Objective-Free Local Learning and Emergent Language Structure in Thinking Machines (2506.23293v1)

Abstract: We present a neuro-symbolic framework for generative language modeling based on local, event-driven emergent learning. At its core is a hierarchical Hopfield memory chain acting as a compositional short-term memory and dynamic tokenizer (retokenizer). Rather than relying on predefined tokens or supervision, the model builds structure from scratch, learning symbol sequences as multi-scale representations. It constructs projection tensors that bind co-occurring features into hierarchical tokens, introducing redundancy (i.e an emergent gauge structure) and enabling compression of local activations into long-range dependencies. Curiously, we find that the retokenizer can filter natural language patterns from noise, generating synthetic languages with coherent internal morphology -- quantifiably the same as human language. Language is learned in a local (Hebbian) fashion, where model constraints dictate allowed emergent structure, and new information is retained in alignment with this structure. The absence of a global objective enables a form of plasticity not found in conventional LLMs, allowing the system to generalize beyond its initial inference class -- even without explicit data. We demonstrate that briefly activating a new neuron during inference binds distributed multi-scale token features into a symbolic embedding. These emergent embedding neurons act as long-term memory and support a key-value mechanism for compositional inference and generalization. This architecture provides a methodological foundation for studying how symbolic structure can emerge from local neural learning. It offers a new pathway for building scalable, interpretable neuro-symbolic systems -- where tokens, grammar, and reasoning arise as compressed memory traces within a Hopfield hierarchy. This approach advances the development of neuromorphic architectures for generative LLMs.