AI RESEARCH

Compositional Neuro-Symbolic Reasoning

arXiv CS.AI

ArXi:2604.02434v1 Announce Type: new We study structured abstraction-based reasoning for the Abstraction and Reasoning Corpus (ARC) and compare its generalization to test-time approaches. Purely neural architectures lack reliable combinatorial generalization, while strictly symbolic systems struggle with perceptual grounding. We. therefore.