Paper separates task performance from compositional feature learning

A new arXiv preprint argues that strong benchmark performance does not by itself show that a model has learned compositional, environment-invariant features. The authors aim to disentangle measured accuracy from the underlying representations that support out-of-distribution generalisation, a capability often treated as a marker of biological intelligence. Their analysis is framed around how systems can transfer invariant properties from training mappings to novel compositions.