PhD Thesis: Interpretability and Evaluation of LLMs and Agentic Workflows

Starting date: November 1st, 2024 (flexible)

Salary: 2,135€ gross / month (social security included)

Place of work (no remote): Laboratoire d'Informatique de Grenoble, CNRS, Grenoble, France

Natural language processing (NLP) has undergone a paradigm shift in recent years, owing to the remarkable breakthroughs achieved by large language models (LLMs). These models have completely altered the landscape of NLP by demonstrating impressive results in language modeling, translation, and summarization. Nonetheless, the use of LLMs has also surfaced crucial questions regarding their reliability and transparency. As a result, there is now an urgent need to gain a deeper understanding of the mechanisms governing the behavior of LLMs, to interpret their decisions and outcomes in scientifically grounded ways, and to precisely evaluate their abilities and limitations. Adding to the complexity, LLMs are often involved as only one small component of larger, more ambitious, extit{agentic workflows} [SemEra]. In an agentic workflow, LLMs collaborate with other LLMs, humans, and tools by exchanging natural language messages to solve complex problems beyond the capabilities of an LLM alone.

Evaluation of LLMs has become particularly challenging as they consume most of the internet during their pre-training, including most of the test splits of evaluation benchmarks [LeakCheatRepeat]. Furthermore, the landscape of available LLMs is changing fast and they have access to web via tools as part of agentic workflows. Therefore, new evaluation methodologies beyond assessing models' skills on a fixed test set are needed to consider these novel properties [Flows]. 

A promising direction to carry out evaluation and interpretability analysis is to take inspiration from the field of Neuroscience which, over the years, has crafted experimental setups to undercover how the human brain computes and represents useful information for tasks of interest [RepEng]. Additionally, we can get help from causal analysis and causal inference toolkits [CausalAbstraction]. Examining the causal relationships between the inputs, outputs, and hidden states of LLMs, can help to build scientific theories about the behavior of these complex systems. Furthermore, causal inference methods can help uncover underlying causal mechanisms behind the complex computations of LLMs, giving hope to better interpret their decisions and understand their limitations [Glitch].

As a Ph.D student working on such a project, you will be expected to develop a strong understanding of the evaluation of complex systems, the principles of causal inference, and their application to machine learning. You will have the opportunity to work on cutting-edge research projects in NLP, contributing to the development of more reliable and interpretable LLMs. It is important to note that the Ph.D. research project should be aligned with your interests and expertise. Therefore, the precise direction of the research can and will be influenced by the personal taste and research goals of the student. It is encouraged that you bring your unique perspective and ideas to the table.

Master degree in Natural Language Processing, computer science or data science.

Mastering Python programming and deep learning frameworks.

Experience in causal inference or working with LLMs

Very good communication skills in English, (proficiency in French  not mandatory).

The thesis will be conducted within the Getalp teams of the LIG laboratory (https://lig-getalp.imag.fr/). The GETALP team has a strong expertise and track record in Natural Language Processing. The recruited person will be welcomed within the team which offer a stimulating, multinational and pleasant working environment.

The means to carry out the PhD will be provided both in terms of missions in France and abroad and in terms of equipment. The candidate will have access to the cluster of GPUs of both the LIG. Furthermore, access to the National supercomputer Jean-Zay will enable to run large scale experiments.

The Ph.D. position will be co-supervised by Maxime Peyrard and François Portet.

Additionally, the Ph.D. student will also be working with external academic collaborators at EPFL and Idiap (e.g., Robert West and Damien Teney) and external industry partners (Microsoft Research)