INTERNAL WORLD MODELS IN HUMANS, MICE AND AI: COGNITIVE MAPS OF MULTI-DIMENSIONAL EVENT SEQUENCES

To navigate dynamic environments, humans must extract latent structures from sensory input to predict future states. Environmental features often change at different rates, with some dimensions remaining stable while others fluctuate. To test how feature stability shapes how we learn such hidden patterns, we developed a task where subjects observe sequences of objects. These objects are defined by three visual features—color, shape, and texture—each evolving according to independent hidden graph structures and transition probabilities, producing fast-, medium-, and slow-changing dimensions. Choice trials are interleaved between sequences, requiring subjects to choose between a correct target stimulus and a distractor that is inconsistent with the underlying graph structures on at least one dimension, with immediate feedback on the outcome. Pilot human data show that choice accuracy improves across sessions. Notably, performance was increasingly influenced by the distance between feature values of the distractor and the preceding stimulus object on all three stimulus dimensions, indicating an emergence of structured representations across all temporal scales. Large language models (LLMs) and tabular foundation models (TabPFN) tested through in-context learning revealed shared behavioral signatures; performance was modulated by the number of plausible dimensions of the distractor. We further adapted this task for behaving mice compatible with 2-photon microscopy, to establish them as a viable model for future investigation of the underlying neural circuits. This integrative approach provides a robust platform for assessing how feature stability influences the acquisition of latent structures and multi-scale prediction across species and AI.