Inside the Black Box

WEBFARE SEMINARS

📅 15th June 2026
🕓 15:00
📍 Scienza Nuova Institute, Corso Montevecchio 38, Torino

In this Webfare webinar, Samuele Tonati will discuss counterfactual explanations. This is a fundamental paradigm in the field of explainable AI, offering insights into model decisions by identifying the smallest input perturbations capable of altering predictions. In this session, we will examine their general formulation, the most common evaluation metrics, and the challenges that arise when moving to textual data. We will then introduce SCCoT, a new supervised method in which the model is guided by control tokens to generate counterfactual explanations of its own predictions. More generally, we will discuss the connections with counterfactual approaches in mechanistic interpretability, where activation patching is used to perform causal interventions on the models’ internal representations.