Registration form
Attendance is free and open to everyone interested. Please register via the link above, and you will receive the Zoom meeting details one day before the seminar.
Prof. Frédéric Precioso, Université Côte d'Azur
By presenting the key principles of LLMs, we will expose how core mechanisms are finally not so complex and how they allow many failures to arise. We will present some of the most recent fancy techniques, such as self-distillation and Reinforcement Learning from Verifiable Reward (RLVR), and how they do not really solve the known weaknesses but potentially open new fields of application and pave the way to apply LLMs efficiently. We then provide insights on evaluation approaches and the persistent limited reliability of the models. We finally expose the ethical implication of such technical choices, presenting the studies of AI biases and fairness.