Comparative viability of neuro-symbolic hybrids versus strictly neural approaches

Determine whether neuro-symbolic hybrid architectures that incorporate explicit symbolic representations and operations ultimately outperform strictly neural architectures on compositional generalization—especially at scale—and specify the conditions under which hybrids “win out” relative to data-driven approaches such as metalearning and large-scale pretraining.

Background

The review contrasts neuro-symbolic hybrids, which embed explicit symbolic modules, with strictly neural architectural biases, metalearning, and pretraining-based approaches. While recent results suggest explicit symbolic components may not be necessary for specific compositional tasks, the authors note that it remains unsettled whether hybrids will prove superior as tasks and scales increase. Resolving this question would clarify the role of symbolic structure in practical AI and cognitive modeling.