Data attribution for LLMs

Contact person: Martin Jullum

Status: Project

What parts of an LLM’s training data are responsible for the answer it generates to a given prompt? This project develops scalable, fine-grained methods for tracing the outputs and behaviour of large language models back to influential training examples and sources. Combining explainable AI, machine learning, natural language processing and statistical methods, we aim to move beyond simple influence scores and distinguish different contributions, such as factual content and linguistic style. The methods will be implemented and tested on modern LLMs and across training settings, including fine-tuning and learning from examples. The results could help model developers, auditors and other users understand and verify model outputs, diagnose errors and bias, improve training-data curation, support targeted unlearning, and build fairer, safer and more trustworthy language models.