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

# Instruction-Tuned Models Locally Reuse Human Syntax More Than Humans Do
Source: [https://arxiv.org/html/2607.26015](https://arxiv.org/html/2607.26015)
###### Abstract

Syntactic convergence — the tendency of speakers to adapt in language towards the grammatical profiles of their interlocutors — is a well\-documented feature of human dialogue widely considered to operate below conscious awareness\. Whether large language models exhibit analogous syntactic convergence towards human users relative to human baselines \(and across a broad range of syntactic constructions\) remains an open question\. Using substitution\-paradigm data from Blevins et al\. \[2026\] in which model generations replace one speaker’s turns in pre\-existing human dialogues, this study measures turn\-adjacent reuse of context\-free grammar \(CFG\) rules across sixteen open\-weight Llama and Gemma models \(1B–70B, pretrained and instruction\-tuned\) at 1,901 matched positions per model\. Every model showed greater CFG\-rule overlap with the preceding human turn than with a sampled unrelated human prime, and in every model this actual\-versus\-random difference was larger for lower\-frequency rules\. Each instruction\-tuned model also showed greater natural\-output overlap with the actual prime than the human response it replaced, and all eight matched architecture pairs exhibited greater actual\-prime overlap after instruction tuning\. However, relative to pretrained variants, instruction\-tuned outputs overlapped more with unrelated primes, showed a smaller actual\-versus\-random increment, and had lower conditional rule\-reuse odds once target rule\-set size was held constant\. In exploratory analyses, each model exhibited greater mean lexical and semantic similarity to the preceding turn than the matched human responses did\. Instruction\-tuned models additionally produced responses with greater mean semantic similarity than their pretrained counterparts in all eight architecture pairs, whereas the lexical similarity results were more heterogeneous\.

## 1Introduction

When a speaker mirrors the grammatical constructions of her interlocutor, she engages in a process called syntactic convergence\. For instance, after hearing “I gave you the book,” she may be more likely to respond with “you sent your mother the letter” \(reusing the double\-object dative construction\) than with the equally grammatical “you sent the letter to your mother”\(Bock,[1986](https://arxiv.org/html/2607.26015#bib.bib1); Braniganet al\.,[2000](https://arxiv.org/html/2607.26015#bib.bib6)\)\. Syntactic convergence is widely considered to be relatively impervious to conscious awareness and is often explained through structural priming mechanisms\(Bock,[1986](https://arxiv.org/html/2607.26015#bib.bib1); Pickering and Garrod,[2004](https://arxiv.org/html/2607.26015#bib.bib13)\)\.

As entities capable of producing syntactically coherent text, large language models \(LLMs\) are now routine interlocutors at population scale\. Symmetrical LLM syntactic alignment\(Kandraet al\.,[2025](https://arxiv.org/html/2607.26015#bib.bib17); Mayoret al\.,[2025](https://arxiv.org/html/2607.26015#bib.bib8); Chenet al\.,[2026](https://arxiv.org/html/2607.26015#bib.bib19)\)and structural priming\(Sinclairet al\.,[2022](https://arxiv.org/html/2607.26015#bib.bib31); Michaelovet al\.,[2023](https://arxiv.org/html/2607.26015#bib.bib9); Jumeletet al\.,[2024](https://arxiv.org/html/2607.26015#bib.bib18); Caiet al\.,[2024](https://arxiv.org/html/2607.26015#bib.bib30)\)have been studied before, as has stylometric accommodation of LLMs\(Blevinset al\.,[2026](https://arxiv.org/html/2607.26015#bib.bib15)\)\(including towards humans\(Blevins,[2026](https://arxiv.org/html/2607.26015#bib.bib16)\)\)\. However, whether and how LLMs syntactically converge directionally towards humans – relative to matched human responses and across a broad inventory of syntactic constructions – remains, to date, unanswered\.

