Study challenges core assumption in neuro-symbolic AI: symbol grounding alone does not produce compositional reasoning

Researchers introduced the Iterative Logic Tensor Network (iLTN), a fully differentiable architecture designed for multi-step deduction, and used it to systematically test whether symbol grounding automatically produces reasoning capability.

A model trained solely on grounding failed to generalize, but iLTN trained jointly on perceptual grounding and multi-step reasoning achieved high zero-shot accuracy across tasks testing novel entities, unseen relations, and complex rule compositions (a taxonomy of generalization).

The findings establish that symbol grounding, while necessary, is insufficient for generalization — reasoning is a distinct capability requiring an explicit learning objective, not an emergent property.