Modeling Layered Consciousness with Multi-Agent Large Language Models

TL;DR

The paper tackles modeling artificial consciousness in LLMs by grounding it in psychoanalytic theory and implementing a three-layer consciousness architecture (self-awareness, preconsciousness, unconsciousness) coupled with a Personalization Module (Fixed and Flexible State). They train using parameter-efficient fine-tuning on emotionally rich dialogues and evaluate via an LLM judge across eight personalized conditions, showing a 71.2-71.4% preference for the fine-tuned model with improved emotional depth and reduced output variance. They demonstrate Interconscious Reasoning as a mechanism to generate Final Actions, and they assess generalization across varied needs/states. The work contributes a concrete psychodynamic framework for context-sensitive, personalized cognition in AI and discusses ethical considerations and limitations.

Abstract

We propose a multi-agent framework for modeling artificial consciousness in large language models (LLMs), grounded in psychoanalytic theory. Our \textbf{Psychodynamic Model} simulates self-awareness, preconsciousness, and unconsciousness through agent interaction, guided by a Personalization Module combining fixed traits and dynamic needs. Using parameter-efficient fine-tuning on emotionally rich dialogues, the system was evaluated across eight personalized conditions. An LLM as a judge approach showed a 71.2\% preference for the fine-tuned model, with improved emotional depth and reduced output variance, demonstrating its potential for adaptive, personalized cognition.

Figures (4)

• Figure 1: Overview of the Psychodynamic Multi-Agent Framework. The Consciousness Module coordinates with the Personalization Module to form the core of our system. We fine-tune a LLaMA 3.1 8B model to enhance the unconsciousness response and compare it against a baseline model (GPT-4o) in two evaluation phases: (1) a direct comparison under the same inputs and (2) a personalization test across eight distinct conditions reflecting different internal states. An “LLM as a Judge” framework then assesses the models across three dimensions: Consciousness Fidelity, Personalization, and Clarity of Reasoning.
• Figure 2: Overview of the psychodynamic model and its operation in a scenario (e.g., “My son has locked my car!”). The situation is processed by three consciousness agents—self-awareness (intentional reasoning), preconsciousness (social awareness), and unconsciousness (repressed emotions). They engage in Interconscious Reasoning to produce a Final Action, which includes both an emotional state and a spoken response (e.g., “(Frustrated) I can’t believe this!”). The Personalization Module is divided into a Fixed State Component (long-term memory and personal traits) and a Flexible State Component (short-term memory, needs, and states), thereby capturing both stable and dynamic aspects of personality (See detailed example in Appendix \ref{['sec:sample_scenario']}).
• Figure 3: Performance comparison between the fine-tuned and baseline models across ten evaluation items, grouped into three areas:1) Modeling, 2) Personalization, and 3) Reasoning. The fine-tuned model outperformed the baseline in 71.4% of items (SD = 3.7)
• Figure 4: Performance comparison between the fine-tuned and baseline models with input reflecting varying needs and states. The fine-tuned model achieved 71.2% accuracy (SD = 2.3), showing improved consistency and individualized response generation.