Large language models increasingly adapt using external feedback to alter inference-time behaviour, persistent state or model parameters. However, the signals that drive these changes may be noisy, biased, incomplete, delayed, correlated with the model’s own errors, or progressively underused during long-context and multi-step processing. This project studies reliable adaptation as a general feedback-mediated problem by distinguishing failures at the feedback source, during signal transmission and during model updating.
In this project, we consider Biomedical NLP as a domain that provides a demanding validation setting. The intended outcome is a framework for LLMs that improve from imperfect feedback without amplifying cumulative error.
Required knowledge
Strong background in NLP (Natural Language Processing) and Machine Learning, especially LLMs.
Strong coding skills in Python and knowledge of agentic environments.
