# ART: Automatic multi-step reasoning and tool-use for large language models

**Authors:** Bhargavi Paranjape, Scott Lundberg, Sameer Singh, Hannaneh Hajishirzi, Luke Zettlemoyer, Marco Tulio Ribeiro  
**arXiv:** [2303.09014](https://arxiv.org/abs/2303.09014)  
**Date:** 2023-03-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2303.09014v1/

## Key points

- Large language models (LLMs) can perform complex reasoning in few- and zero-shot settings by generating intermediate chain of thought (CoT) reasoning steps.
- Automatic Reasoning and Tool-use (ART) is a framework that uses frozen LLMs to automatically generate intermediate reasoning steps as a program.
- ART selects demonstrations of multi-step reasoning and tool use from a task library when given a new task to solve.
- ART seamlessly pauses generation whenever external tools are called, integrates their output before resuming generation, and achieves substantial improvement over few-shot prompting and automatic CoT on unseen tasks in the BigBench and MMLU benchmarks.
- ART matches performance of hand-crafted CoT prompts on a majority of these tasks.
- The framework is also extensible, making it easy for humans to improve performance by correcting errors in task-specific programs or incorporating new tools.
- ART represents an important step towards automating multi-step reasoning and tool-use for large language models without requiring extensive manual effort.

## Summary

Large language models (LLMs) have demonstrated the ability to perform complex reasoning in few- and zero-shot settings by generating intermediate chain of thought (CoT) reasoning steps. To address the limitation of requiring hand-crafted task-specific demonstrations and carefully scripted interleaving of model generations with tool use, the authors introduce Automatic Reasoning and Tool-use (ART), a framework that uses frozen LLMs to automatically generate intermediate reasoning steps as a program. ART selects demonstrations of multi-step reasoning and tool use from a task library when given a new task to solve. At test time, ART seamlessly pauses generation whenever external tools are called, integrates their output before resuming generation, and achieves substantial improvement over few-shot prompting and automatic CoT on unseen tasks in the BigBench and MMLU benchmarks. Moreover, ART matches performance of hand-crafted CoT prompts on a majority of these tasks. The framework is also extensible, making it easy for humans to improve performance by correcting errors in task-specific programs or incorporating new tools. The authors demonstrate this capability by drastically improving performance on select tasks with minimal human intervention. Overall, ART represents an important step towards automating multi-step reasoning and tool-use for large language models without requiring extensive manual effort. This has significant implications for fields such as natural language processing, where LLMs have shown great promise but require further development to achieve their full potential.