Reasoning foundation models

Contact person: Benjamin Ricaud

Status: Project

In this project, we are interested in how the thinking and decision process takes place inside a Large Language Models as it generates text. Understanding it will give us the ability to control, steer and align an LLM with more accuracy and flexibility than doing it during the training phase. We propose to act directly inside the neural network layers to modify the signals and change the model behavior. This direction of research is developing fast but is so far purely empirical with no clear principles. Progress in this direction can lead to better alignment of models, blocking the generation of harmful content, improving reasoning or reducing hallucinations.