Researchers from Carnegie Mellon University have investigated the integration of Large Language Models (LLMs) and Cognitive Architectures (CAs) in the development of intelligent artificial agents. They propose three integration approaches: the modular approach, the agency approach, and the neurosymbolic approach. Each approach is based on theoretical models and supported by preliminary empirical evidence. LLMs excel in natural language tasks and show potential in interactive decision making, but face interpretability and scalability issues. CAs propose hypotheses about the structures governing minds, facilitating intelligent behavior, but face challenges in knowledge representation and scalability. The research aims to integrate these approaches to overcome their individual limitations.
Integration of Large Language Models and Cognitive Architectures for AI Development
A team of researchers from Carnegie Mellon University, including Oscar J. Romero, John Zimmerman, Aaron Steinfeld, and Anthony Tomasic, have explored the integration of two AI subdisciplines, Large Language Models (LLMs) and Cognitive Architectures (CAs), in the development of artificial agents that exhibit intelligent behavior. They present three integration approaches, each grounded in theoretical models and supported by preliminary empirical evidence.
The Three Integration Approaches
The first approach, the modular approach, introduces four models with varying degrees of integration. It 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 second approach, the agency approach, is motivated by the Society of Mind theory and the LIDA cognitive architecture. It proposes the formation of agent collections that interact at micro and macro cognitive levels, driven by either LLMs or symbolic components.
The third approach, the neurosymbolic approach, takes inspiration from the CLARION cognitive architecture. It 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.
Strengths and Weaknesses of LLMs and CAs
LLMs, like ChatGPT, GPT-4, and PaLM, are generative models that excel in a variety of natural language tasks and even show promise in interactive decision making, reasoning, and modeling aspects of artificial general intelligence (AGI). However, they face interpretability, consistency, and scalability issues, partly due to limitations in context window size and sensitivity to prompt structure.
On the other hand, CAs propose hypotheses about the fixed structures governing the operation of minds, whether in natural or artificial systems, facilitating intelligent behavior in complex environments. However, CAs face challenges in knowledge representation and scalability.
Relevant Work in the Field
Chain-of-thought prompting (CoT) enhances LLM reasoning, leading to improved performance in various reasoning and natural language processing tasks. Augmented Language Models combine enhanced reasoning skills of an LLM with tools like APIs, DBs, and code interpreters for improved knowledge retrieval, reasoning, and action execution.
Conclusion
The main contribution of this work lies in characterizing plausible approaches to integrating CAs and LLMs, viewing them through a hybrid and synergetic lens. Both LLMs and CAs have made valuable and sound contributions to the construction of complex autonomous AI agents, however, each approach has its strengths and weaknesses. The researchers discuss the trade-offs and challenges of each approach.
“Synergistic Integration of Large Language Models and Cognitive Architectures for Robust AI: An Exploratory Analysis” is an article authored by Oscar J. Romero, John Zimmerman, Aaron Steinfeld, and Anthony Tomasic. Published on January 22, 2024, in the Proceedings of the AAAI Symposium Series, the paper explores the integration of large language models and cognitive architectures for the development of robust AI. The article can be accessed via its DOI: https://doi.org/10.1609/aaaiss.v2i1.27706.