that memory fine-tuning further enhances theadaptability of the proposed architecture.
URL
Affiliations
Abstract
Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: