Learning Through Explanations: Cooperative Dialogue Between Models

Day - Time: 16 December 2026, h.11:00

Place: Area della Ricerca CNR di Pisa - Room: C-40

Speakers

• Carlo Metta

Referent

Giulio Del Corso

Abstract

Explainability in machine learning has mostly been used in one direction: either to communicate a model’s reasoning to humans or to regularize a model through its own explanations. Here we study a different setting, in which multiple models exchange explanations during cooperative learning. We treat explanations astraining signals that allow heterogeneous models to influence one another through the way they justify their predictions. We introduce a cooperative training framework based on directional explanation alignment. An external reference model acts as a devil’s advocate and moderates the exchange, limiting collapse toward overly similar explanations or unbalanced influence. Our results suggest that explanatory interaction can improve model ensemble performance, the quality and robustness of model explanations and can also support cooperation between models during learning.