Open project: Instruction Following for Resource Efficient AI Models

Large language models (LLMs) are known for their ability to follow instructions and learn from examples provided inside a prompt. This provides an extremely powerful interface for people to interact with machines and ask them to perform tasks, such as to collate and summarise information in a particular format. However, LLMs demand a huge amount of resources in terms of computing power, memory and energy, and can be outperformed by smaller, specialised models for tasks like text classification in a specific domain [1]. This project explores how to develop instruction following capabilities for smaller neural network models, aiming to create efficient models that can be adapted to specialised tasks by end users through prompts. The project will investigate approaches such as distillation from LLMs to small models [2] to combine multiple user instructions, and the potential to learn from explanations [3] provided by a human-in-the-loop to correct a model’s errors. The supervisor has several ongoing industry and academic partnerships, which will allow the PhD student to target their research towards a real-world information processing application in healthcare, medical research, finance, or intelligence analysis.

References:

• [1] J. Kocoń, et al., “ChatGPT: Jack of all trades, master of none,” Information Fusion, Volume 99, p. 101861, 2023.
• [2] L. Vöge, V. Gurgul, S. Lessmann, “Leveraging Zero-Shot Prompting for Efficient Language Model Distillation,” arXiv preprint arXiv:2403.15886, 2024
• [3] R. Menon, S. Ghosh, S. Srivastava, “CLUES: A benchmark for learning classifiers using natural language explanations,” ACL, 2022.
• Related recent work from our lab on interactive machine learning for NLP:
• H. Fang, J. Gor, E. Simpson, “Efficiently acquiring human feedback with Bayesian deep learning,” in Proc. 1st Workshop on Uncertainty-aware NLP, 2024.
• Y. Ye, E. Simpson, “Towards abstractive timeline summarisation using preference-based reinforcement learning,” in Proc. ECAI, 2023.

Applying

Please contact Edwin Simpson by email if you are interested in this topic, or other aspects of interactive NLP, uncertainty quantification, or applications to healthcare, with a brief summary of your interests, background and ambitions for your PhD. We are currently considering applications for University of Bristol scholarships and DTP studentships. It is also possible to apply for a position on the Centre for Doctoral Training in Practice-oriented AI or apply for a national scholarship such as the China Scholarship Council scholarship (applications for Sep 2025 entry must be received by 2nd December 2024).