Brain-Inspired AI Model Learns Abstract Structures Like Humans

In a new preprint on arXiv, Tianqiu Zhang, Muyang Lyu, Xiao Liu, and Si Wu present a world model inspired by the hippocampal-entorhinal circuit, which is known to represent both spatial and conceptual spaces in the brain. Their architecture simultaneously infers latent transitions and constructs a predictive visual world model from raw video streams. The key innovation is a dual-component design: an inverse model extracts structural abstractions, while a HPC-MEC coupling module dissociates relational structures (represented by the MEC-like component) from integrated episodic scenes (represented by the HPC-like component). This mirrors how the brain separates "where" and "what" information.

The researchers benchmarked their model on primitive transformation dynamics (e.g., moving shapes, rotating objects). Results show the model can abstract invariant structures from changing visual inputs and reuse those structures to predict outcomes in entirely new contexts — a phenomenon known as structural generalization. By leveraging velocity-driven path integration, the model maintains robust predictions even when visual features shift. This work provides a novel computational framework for self-supervised learning of world models, suggesting a path toward AI systems that acquire reusable abstract knowledge without massive labeled datasets.