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

Problem Statement

Syntactic convergence—the tendency of speakers to adapt their grammatical constructions to match their interlocutors—is a well-documented phenomenon in human dialogue, typically operating below conscious awareness via structural priming mechanisms. While Large Language Models (LLMs) are now routine interlocutors, it remains an open question whether they exhibit analogous directional syntactic convergence toward human users relative to human baselines. Specifically, it is unclear if LLMs reuse the context-free grammar (CFG) rules of preceding human turns more than humans do, and how this behavior varies across model families, parameter scales, and training regimes (pretrained vs. instruction-tuned).

Methodology

The study utilizes a substitution paradigm based on data from Blevins et al. [2026], where LLM-generated turns replace one speaker's turns in pre-existing human-human dialogues from the DailyDialog corpus. This setup allows for direct comparison between model outputs and the original human responses at identical conversational positions.

Data and Models

• Corpus: 707 dyadic English conversations (restricted to turns 6–end), yielding 1,901 matched positions per model.
• Models: Sixteen open-weight models across two families (Llama-3.1/3.2 and Gemma-3), ranging from 1B to 70B parameters. Each family includes both pretrained (PT) and instruction-tuned (IT) variants.
• Conditions:
  1. HUMAN: Original human response ($h_i$) to human prime ($p_i$).
  2. ACTUAL: Model generation ($m_i$) to human prime ($p_i$).
  3. RANDOM: Model generation ($m_i$) to an unrelated human prime ($\tilde{p}_i$) sampled from a different conversation.

Measurement Pipeline

• Syntactic Annotation: Utterances were parsed using benepar to extract Context-Free Grammar (CFG) production rules. Eligible rules were non-unary and occurred more than once in the human corpus.
• Metrics:
  • Total Natural-Output Alignment: Overlap of CFG rules between the target and the actual prime.
  • Actual-vs-Random Increment: The added overlap attributable to the true conversational prime versus an unrelated one.
  • Conditional Reuse Propensity: Overlap odds when holding the target rule-set size constant.
• Statistical Modeling: Logistic mixed-effects models (GLMMs) were fitted for each model and pooled across architecture pairs. Random intercepts accounted for variation across positions and rules. Covariates included rule frequency and prime structural size.
• Exploratory Analyses: TF-IDF cosine similarity (lexical) and L2-normalized sentence embedding cosine similarity (semantic) were calculated to compare model and human similarity to the preceding turn.

Key Results

1. Universal Syntactic Reuse Above Baseline (RQ1)

Every model (all 16) demonstrated greater CFG-rule overlap with the actual preceding human turn than with an unrelated prime. The actual-versus-random coefficient ranged from $\beta = 0.706$ to $1.453$($p \le 6.5 \times 10^{-31}$), confirming that all evaluated models exhibit directional syntactic convergence.

2. Instruction-Tuned Alignment vs. Human Baseline (RQ2)

• Natural Output: Every instruction-tuned model showed greater total actual-prime overlap than the matched human response ($\beta = 0.398$ to $0.796$). Pretrained models were closer to the human baseline, with mixed results (four significantly positive, four indistinguishable from zero).
• Architecture Pairs: All eight matched architecture pairs showed greater total actual-prime reuse under instruction tuning compared to their pretrained counterparts.
• The "Structure" Confound: When decomposing these results, instruction-tuned models produced responses with more eligible syntactic structure (larger target rule sets), which inherently affords more opportunities for overlap.
  • Actual-vs-Random Increment: Instruction-tuned models showed a smaller incremental boost from the true prime compared to pretrained models ($\beta = -0.310$).
  • Conditional Reuse: When target rule-set size was held constant, the advantage of instruction tuning reversed. Pretrained models exhibited a stronger tendency to reuse available rules ($\beta = -0.505$ for IT vs. PT), while instruction-tuned models showed lower conditional reuse odds.

3. Frequency Profiles (RQ3)

• Rule Frequency: For all models, the actual-versus-random increment was larger for lower-frequency rules.
• Divergent Profiles: Humans and pretrained models showed negative interactions with frequency (stronger priming for rare rules). Instruction-tuned models showed more varied frequency interactions, with some exhibiting positive interactions, suggesting a different mechanism in how they weight rule reuse based on frequency.

4. Lexical and Semantic Similarity (RQ4)

• All sixteen models were more lexically and semantically similar to the preceding turn than the matched human responses were.
• Instruction-tuned models consistently produced responses with greater mean semantic similarity than their pretrained counterparts. Lexical similarity results were heterogeneous across architecture pairs.

Significance and Claims

The paper claims to provide the first measurement of directional local syntactic convergence in human-LLM dialogue across a broad inventory of constructions and model variants.

• Mechanism of Convergence: The findings suggest that instruction tuning optimizes models for local "responsiveness," leading to higher total alignment with the preceding turn. However, this is partly driven by the generation of responses containing more syntactic structure rather than a pure increase in the propensity to reuse specific rules.
• Reconciliation with Stylometry: The results reconcile with prior work (Blevins et al. [2026]) showing instruction-tuned models may converge less on certain stylometric features; the current study suggests this may reflect a tendency to introduce novel material while still maintaining high structural alignment through increased output complexity.
• Implications for Alignment Risks: The study operationalizes a local linguistic component relevant to the "amplification spiral" hypothesis (Augustin et al. [2026]), where linguistic alignment combined with hyperpersonalized generation may contribute to maladaptive belief construction. The data confirms that LLMs, particularly instruction-tuned ones, exhibit strong, measurable syntactic accommodation to human users.
• Psycholinguistic Insight: The presence of frequency-dependent priming effects (stronger for rare rules) in all models, including small pretrained ones, suggests that sensitivity to the local syntactic environment is a fundamental property of next-token prediction, potentially analogous to error-driven implicit learning in humans, though the paper notes this analogy requires further mechanistic investigation.

The study concludes that while instruction-tuned models align more strongly with human syntax in natural output, this alignment is complex, involving both increased structural opportunity and distinct frequency-weighting mechanisms compared to both humans and pretrained models.