Their work, presented at the 2024 Conference on Language Modeling in Philadelphia, focused on nonverbal abstract reasoning tasks—challenges requiring both visual understanding and logical deduction.