AVA Framework Reveals Embedding Limitations in Ontological Reasoning

The AVA framework uses 171,007 contrastive triplets across 163 ontologies to test if embeddings capture symbolic logic. Evaluation of 25 state-of-the-art models shows the best achieves only 0.739 triplet accuracy, with hard negative accuracy dropping to 0.135.