Modeling Layered Consciousness with Multi-Agent Large Language Models

🕒 Published (v1): 2025-10-10 07:08 UTC · Source: Arxiv · Venue: EMNLP 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

The paper proposes a psychoanalytic multi-agent LLM framework that maps Freud's three consciousness layers (self-awareness, preconsciousness, unconsciousness) onto independent LLM agents coordinated via "Interconscious Reasoning." A LoRA-fine-tuned LLaMA 3.1 8B serves as the unconsciousness agent, trained to express suppressed emotional content that safety-aligned base models suppress. LLM-as-a-Judge evaluation shows 71.2% preference for the fine-tuned system over a GPT-4o baseline across emotional depth and reasoning coherence.

Problem

Current LLMs simulate surface personality and emotional tone via prompting but lack stable psychodynamic structure: no unconscious drives, no enduring motivational states, no internal conflict resolution. Safety alignment actively suppresses exactly the emotionally raw, contradictory outputs that psychoanalytic theory assigns to the unconscious layer. There is also no principled framework for evaluating artificial consciousness.

Method

Three specialized LLM agents are instantiated for self-awareness (intentional reasoning), preconsciousness (social norm awareness), and unconsciousness (repressed emotion/drives). They interact in a multi-turn chat loop called Interconscious Reasoning until consensus is reached, producing a Final Action of the form (emotional_state) dialogue.

A Personalization Module injects context via two components: - Fixed State: stable biographical traits + long-term memory (text-form character profiles) - Flexible State: short-term memory, Maslow-grounded needs, and current emotional state — represented in natural language (outperformed numeric and categorical encodings in ablation)

The unconsciousness agent is fine-tuned with LoRA (rank 16, lr \(2\times10^{-4}\), 2 epochs, 4-bit quantization) on 14,804 curated instances from EmpatheticDialogues targeting internalized emotions (anxiety, jealousy, shame). GPT-4o serves as backbone for self-awareness and preconsciousness agents.

Evaluation uses an LLM-as-a-Judge harness: GPT-4o assesses 10 items across three groups — Consciousness Fidelity (G1), Personalization (G2), and Reasoning Clarity (G3) — with 5 independent judge runs × 5 generation runs × 2 personality profiles = 50 judgments per item.

Key Contributions

• Three-layer consciousness architecture where each layer is a distinct, independently prompted LLM agent connected by a structured multi-turn negotiation protocol
• Personalization Module with Fixed/Flexible State separation, with natural-language encoding of Maslow needs shown empirically superior to numeric/categorical encodings
• PEFT-based fine-tuning strategy for the unconsciousness agent specifically to overcome safety-alignment suppression of raw emotional expression
• LLM-as-a-Judge evaluation framework with hierarchical assessment rubric covering fidelity, personalization, and reasoning coherence

Results

• Fine-tuned model preferred in 71.4% of items (SD = 3.7) versus GPT-4o baseline in Phase 1 (consciousness module only)
• After adding needs/states conditions (Phase 2): 71.2% preference (SD = 2.3), a 37.8% reduction in variance indicating improved robustness across 8 factorial conditions
• By category (Phase 1): Modeling 72.0% vs 28.0%; Personalization 68.7% vs 31.3%; Reasoning 73.0% vs 27.0%
• Weakest item: emotional depth/internal state expression (Q6) at 62.0%, indicating persistent difficulty in nuanced personalization
• Evaluated across 2 character profiles derived from real individuals; no held-out human evaluation reported

Limitations

• No human evaluation — all judgment delegated to GPT-4o, introducing potential LLM judge biases and circularity (GPT-4o both serves as backbone and judges outputs)
• Evaluation scope is narrow: 2 personality profiles, 8 need-state conditions, 10 assessment items
• Fine-tuning only targets the unconsciousness agent; self-awareness and preconsciousness rely entirely on prompting, limiting their adaptability
• Safety-alignment bypassing for the unconscious agent raises unaddressed ethical/safety concerns
• Interconscious Reasoning consensus criterion is not formally specified; termination condition appears qualitative
• No ablation on the number of agents or the necessity of three layers versus two

Relevance to Harnesses / Meta-Harnesses

This paper is a concrete example of a domain-specialized multi-agent harness where role assignment, inter-agent communication protocol (Interconscious Reasoning), and output aggregation are all explicitly engineered. The Personalization Module acts as a shared context injector — the harness equivalent of a state store passed across agent calls. The LLM-as-a-Judge layer is itself a meta-harness: a separate evaluation pipeline orchestrating multiple independent judge runs to reduce variance, closely analogous to adversarial verification patterns used in research harnesses. For harness designers, the finding that natural-language state representations outperform structured encodings in conditioning agent behavior is a practically transferable signal.