A Powerful New Method for Large Language Models
Researchers from the Hong Kong University of Science and Technology and the University of Illinois Urbana-Champaign joined forces to tackle a big problem facing large language models (LLMs): hallucination. Hallucination happens when these models generate fake facts. The new approach developed by the researchers is called Refusal-Aware Instruction Tuning (R-Tuning).
R-Tuning works by recognizing the gap between the knowledge of LLMs and the instruction data, then creating a refusal-aware dataset. This dataset identifies uncertain questions and trains the model to decline to answer questions that fall outside of its knowledge base.
They ran experiments on seven datasets, showing that R-Tuning is effective at refusing uncertain questions. It also outperformed other models in terms of accuracy. They also discovered that bigger models perform better, and learning uncertainty during training improves the model’s performance.
Overall, R-Tuning is a powerful method for teaching LLMs to refuse unknown questions. This refusal skill can be applied to various tasks, potentially improving the reliability and performance of large language models.
To learn more, check out the research paper. And don’t forget to join the ML SubReddit, Facebook Community, Discord Channel, LinkedIn Group, Twitter, and Email Newsletter for more AI research news and cool AI projects.
If you’re into AI, you’ll love our newsletter.