Neurosymbolic reasoning shortcuts under the independence assumption

Neurosymbolic predictors often assume independence among symbolic concepts to simplify probabilistic reasoning. That shortcut speeds inference, but it can also mask uncertainty in the combinations that matter most.

This work formalizes the limitation: if you assume independence among concepts, the model cannot represent uncertainty over certain concept combinations. As a result, the system can appear correct on downstream tasks while relying on brittle reasoning shortcuts.

The paper clarifies when the independence assumption actually constrains NeSy systems, and why that matters for trustworthy reasoning under uncertainty.

Paper: https://proceedings.mlr.press/v284/krieken25a.html