# Language Models can Solve Computer Tasks

**Authors:** Geunwoo Kim, Pierre Baldi, Stephen McAleer  
**arXiv:** [2303.17491](https://arxiv.org/abs/2303.17491)  
**Date:** 2023-03-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.17491v1/

## Key points

- Language models (LMs) can automate computer tasks and improve efficiency and productivity
- The authors propose a novel approach called recursive criticism and improvement (RCI) using a pre-trained large language model (LLM) agent guided by natural language commands
- RCI significantly outperforms existing LLM methods for automating computer tasks
- RCI surpasses supervised learning (SL) and reinforcement learning (RL) approaches on the MiniWoB++ benchmark
- RCI requires only a handful of demonstrations per task instead of tens of thousands and does not rely on task-specific reward functions
- RCI prompting enhances LLMs' reasoning abilities on various natural language reasoning tasks, outperforming chain of thought (CoT) prompting
- Combining RCI with CoT prompts yields even better results than using either method separately
- Pre-trained LLM agents have potential in solving computer tasks through natural language commands
- The RCI approach reduces reliance on expert demonstrations and task-specific rewards while achieving impressive performance on automation and reasoning tasks.

## Summary

In their paper titled "Language Models can Solve Computer Tasks," Geunwoo Kim, Pierre Baldi, and Stephen McAleer explore the potential of language models (LMs) in automating computer tasks. They highlight the importance of agents capable of carrying out general tasks on a computer to improve efficiency and productivity by automating repetitive tasks and assisting in complex problem-solving. To address the issue of relying on large amounts of expert demonstrations and task-specific reward functions for new tasks, the authors propose a novel approach using a pre-trained large language model (LLM) agent guided by natural language commands through a simple prompting scheme called recursive criticism and improvement (RCI). The RCI approach significantly outperforms existing LLM methods for automating computer tasks and surpasses supervised learning (SL) and reinforcement learning (RL) approaches on the MiniWoB++ benchmark. Notably, RCI achieves competitive performance with state-of-the-art SL+RL methods while requiring only a handful of demonstrations per task instead of tens of thousands and without relying on task-specific reward functions. Moreover, the authors demonstrate that RCI prompting enhances LLMs' reasoning abilities on various natural language reasoning tasks, outperforming chain of thought (CoT) prompting. Interestingly, combining RCI with CoT prompts yields even better results than using either method separately. Overall, this research showcases the potential of pre-trained LLM agents in solving computer tasks through natural language commands. The RCI approach offers an effective solution that reduces reliance on expert demonstrations and task-specific rewards while achieving impressive performance on both automation and reasoning tasks.