# OlaGPT: Empowering LLMs With Human-like Problem-Solving Abilities

**Authors:** Yuanzhen Xie, Tao Xie, Mingxiong Lin, WenTao Wei, Chenglin Li, Beibei Kong, Lei Chen, Chengxiang Zhuo, Bo Hu, Zang Li  
**arXiv:** [2305.16334](https://arxiv.org/abs/2305.16334)  
**Date:** 2023-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16334v1/

## Key points

- Large language models (LLMs) have made progress in reasoning tasks through generating chains of thought (COT)
- There is still a gap between LLMs and human problem-solving capabilities
- Current approaches focus on COT and tool use, neglecting human cognitive frameworks
- OlaGPT is a new intelligent framework that simulates aspects of human cognition such as attention, memory, reasoning, learning, scheduling, and decision-making mechanisms
- OlaGPT proposes a learning unit to record previous mistakes and expert opinions dynamically to strengthen its ability to solve similar problems
- The paper outlines effective reasoning frameworks for human problem-solving and designs COT templates accordingly
- A comprehensive decision-making mechanism is proposed to maximize model accuracy
- OlaGPT surpasses state-of-the-art benchmarks with superior performance on multiple reasoning datasets
- The implementation of OlaGPT is available on GitHub for further exploration by researchers

## Summary

The field of large language models (LLMs) has made significant strides in performing reasoning tasks by generating chains of thought (COT) through specific prompts. However, there is still a considerable gap between the problem-solving capabilities of LLMs and humans. Current approaches to improving LLMs focus on COT and tool use, neglecting the adoption and application of human cognitive frameworks. Humans employ various cognitive abilities when facing complex reasoning challenges, necessitating interaction with tools, knowledge, and external environment information to accomplish intricate tasks. To address this issue, a novel intelligent framework called OlaGPT has been introduced that simulates certain aspects of human cognition by approximating different cognitive modules such as attention, memory, reasoning, learning, scheduling and decision-making mechanisms. Inspired by the active learning mechanism of humans, OlaGPT proposes a learning unit to record previous mistakes and expert opinions dynamically to strengthen its ability to solve similar problems. The paper outlines common effective reasoning frameworks for human problem-solving and designs COT templates accordingly. A comprehensive decision-making mechanism is also proposed to maximize model accuracy. The efficacy of OlaGPT has been stringently evaluated on multiple reasoning datasets and surpasses state-of-the-art benchmarks with superior performance. The implementation of OlaGPT is available on GitHub for further exploration by researchers in the field. This research presents an innovative approach towards bridging the gap between LLMs' problem-solving capabilities and those of humans by incorporating human-like problem-solving abilities into LLMs through simulation of certain aspects of human cognition such as attentional processes, memory recall strategies or decision making algorithms which are essential components in any successful problem solving process.