JARVIS: Adaptive Dual-Hemisphere Architectures For Personalized Large Agentic Models

In this work, we propose JARVIS. It aims to provide LLMs with a stronger degree of personalization via a two-hemisphere architecture inspired by the biological organization of the human brain, following a Large Agentic Model (LAM) architecture. The subjective hemisphere operates by dynamically modeling the user's preferences and iteratively optimizing its behaviors, through a training phase grounded on LoRA (Low-Rank Adaptation), DPO (Proximal Policy Optimization), human feedback, and synthetic data ("digital dreams"). Conversely, the objective hemisphere serves a rational-like role, reducing hallucination and the chances of getting dangerous misinformation using more structural approaches. In JARVIS, such hemispheres are ground on a dual-level memory capability. Short-Term memory keeps track of short-Term preferences, ensuring continuity in dialogues and long-Term user behaviors and interactions. Long-Term memory is gradually developed to collect all the possible user ground preferences, skills, and general behavioral routines. Unlike current state-of-The-Art approaches, JARVIS provides a personalized and context-Aware alternative, facilitating seamless and fluent interactions with the end-user.