Analyzing data fromBlevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\)’s substitution paradigm, I use a measurement pipeline for directional local syntactic convergence in human\-LLM dialogue and apply it to dyadic conversations from sixteen open\-weight models spanning two families, parameter counts ranging from 1B to 70B, with both pretrained and instruction\-tuned counterparts\. Specifically, I measure whether LLM\-generated turns reuse the context\-free grammar \(CFG\) rules of the human turns to which they are responding, both relative to an unrelated\-prime baseline and relative to the original human responses that the LLM output replaces\.

This comparison yields four principal results\. First, all sixteen models reuse actual\-prime CFG rules above an unrelated\-prime baseline \(RQ1\)\. Second, every instruction\-tuned model shows greater natural\-output syntactic alignment than the matched human response \(although decomposition by unrelated\-prime overlap and target structural opportunity qualifies a simple account of stronger copying under instruction tuning\) \(RQ2\)\. Third, rule frequency appears to shape actual\-prime reuse differently across humans, pretrained models, and instruction\-tuned models \(RQ3\)\. Finally, each of the sixteen models showed greater mean lexical and semantic similarity to the preceding turn than the matched human responses did \(RQ4\)\.

It is worth being precise about the terminology used in this paper\. I use*convergence*\(which may have both long\-term and short\-term instantiations – priming being an example of the latter\) to refer specifically to accommodation towards another interlocutor, and*alignment*to describe the resulting state or degree of similarity\. Importantly, greater rule overlap with an actual prime can arise because a response reuses syntax more strongly, because it contains more structural material and therefore affords more opportunities for overlap, or both\. I use*total natural\-output alignment*for the amount of actual\-prime overlap in responses as generated,*actual\-versus\-random increment*for the added overlap attributable to the true conversational prime relative to an unrelated one, and*conditional reuse propensity*for analyses that hold target rule\-set size constant\.

## 2Background

While communicative mirroring occurs across a number of behaviors \(both linguistically and paralinguistically\), syntactic convergence is a theoretically distinctive convergence dimension\. Grammatical form can recur independently of lexical content\(Bock,[1986](https://arxiv.org/html/2607.26015#bib.bib1)\)\. Historically, a dominant experimental paradigm for measuring syntacticpriminghas involved manipulating pre\-specified constructions and coding, in a binary fashion, whether a participant subsequently produced the same form\(Bock,[1986](https://arxiv.org/html/2607.26015#bib.bib1); Braniganet al\.,[2000](https://arxiv.org/html/2607.26015#bib.bib6)\)\. Context\-free grammar \(CFG\)\-based corpus methods instead represent utterances as production rules extracted from constituency parse trees, permitting broad\-coverage measurement without pre\-specifying constructions for the analysis\(Reitter and Moore,[2014](https://arxiv.org/html/2607.26015#bib.bib14)\)\.

To measure LLM linguistic accommodation,Blevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\)introduce a substitution paradigm that replaces one speaker’s turns in pre\-existing human–human dialogues, permitting direct comparison between an LLM\-generated response and what a human speaker actually said at the identical conversational position\. Their stylometric analyses \(including of token novelty and utterance length\) show that model and human convergence profiles differ by feature and tuning regime\. By contrast,Kandraet al\.\([2025](https://arxiv.org/html/2607.26015#bib.bib17)\)examine syntactic structure, restrict their analysis tointra\-LLM conversations, and use a symmetrical aggregate measure\. Other experiments have provided complementary evidence of human\-like structural priming signatures in language models including inverse\-frequency effects\(Jumeletet al\.,[2024](https://arxiv.org/html/2607.26015#bib.bib18)\), but have not compared generated dialogue responses against matched human respondents across a range of syntactic constructions\. The need to understand LLM \(syntactic\) accommodation is increasingly pressing, asAugustinet al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib32)\)propose an "amplification spiral" in which linguistic alignment, hyperpersonalized generation, and sycophancy may co\-occur in the construction of delusional user beliefs\.

