In the minds of stochastic parrots: Benchmarking, evaluating and interpreting large language models

Abstract

The arrival of large language models (LLMs) in recent years has changed the landscape of natural language processing (NLP). Their impressive performance on popular benchmarks, ability to solve a range of different tasks and their human-like linguistic interactional abilities, have prompted a debate into whether these are just "stochastic parrots" who are cleverly repeating what humans say without understanding its meaning or whether they are acquiring essential language capabilities, which would be an important stepping stone towards artificial general intelligence.

To tackle this question, developing analysis methods to measure and understand the language capabilities of LLMs has become a defining challenge. These include developing benchmarks to reliably measure their performance as well and interpretability methods to gauge their inner-workings. This is especially relevant at a time when these models already are having a considerable impact on our society. An increasing amount users are affected by the technology and calls are made for transparent, regulated and thorough evaluation of AI. In these efforts, it is important to estimate the possibilities and limitations of these analysis methods since they will play an important role in holding technologies in AI accountable.

In this compilation thesis, I expound on the components and processes involved in analyzing LLMs. The articles included in this compilation thesis use different approaches for analyzing LLMs, from introducing a multi-task benchmark Superlim for Swedish NLU to investigating LLMs' ability to predict language variation. To this effort I explore what the possibilities and limitations are of popular analysis methods and what implications these have for developing LLMs. I argue that integrating explanatory approaches from empirical linguistic research is important to understand the role of both the data and the linguistic features used when analyzing LLMs. Doing so does not only help guide the development of LLMs, but also bring insights into linguistics.