On the Emergence of Syntax by Means of Local Interaction

The more striking detail buried in the framing is that no syntactic structure was explicitly taught: the model received only a binary boundary signal and self-organized into something functionally resembling a classical parsing algorithm. That's a claim about emergence, not just compression, and it puts the burden of proof on whether the internal structure is genuinely algorithmic or a statistical artifact of the arithmetic grammar's regularity.

This sits in productive tension with the finding from 'Heterogeneity in Formal Linguistic Competence of Language Models' (arXiv, April 20), which showed that GPT-2 Small fails specific grammatical constructions not because of architectural limits but because of data gaps. That paper argues scale and data drive formal competence. This paper argues the opposite direction: that a tiny model with almost no signal can recover grammatical structure through local interaction alone. The two results don't contradict each other, but they're asking different questions about where linguistic structure lives, in the data, in the architecture, or in the dynamics of learning itself.

The real test is whether the same emergent CKY-like structure appears when the grammar is swapped for a natural-language fragment with genuine ambiguity. If it does, the local-interaction hypothesis has legs; if the model collapses on ambiguous inputs, the result is specific to deterministic arithmetic grammars and its scope is narrow.