Kazlaris Ioannis (Phd Candidate)

Thesis title: Mitigation of Hallucinations in Large Language Models.

Supervisor: Antoniou Efstathios

Advisory Committee Members:

Konstantinos Diamantaras, Professor, Dept. of Information and Electronic Engineering, IHU
Bratsas Charalampos, Assistant Professor, Dept. of Information and Electronic Engineering, IHU

Abstract:

The rapid development of deep learning has led to models that can understand and produce text and visual data with great accuracy. Despite the progress, the phenomenon of “hallucinations”, where models produce inaccurate or arbitrary content, continues to concern the research community. This PhD proposal aims to investigate and use deep learning methods (e.g. Reinforcement Learning, Meta-Prompting, Chain of Thoughts, Tree of Thoughts, Contrastive Learning, etc.) to improve the alignment and uniformity of vector representations and reduce the frequency of the phenomenon of hallucinations in large language models.

The main objectives of the thesis include the development of the theoretical background of new approaches, their implementation with deep learning techniques, the conduct of extensive experiments on appropriate datasets and the evaluation of the results. It is expected that the use of existing approaches, as well as the exploration of innovative methods, will contribute to limiting the production of model illusions in tasks that require a detailed understanding of the relationships between data.

The challenges that will be addressed include the increased computational complexity, the need for large datasets of qualitative data, ensuring that the model will generalize effectively to new, unknown data, as well as the inherent peculiarities of the aforementioned methods.