# EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling

**Authors:** Shimao Zhang, Yu Bao, Shujian Huang  
**arXiv:** [2403.14541](https://arxiv.org/abs/2403.14541)  
**Date:** 2024-03-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2403.14541v2/

## Key points

- Authors Shimao Zhang, Yu Bao, and Shujian Huang address fixed temperature parameters in Large Language Models (LLMs)
- LLMs have shown remarkable performance in language tasks with temperature sampling
- Fixed temperature parameter may not always result in optimal balance between generation quality and diversity
- Proposed approach: Entropy-based Dynamic Temperature (EDT) Sampling
- EDT dynamically selects temperature parameter for balanced performance in generation quality and diversity
- Model performance evaluations and analyses across four different benchmarks demonstrate effectiveness of EDT
- EDT significantly outperforms existing strategies across various tasks
- Proposed method enhances overall performance of LLMs in generating diverse and high-quality outputs
- Research contributes insights to optimizing LLMs' generation processes and importance of adaptive strategies for improving language model capabilities

## Summary

In their paper titled "EDT: Improving Large Language Models' Generation by Entropy-based Dynamic Temperature Sampling," authors Shimao Zhang, Yu Bao, and Shujian Huang address the issue of fixed temperature parameters in Large Language Models (LLMs) during the generation process. LLMs have shown remarkable performance in various language tasks, with temperature sampling being a common decoding strategy. However, the use of a fixed temperature parameter may not always result in an optimal balance between generation quality and diversity. To tackle this challenge, the authors propose an innovative approach called Entropy-based Dynamic Temperature (EDT) Sampling. This method dynamically selects the temperature parameter to achieve a more balanced performance in terms of both generation quality and diversity. The study includes model performance evaluations and comprehensive analyses across four different generation benchmarks to demonstrate the effectiveness of EDT. The experiments conducted by Zhang, Bao, and Huang reveal that EDT significantly outperforms existing strategies across various tasks. By introducing dynamic temperature selection based on entropy considerations, the proposed method enhances the overall performance of LLMs in generating diverse and high-quality outputs. This research contributes valuable insights to optimizing LLMs' generation processes and highlights the importance of adaptive strategies for improving language model capabilities.