The work “Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models” with Samuele Bortolotti, Emanuele Marconato, Andrea Passerini and Stefano Teso has been accepted at NeurIPS 2026!
This paper studies how to quantify uncertainty in neuro-symbolic concept-based models while respecting the logical constraints linking concepts and labels. We identify three key requirements for conformal prediction in this setting: consistency, coverage and conciseness. We then introduce COCOCO, a post-hoc framework that constructs prediction sets for concepts and labels jointly and reconciles them through logical reasoning, retaining distribution-free coverage guarantees. We finally showcase the effectiveness of our approach across eight datasets, obtaining competitive performance and compact prediction sets.
A final thanks to all the amazing co-authors involved in this project.
See you in Paris 🗼🥐🇫🇷!