# Making Language Models Better Tool Learners with Execution Feedback

**Authors:** Shuofei Qiao, Honghao Gui, Huajun Chen, Ningyu Zhang  
**arXiv:** [2305.13068](https://arxiv.org/abs/2305.13068)  
**Date:** 2023-05-22  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2305.13068v1/

## Key points

- The paper proposes a framework called Tool leaRning wIth exeCution fEedback (TRICE) to enable language models to selectively use tools while enhancing performance.
- TRICE is an end-to-end framework that enables the model to continually learn through feedback derived from tool execution.
- The proposed two-stage approach includes training the model on simple tasks that it can readily resolve without tools and then gradually introducing more complex tasks where tools are necessary.
- Experimental results show that TRICE outperforms existing methods in terms of accuracy while reducing the number of times a tool is used.
- The code and datasets for TRICE will be available on GitHub at https://github.com/zjunlp/trice.
- This paper provides a promising solution for improving language models' ability to use tools effectively, which has significant implications for various natural language processing applications.

## Summary

The paper "Making Language Models Better Tool Learners with Execution Feedback" proposes a framework called Tool leaRning wIth exeCution fEedback (TRICE) to enable language models to selectively use tools while enhancing performance. TRICE is an end-to-end framework that enables the model to continually learn through feedback derived from tool execution. The proposed two-stage approach includes training the model on simple tasks that it can readily resolve without tools and then gradually introducing more complex tasks where tools are necessary. Experimental results show that TRICE outperforms existing methods in terms of accuracy while reducing the number of times a tool is used. The code and datasets for TRICE will be available on GitHub at https://github.com/zjunlp/trice. This paper provides a promising solution for improving language models' ability to use tools effectively, which has significant implications for various natural language processing applications.