Interpreting Behaviour Emergence in Large Language Models via Training Dynamics

NLP has undergone major advancements in recent years, and there are no signs of this
trend slowing down. The key to these improvements has been scale: LLMs are growing
larger, training for longer, and being fed more data. While this has unlocked groundbreaking performance
in numerous domains, it has also rendered the inner workings of LLMs much more difficult to understand.
Mechanistic interpretability is a promising area of research that seeks to reverse engineer the operations performed
by transformer-based LLMs, but it is struggling to contend with the immense size models have reached. The
prediction of a single token is governed by up to trillions of parameters, which form complex circuits and
exhibit very unintuitive behaviours. Even so, the field has an ethical responsibility to comprehend the technology
it creates, especially as LLMs are deployed to increasingly sensitive parts of our lives. Therefore, I seek to
contribute towards our understanding of LLMs by investigating their training dynamics. Specifically, I want
to draw closer connections between the data models are trained on, the capabilities and behaviours that emerge
as a result, and the mechanistic underpinnings that govern the learning process.