Mitigating Hallucinations and Enhancing Confidence Estimation in LLMs for reliable Information [...]

10–12-week development experience running through the summer months each year. Internships are meant to be a catalyst in helping students gain hands-on learning and do not require sponsorship to work in the UK. Students pursuing undergraduate, masters or doctoral degrees are eligible. A salary will be paid.

Large Language Models (LLMs) such as GPT have gained significant attention for their remarkable ability to generate human-like text, making them indispensable tools for various natural language applications. Multiple teams within AZ are already involved in utilising these powerful LLMs for various applications including content/report generation, summarisation, chat assistant, multi-modal biomarkers, information extraction from unstructured documents, etc.

However, the presence of hallucinations in LLM-generated content can pose significant risks to medical decision-making, patient safety, and data integrity. Hence, it is critical to investigate and develop methods to improve the reliability of these powerful LLMs across AZ.

This work delves into the critical importance of mitigating hallucinations in LLMs and thereby improving the interpretability of LLM responses by estimating well-calibrated confidence scores for LLM predictions.

The key objectives of this project are:

• Develop a method to utilise a combination of expert models to provide confidence bounds at an output level and mitigate hallucination in LLMs primarily for information extraction from unstructured and semi-structured data sources.
• Develop confidence estimation techniques that provide a clear understanding of the reliability of AI-generated responses across AZ applications.
• Conduct extensive experiments, collect, and analyse data (internal dataset available), and publish the research findings to further increase the impact.