Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do

This study investigates whether large language models (LLMs) exhibit syntactic convergence—the tendency to mirror the grammatical structures of an interlocutor—when interacting with humans. Using a substitution paradigm, the author replaced human turns in existing dialogues with generations from sixteen open-weight models (Llama and Gemma families). By comparing these model outputs to both the original human responses and unrelated conversational primes, the study measures the reuse of context-free grammar (CFG) rules to determine if models adapt their syntax to the preceding human turn.

All sixteen models demonstrated significant syntactic convergence, reusing rules from the preceding human turn more frequently than from unrelated random primes. Instruction-tuned models consistently showed higher total syntactic overlap with human interlocutors than the original human respondents did. However, this increased overlap is nuanced: instruction-tuned models also showed higher overlap with unrelated primes, suggesting a general tendency toward syntactic mimicry rather than specific, context-sensitive adaptation. When the analysis controlled for the total amount of syntactic structure (target size) in the model's output, instruction-tuned models actually displayed lower conditional propensity to reuse specific human rules compared to their pretrained variants.

These results challenge the assumption that instruction tuning simply makes models more "human-like" in their conversational dynamics. While instruction-tuned models appear more syntactically aligned with users, this behavior is partially an artifact of their generation style rather than a deeper, context-driven convergence mechanism. Understanding these differences is critical for researchers studying how LLMs influence human communication patterns, particularly regarding potential "amplification spirals" where models and users inadvertently reinforce each other's linguistic habits.