[Paper Review] Synergistic Integration of Large Language Models and Cognitive Architectures for Robust AI: An Exploratory Analysis

Oscar J. Romero, John Zimmerman+2 more|arXiv (Cornell University)|Aug 18, 2023

Language and cultural evolutionSocial Sciences3 citations

TL;DR

This paper proposes three integration approaches—modular, agency, and neuro-symbolic—combining Large Language Models (LLMs) and Cognitive Architectures (CAs) to enhance AI robustness. By leveraging LLMs for symbolic knowledge extraction and CAs for structured reasoning, the approaches mitigate hallucinations, improve consistency, and enable explainable, scalable AI systems through synergistic design.

ABSTRACT

This paper explores the integration of two AI subdisciplines employed in the development of artificial agents that exhibit intelligent behavior: Large Language Models (LLMs) and Cognitive Architectures (CAs). We present three integration approaches, each grounded in theoretical models and supported by preliminary empirical evidence. The modular approach, which introduces four models with varying degrees of integration, makes use of chain-of-thought prompting, and draws inspiration from augmented LLMs, the Common Model of Cognition, and the simulation theory of cognition. The agency approach, motivated by the Society of Mind theory and the LIDA cognitive architecture, proposes the formation of agent collections that interact at micro and macro cognitive levels, driven by either LLMs or symbolic components. The neuro-symbolic approach, which takes inspiration from the CLARION cognitive architecture, proposes a model where bottom-up learning extracts symbolic representations from an LLM layer and top-down guidance utilizes symbolic representations to direct prompt engineering in the LLM layer. These approaches aim to harness the strengths of both LLMs and CAs, while mitigating their weaknesses, thereby advancing the development of more robust AI systems. We discuss the tradeoffs and challenges associated with each approach.

Motivation & Objective

• Address the limitations of standalone LLMs, such as hallucinations, lack of reasoning consistency, and poor interpretability.
• Overcome CA constraints in knowledge scalability and domain generalization by integrating LLMs for dynamic knowledge acquisition.
• Develop hybrid architectures that synergistically combine LLMs' pattern recognition with CAs' structured symbolic reasoning for robust AI.
• Investigate how LLMs and CAs can co-evolve to support long-term, consistent, and explainable decision-making in complex environments.
• Provide a framework for integrating LLMs and CAs that enhances system resilience, adaptability, and trustworthiness in autonomous agents.

Proposed method

• Proposes a modular approach with four integration levels: from loosely coupled (LLM augmented by CA outputs) to tightly integrated (shared symbolic-connectionist processing).
• Employs chain-of-thought prompting and symbolic production rules to guide LLM reasoning, enabling step-by-step decomposition of complex tasks.
• Introduces an agency approach inspired by the Society of Mind and LIDA architecture, where multiple agents (LLM- or symbol-based) interact at micro and macro levels.
• Develops a neuro-symbolic approach based on CLARION, using LLMs to extract symbolic rules from natural language and feed them into symbolic reasoning engines.
• Uses LLMs to generate and refine symbolic representations for world modeling, planning, and memory management in CA frameworks.
• Applies prompt engineering and automated program generation (e.g., via PAL) to align LLM outputs with CA-executable symbolic logic.

Experimental results

Research questions

• RQ1 How can LLMs and CAs be integrated to improve reasoning consistency and reduce hallucinations in AI agents?
• RQ2 What architectural configurations enable effective synergy between connectionist (LLM) and symbolic (CA) processing in autonomous systems?
• RQ3 To what extent can LLMs enhance CA knowledge representation and scalability while preserving interpretability and bounded rationality?
• RQ4 How do different integration levels (modular, agency, neuro-symbolic) affect system robustness, error propagation, and planning reliability?
• RQ5 Can hybrid systems leverage LLMs for dynamic world modeling and CA components for deliberative, goal-directed planning in complex environments?

Key findings

• The modular approach enables incremental integration, with the most loosely coupled model using CA outputs for in-context learning, though it limits high-level reasoning.
• Tightly integrated modular models allow LLMs to populate ontologies and generate syntactically correct CA productions, improving world modeling and symbolic reasoning.
• The agency approach enhances system resilience through redundant agent roles and self-reflective role-playing, reducing cascading failures in reasoning chains.
• Neuro-symbolic integration enables bidirectional translation between natural language and symbolic logic, supporting continuous rule generalization and explainable decision-making.
• All approaches reduce reliance on purely prompt-based LLM execution, with symbolic components providing structure and consistency to mitigate hallucinations.
• Despite benefits, all approaches face challenges related to LLM reliability, including cascading failures from hallucinated or biased outputs in planning and knowledge extraction.