The present study, then, asks four questions:

1. RQ1:Do LLMs reuse the CFG rules of the immediately preceding human turn above an unrelated\-prime baseline?
2. RQ2:Do LLMs reuse CFG rules more or less than a matched human respondent would at the same turn position?
3. RQ3:Does reuse vary with rule frequency in the same way across human, pretrained, and instruction\-tuned responses?
4. RQ4:Do human–model differences in similarity to the preceding turn appear at the lexical and semantic levels as well as the syntactic one?

## 3Methods

### 3\.1Data and substitution paradigm

Human turns are drawn from DailyDialog\(Liet al\.,[2017](https://arxiv.org/html/2607.26015#bib.bib20)\), a dataset of dyadic English conversations, along with the substituted versions released byBlevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\)in which LLM\-generated completions replace one speaker’s utterances from turn six onward\. Following their preprocessing, conversations were restricted to two\-speaker dialogues with at least six turns, yielding 707 conversations; their model generations and preprocessing code are publicly released\. The full set of 707 conversations was analyzed for every model\.

Sixteen open\-weight language models are evaluated: the Llama\-3\.1 \(8B, 70B\) and Llama\-3\.2 \(1B, 3B\) families\(Grattafioriet al\.,[2024](https://arxiv.org/html/2607.26015#bib.bib2)\)and the Gemma\-3 family \(1B, 4B, 12B, 27B\)\(Gemma Teamet al\.,[2025](https://arxiv.org/html/2607.26015#bib.bib3)\), each in both its pretrained and instruction\-tuned variant\. Because every model completes the identical conversations and because the human turns are held fixed across all versions, responses are directly comparable across models and against the original human responses in the same conversational contexts\. Generated responses were capped at 40 new tokens\. From the second substituted position onward, a model’s generation is conditioned in part on its own earlier generations; the immediately preceding turn itself is nevertheless always human in the analyzed prime–target pair\.

Ultimately, there were 1,918 substituted positions per model\. Seventeen positions had no eligible rule in the immediately adjacent human prime and were excluded, leaving 1,901 exactly matched positions per model\. No non\-adjacent prime was substituted for a missing one\. Human and model targets with no eligible rules were retained and contributed zero outcomes for every eligible prime rule\.

### 3\.2Syntactic annotation and rule eligibility

All utterances, both human and LLM\-generated, were parsed using benepar’sbenepar\_en3\_largeconstituency parser\(Kitaev and Klein,[2018](https://arxiv.org/html/2607.26015#bib.bib21); Kitaevet al\.,[2019](https://arxiv.org/html/2607.26015#bib.bib22)\), with spaCy\(Montaniet al\.,[2023](https://arxiv.org/html/2607.26015#bib.bib5)\)used for tokenization and segmentation\. Prior to parsing, utterance text was normalized by collapsing whitespace sequences to a single space\.

A constituency parse represents an utterance as a tree whose nonterminal nodes encode syntactic categories such as S, NP, and VP\. For example, a parse of “The reader reviewed the paper” contains the productions

S→NP​VP,NP→DT​NN,VP→VBD​NP\.\\mathrm\{S\}\\rightarrow\\mathrm\{NP\}\\ \\mathrm\{VP\},\\qquad\\mathrm\{NP\}\\rightarrow\\mathrm\{DT\}\\ \\mathrm\{NN\},\\qquad\\mathrm\{VP\}\\rightarrow\\mathrm\{VBD\}\\ \\mathrm\{NP\}\.
A CFG production records how one nonterminal node expands into its immediate children\. Preterminal part\-of\-speech nodes expand directly to terminal words, as inDT→\\rightarrowThe\. For an utteranceuu, letρ​\(u\)\\rho\(u\)denote the multiset of CFG productions extracted from its sentence\-level parse trees\.R​\(u\)R\(u\)is the set of distinct eligible rule types appearing inρ​\(u\)\\rho\(u\), where a rule is eligible if it is non\-unary and occurs more than once in the original human corpus\. Thus, repeated occurrences of the same rule within an utterance contribute only once\. Frequencies were calculated once from human turns only, preventing model outputs from changing rule eligibility or the frequency scale\. \(Lexical preterminal productions are unary and are therefore already excluded\)\. The human corpus contained 10,887 rule types in total, of which 1,405 met the global eligibility criterion; 1,138 appeared in the analyzed prime sets\.

