A brain-inspired agentic architecture to improve planning with LLMs

Area of Science:

• Artificial Intelligence
• Cognitive Neuroscience

Background:

• Large language models (LLMs) excel at many tasks but falter in multi-step reasoning and goal-directed planning.
• Human planning involves specialized cognitive processes like conflict monitoring, state prediction, and task coordination, often linked to specific brain regions.

Purpose of the Study:

• To develop an improved planning capability for LLMs by integrating brain-inspired modular processes.
• To address the limitations of standard LLMs in autonomously coordinating planning functions.

Main Methods:

• Proposed a Modular Agentic Planner (MAP) architecture, utilizing specialized, brain-inspired LLM modules.
• Evaluated MAP on graph traversal, Tower of Hanoi, PlanBench benchmark, and the strategyQA NLP task.

Main Results:

• MAP significantly outperformed standard LLMs and agentic baselines on planning tasks.
• MAP demonstrated effective combination with smaller, cost-efficient LLMs and superior cross-task transfer.
• The architecture showed improved multi-step reasoning and goal-directed planning.

Conclusions:

• Integrating cognitive neuroscience principles into LLM architectures enhances planning abilities.
• Modular, brain-inspired design is a promising direction for advancing LLM performance in complex reasoning tasks.