# A Prefrontal Cortex-inspired Architecture for Planning in Large Language Models

**Authors:** Taylor Webb, Shanka Subhra Mondal, Chi Wang, Brian Krabach, Ida Momennejad  
**arXiv:** [2310.00194](https://arxiv.org/abs/2310.00194)  
**Date:** 2023-09-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.00194v3/

## Key points

- Researchers explore limitations of Large Language Models (LLMs) in tasks requiring multi-step reasoning or goal-directed planning
- Proposal of a novel black box architecture, GPT-4, with multiple LLM-based modules inspired by the human brain's prefrontal cortex (PFC)
- Modules mimic functions such as conflict monitoring, state prediction, task decomposition, and coordination found in the PFC
- New architecture improves planning by breaking down complex problems into automated calls to the LLM through specialized PFC-inspired modules
- Evaluation of combined architecture on challenging planning tasks like graph traversal, Tower of Hanoi, and logistics shows significant performance improvements compared to standard LLM methods and competitive baselines
- Study demonstrates potential benefits of integrating knowledge from cognitive neuroscience to enhance planning capabilities in LLMs

## Summary

In this study, the researchers explore the limitations of Large Language Models (LLMs) in tasks that require multi-step reasoning or goal-directed planning. Drawing inspiration from the human brain's prefrontal cortex (PFC), which utilizes specialized modules for planning, they propose a novel black box architecture with multiple LLM-based modules (GPT-4). These modules mimic functions such as conflict monitoring, state prediction, task decomposition, and coordination found in the PFC. The new architecture improves planning by breaking down complex problems into automated calls to the LLM through specialized PFC-inspired modules. The researchers evaluate their combined architecture on challenging planning tasks like graph traversal, Tower of Hanoi, and logistics. Comparing it against standard LLM methods and competitive baselines like zero-shot prompting and chain-of-thought approaches, they find significant improvements in performance. By leveraging insights from cognitive neuroscience to enhance planning capabilities in LLMs, this study demonstrates the potential benefits of integrating knowledge from different domains to advance artificial intelligence research.