# 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/2310.00194v1/

## Key points

- Large Language Models (LLMs) struggle with tasks requiring multi-step reasoning or goal-directed planning
- Authors propose a black box architecture inspired by the human brain's prefrontal cortex (PFC)
- PFC-inspired modules perform functions like conflict monitoring, state prediction, state evaluation, task decomposition, and task coordination
- LLMs can perform these functions individually but struggle to coordinate them towards a goal
- Proposed architecture combines LLM-based modules (GPT-4) with specialized PFC-inspired modules to break down complex problems into automated calls to the LLM
- Evaluation on planning tasks shows significant improvements compared to standard LLM methods like zero-shot prompting or in-context learning
- Incorporating knowledge from cognitive neuroscience enhances planning capabilities of LLMs

## Summary

Large Language Models (LLMs) have shown impressive performance on various tasks, but they often struggle with tasks that require multi-step reasoning or goal-directed planning. To address this limitation, the authors propose a black box architecture inspired by the human brain's prefrontal cortex (PFC), which is responsible for planning through recurrent interaction of specialized modules. These PFC-inspired modules perform functions such as conflict monitoring, state prediction, state evaluation, task decomposition, and task coordination. While LLMs can perform these functions in isolation, they struggle to coordinate them autonomously towards a goal. The proposed architecture utilizes multiple LLM-based modules (GPT-4) that interact with specialized PFC-inspired modules to break down complex problems into brief automated calls to the LLM. The authors evaluate this combined architecture on challenging planning tasks like graph traversal and Tower of Hanoi and find significant improvements compared to standard LLM methods such as zero-shot prompting or in-context learning. This research demonstrates the potential benefits of incorporating knowledge from cognitive neuroscience into LLMs to enhance their planning capabilities.