# Large Language Model Can Continue Evolving From Mistakes

**Authors:** Haokun Zhao, Haixia Han, Jie Shi, Chengyu Du, Jiaqing Liang, Yanghua Xiao  
**arXiv:** [2404.08707](https://arxiv.org/abs/2404.08707)  
**Date:** 2024-04-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.08707v1/

## Key points

- Large Language Models (LLMs) face challenges despite impressive performance in downstream tasks
- LLMs often generate incorrect responses due to knowledge deficiencies and flawed pre-training data
- Continual Learning (CL) is commonly used to address this issue, but traditional methods are task-oriented and costly
- Authors propose the Continue Evolving from Mistakes (CEM) method for iterative refinement of LLMs
- CEM method identifies knowledge deficiencies through incorrect responses and collects relevant data from multiple sources
- Involves iterative supplementary training to continuously update and supplement the model's knowledge
- Two strategies developed to construct supplementary training sets that enhance understanding of corpus and prevent catastrophic forgetting
- Extensive experiments validate effectiveness of CEM method, showing 17.00% improvement in accuracy of LLMs

## Summary

In their paper titled "Large Language Model Can Continue Evolving From Mistakes," authors Haokun Zhao, Haixia Han, Jie Shi, Chengyu Du, Jiaqing Liang, and Yanghua Xiao discuss the challenges faced by Large Language Models (LLMs) despite their impressive performance in various downstream tasks. These models often generate incorrect responses due to knowledge deficiencies and flawed pre-training data. To address this issue, Continual Learning (CL) is commonly used; however, traditional CL methods are task-oriented and require retraining LLMs from scratch using novel or factually accurate data. This can lead to high training costs. To overcome these challenges, the authors propose the Continue Evolving from Mistakes (CEM) method inspired by the 'summarize mistakes' learning skill. This approach aims for iterative refinement of LLMs by identifying knowledge deficiencies through incorrect responses and collecting relevant data from multiple sources. The CEM method involves iterative supplementary training to continuously update and supplement the model's knowledge. Additionally, two strategies are developed to construct supplementary training sets that enhance the LLM's understanding of the corpus and prevent catastrophic forgetting. Extensive experiments were conducted to validate the effectiveness of the CEM method, showing a significant improvement of 17.00% in the accuracy of LLMs. This innovative approach offers a more efficient way to refine large language models by addressing knowledge deficiencies through targeted training updates based on mistakes made during inference.