Reproduce “Representation of syntax in LLMs through the lens of linear distance and similarity-aware entropy” on OpenResearch

About this paper

Juan Pablo Vigneaux, Mary Kennedy, Khalil Iskarous, Robert Frank, Matilde Marcolli

This work disaggregates structural probe performance for individual syntactic relations in large language models, demonstrating that the fidelity of syntactic representation is predictably modulated by linear distance, its dispersion, and the similarity-aware entropy of the dependency's head. It quantifies how these linguistic and geometric factors influence syntactic encoding across various transformer architectures, highlighting the critical role of learned word order.

Key insights

• The representation of syntax in LLMs is not uniformly abstract; the success of structural probes in reconstructing relations is predictably modulated by linear distance, its dispersion, and the lexical diversity of head words.
• Linguistically interpretable groupings of syntactic relations emerge from hierarchical clustering based on their distinct UASL-vs-distance profiles, suggesting different processing patterns for various dependency types.
• The capacity for syntactic encoding, and the influence of the identified predictive factors, is critically dependent on learned word order, as evidenced by a substantial performance drop in a model trained on permuted sentences.

Summary by alphaXiv.