Lexical Consensus: Grounded Word Learning and Shared Meaning in Artificial Agents

The paper introduces Lexical Consensus, a framework for grounded word learning in AI agents, utilizing frozen DINOv2 visual embeddings and Carroll-style nonce words. Key findings reveal a robust perceptual-coherence gradient in lexical acquisition, where native categories are learned most easily, and perceptual distance is a significant predictor of acquisition accuracy, while semantic distance has negligible impact. This research highlights the importance of perceptual geometry in grounding lexical meanings, which is crucial for practitioners developing AI systems that require effective word learning and concept generalization.