Both the human brain and artificial neural networks organize information in low-dimensional representational spaces, yet how this geometry supports learning remains unclear. Recent theoretical results suggest that low-dimensional representations enable faster convergence of empirical to population distributions under the Wasserstein distance, meaning that fewer samples are required to accurately capture the underlying data structure, thereby improving learning efficiency and generalization. We tested this hypothesis in artificial and biological systems. Across small supervised networks and large pretrained foundation models, lower intrinsic dimension was associated with smaller train–test distribution divergence and better generalization. In the human brain, this effect was region-specific: in higher-order cortical areas such as the Angular Gyrus, individuals with lower intrinsic dimension and more stable representational distributions across sessions showed stronger learning outcomes. Together, these findings reveal a shared geometric principle across brains and AI: low-dimensional representational organization accelerates distributional convergence and supports efficient generalization.

Social behavior in mice is critical for understanding their natural interactions and underlying neural mechanisms. Traditional Markov models, however, face limitations in capturing the sequential dynamics of body language associated with social behaviors. To address these challenges, we developed the body language-bidirectional encoder representation from transformers (BL-BERT) framework, which surpasses the Markov model in extracting complex sequential behavioral patterns. BL-BERT effectively differentiates the body language of the mice within different social interaction paradigms and produces results consistent with manual annotations. Notably, BL-BERT achieves higher extraction accuracy than the Markov model by reducing the complexity of the recurrent state transitions in behavior sequences. These advantages enable BL-BERT to accurately quantify high-order sequential behavioral structures in mice, paving the way for more detailed insights into the brain’s mechanisms controlling complex behavior.