Large Language Model Can Continue Evolving From Mistakes

As world knowledge evolves and new task paradigms emerge, Large Language Models (LLMs) often fall short of meeting new demands due to knowledge deficiencies and outdated information. Continual Learning (CL) is crucial for keeping LLMs up-to-date and addressing these deficiencies. However, traditional CL approaches struggle to balance task-width generality with task-depth specificity and often lack efficient data collection strategies, leading to increased training costs without addressing the model's most critical needs. Inspired by the `summarizing mistakes' learning skill, we propose the Continue Evolving from Mistakes (CEM) method. This iterative approach continually evaluates LLMs to identify knowledge deficiencies based on their mistakes, collecting relevant data from multiple sources to supplement training in a targeted manner. To enhance the model's utilization of supplemental knowledge and prevent forgetting, we developed three dataset construction strategies that integrate various types of continual pretraining (CPT) data and continual instruction tuning (CIT) data. Extensive experiments demonstrate the efficacy of the CEM method, achieving up to a 17% improvement in LLM accuracy in the best scenarios. Additionally, further experiments confirm the potential of combining CEM with other catastrophic forgetting mitigation strategies, enabling multi-round iterative optimization.