This research paper proposes a unified framework for Riemannian deep learning, generalizing batch normalization and multinomial logistic regression to broad classes of Lie groups and gyrogroups. It also introduces novel neural networks for geometric representations, including hyperbolic space and full-rank correlation matrices. This work has potential implications for developing more robust and adaptive AI models.

Researchers introduced a diagnostic experiment to isolate failures in multimodal LLMs' perception and reasoning on physics and geometry benchmarks. The results show that incorrect diagram interpretation can lead to performance degradation, even on tasks models can solve from text alone. The study found that recovery under corrected captions is high for some models, allowing for the separation of perception-blocked failures from genuine reasoning bottlenecks.