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

Researchers introduce Lexical Consensus, a framework testing whether AI agents can learn and stabilize new word meanings from visual experience. Results show a perceptual-coherence gradient where learning success depends on visual similarity rather than semantic relatedness, revealing fundamental constraints on how frozen neural representations enable or limit language acquisition.

This research addresses a foundational question in AI cognition: can artificial agents genuinely acquire meaning through grounded experience, or merely perform pattern matching? The Lexical Consensus framework provides empirical evidence that perceptual distance—not semantic relationships—governs how effectively agents learn artificial labels for visual concepts. The pre-registered CIFAR-100 experiment yields particularly strong results, with perceptual distance explaining 24.5% of learning variance while semantic distance contributes virtually nothing.

The work builds on decades of cognitive science examining how humans learn words through perception, applying these principles systematically to neural networks. Prior studies on language emergence in multiagent systems rarely isolate perceptual constraints from semantic reasoning. This paper's contribution lies in demonstrating that frozen visual embeddings from DINOv2 simultaneously enable lexical grounding and fundamentally limit what can be learned without architectural adaptation.

For AI development, these findings suggest that scaling model size or improving general capabilities may not overcome perceptual geometry constraints. The bidirectional evaluation revealing disparities between naming and retrieval introduces a memory-fidelity dimension previously underexamined in language grounding literature. This distinction matters for applications requiring reliable label-to-image mapping versus open-ended naming tasks.

Future work should explore representational restructuring—whether agents can reorganize frozen embeddings through learning mechanisms or require dynamic, trainable perceptual systems. The work implies that multimodal AI systems may require explicit architectural choices to move beyond perceptual constraints, particularly for acquiring concepts with non-obvious visual coherence.

• →AI agents learning new words are fundamentally constrained by visual similarity, not semantic relationships between concepts
• →Frozen perceptual embeddings enable lexical grounding but create ceiling effects preventing acquisition of visually disjunctive categories
• →Naming and retrieval operate through distinct mechanisms, with exemplar-based approaches outperforming prototype centroids in image retrieval tasks
• →Perceptual distance predicts learning accuracy with statistical significance (R² = 0.245, p < 1e-7) across controlled experimental conditions
• →Current multimodal AI architectures may require representational adaptation capabilities to overcome inherent perceptual geometry limitations