# Into the Unknown: Self-Learning Large Language Models

**Authors:** Teddy Ferdinan, Jan Kocoń, Przemysław Kazienko  
**arXiv:** [2402.09147](https://arxiv.org/abs/2402.09147)  
**Date:** 2024-02-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.09147v1/

## Key points

- Authors address a critical issue in self-learning large language models: determining what to learn
- Introduce a novel self-learning LLM framework for acquiring unknown knowledge through evaluating hallucinations
- Propose concept of Points in The Unknown (PiUs) using hallucination score
- Present extrinsic and intrinsic methods for identifying PiUs automatically
- Establish a self-learning loop targeting knowledge gaps represented by PiUs to reduce hallucination score
- Develop evaluation metrics to assess LLM's capacity for self-learning
- Experiments show significant proficiency in self-learning for 7B-Mistral models
- Self-learning concept streamlines LLM updates and enhances public trust in AI systems

## Summary

In their paper titled "Into the Unknown: Self-Learning Large Language Models," authors Teddy Ferdinan, Jan Kocoń, and Przemysław Kazienko address a critical issue in self-learning large language models (LLMs): determining what to learn. They introduce a novel self-learning LLM framework that empowers these models to autonomously acquire previously unknown knowledge by evaluating their own hallucinations. Through the use of a hallucination score, they propose the concept of Points in The Unknown (PiUs) and present one extrinsic and three intrinsic methods for automatically identifying PiUs. This framework establishes a self-learning loop that specifically targets the knowledge gaps represented by PiUs, ultimately leading to a reduction in the hallucination score. Additionally, the authors develop evaluation metrics to assess an LLM's capacity for self-learning. Their experiments demonstrate that 7B-Mistral models, whether fine-tuned or aligned, exhibit significant proficiency in self-learning. The proposed self-learning concept not only streamlines LLM updates but also paves the way for new avenues of knowledge exchange. Furthermore, it has the potential to enhance public trust in artificial intelligence (AI) systems. Overall, this research contributes valuable insights into advancing the capabilities of LLMs through autonomous learning mechanisms and underscores the importance of addressing knowledge gaps to improve model performance and reliability.