The emergence of Large Language Models (LLMs) has redefined benchmarks in natural language processing, however their deployment as complete reliable autonomous agents for multi-step reasoning and planning remains a subject of academic exploration. While state-of-the-art models such as GPT-4 demonstrate high linguistic proficiency, they face challenges in maintaining contextual information across extended interactions and implementing systematic problem-solving strategies characteristic of human cognition. Classical cognitive architectures, such as SOAR and ACT-R, offer robust frameworks for goal-directed decision-making and structured memory management, but lack the probabilistic scalability and semantic flexibility of modern neural approaches. This thesis investigates these archetypes, exploring how cognitive architectural principles can be integrated into LLM-based agents to further enhance reasoning and improve performance on complex tasks. Drawing upon the Cognitive Architectures for Language Agents (CoALA) framework, this study examines the implementation of internal reasoning actions and structured decision-making processes, including methods such as Chain of Thought, Graph of Thought, and iterative refinement. Through comparative evaluation using the UltraInteract and StrategyQA benchmarks, this research investigates how different organizational structures of inference processes align with task-specific characteristics and computational requirements.
Ouadih Jonas (Thu,) studied this question.