We don’t learn the world as it is. We rapidly infer what matters, with belief and attention reciprocally shaping each other as neural geometry adapts.

Thanks to @stephanie-theves.bsky.social, @mikael-johansson.bsky.social, Peter Gärdenfors, @doellerlab.bsky.social.

www.biorxiv.org/content/10.1...

Critically, inference stretches neural distances along relevant dimensions and compresses irrelevant ones right before a decision, predicts faster RTs, and this re-shaping precedes feedback-related frontal theta tracking model-derived PE.

The brain’s representational space flexes with inferred complexity.

Neural effective dimensionality scales up in 2D vs 1D, and is higher on correct vs incorrect trials. In 2D, the two attended features show up as near-orthogonal axes in a shared planar manifold plane.
6/8

Eyes tell the same story
Gaze selectively shifts toward task-relevant features, irrelevant features drop out. Gaze entropy decreases as beliefs stabilise, and negative prediction errors from the HSI model trigger broader sampling (exploration), while positive PEs tighten focus (exploitation).
5/8

A Hidden State Inference (HSI) model best explained choices and inferred contexts, beating Q-learning variants (standard, forgetting, counterfactual).

HSI captures something structurally different from incremental RL.
4/8

Participants adapted fast: first trial after a switch was at chance level, then rapid recovery. RTs drop and accuracy rises within context blocks - they used the structure to take decisions.
3/8

Serial reversal learning task with same cars, same feature space (3 dimensions), but the rule silently flips. Different dimensions matter in different trials. Sometimes one dim matters, sometimes two dims. You only find out via feedback, meaning participants had to infer the latent state.
2/8

Having worked at the Max Planck Institute for Human Development in Berlin, it is hard not to find this advice laughable coming from a group that seems to foster anything but open culture and flat hierarchies.