### 3\.3Matched actual, human, and unrelated\-prime conditions

For matched positionii, writepip\_\{i\}for the prime \(the human turn att−1t\-1\),hih\_\{i\}for the original human response attt,mim\_\{i\}for the model generation attt, andp~i\\tilde\{p\}\_\{i\}for a human prime sampled with a fixed seed from a different conversation\. Onep~i\\tilde\{p\}\_\{i\}was drawn per position and shared across all sixteen models \(to ensure that model comparisons do not inherit different random baselines\)\. The three conditions pair\(pi,hi\)\(p\_\{i\},h\_\{i\}\)in thehumancondition,\(pi,mi\)\(p\_\{i\},m\_\{i\}\)in theactualmodel condition, and\(p~i,mi\)\(\\tilde\{p\}\_\{i\},m\_\{i\}\)in therandomcondition\. Thus, the human and model targets respond to the same potential source of syntactic priming, while the unrelated\-prime condition holds the model response fixed and varies only the prime\.

For each conditionκ\\kappaat positioniiand each eligible prime rulerjr\_\{j\}, the binary outcome is

yi​jκ=𝟏​\[rj∈R​\(targetiκ\)\]\.y\_\{ij\}^\{\\kappa\}=\\mathbf\{1\}\\\!\\left\[r\_\{j\}\\in R\(\\mathrm\{target\}\_\{i\}^\{\\kappa\}\)\\right\]\.\(1\)Rule frequency was represented as mean\-centeredlog⁡FH​\(rj\)\\log F\_\{H\}\(r\_\{j\}\), and prime structural size as mean\-centeredlog⁡\|R​\(primei\)\|\\log\|R\(\\mathrm\{prime\}\_\{i\}\)\|\. No additive constant was required before either logarithm: eligible rules satisfyFH​\(r\)\>1F\_\{H\}\(r\)\>1, and positions with empty prime rule sets were excluded before analysis\. In the structural\-opportunity analyses, target size was represented as mean\-centeredlog⁡\(\|R​\(targeti\)\|\+1\)\\log\(\|R\(\\mathrm\{target\}\_\{i\}\)\|\+1\)\. \(The\+1\+1offset is required only for target size because a valid human or model response may contain no eligible rules\)\. This procedure produced 39,222 rule\-level observations per model, or 627,552 observations across the sixteen per\-model datasets\.

### 3\.4Statistical models

For each model separately, a logistic mixed\-effects model was fitted over conditionκ∈\{human,actual,random\}\\kappa\\in\\\{\\textsc\{human\},\\textsc\{actual\},\\textsc\{random\}\\\}\. Letxj=log⁡FH​\(rj\)~x\_\{j\}=\\widetilde\{\\log F\_\{H\}\(r\_\{j\}\)\}denote the mean\-centered log frequency of rulerjr\_\{j\}, and letqiκ=log⁡\|R​\(piκ\)\|~q\_\{i\}^\{\\kappa\}=\\widetilde\{\\log\|R\(p\_\{i\}^\{\\kappa\}\)\|\}denote the mean\-centered log size of the relevant prime rule set\. Here,piκ=pip\_\{i\}^\{\\kappa\}=p\_\{i\}in thehumanandactualconditions, whereaspiκ=p~ip\_\{i\}^\{\\kappa\}=\\tilde\{p\}\_\{i\}in therandomcondition\.

Usinghumanas the reference condition, the linear predictor was

logit⁡Pr⁡\(yi​jκ=1\)=\\displaystyle\\operatorname\{logit\}\\Pr\\\!\\left\(y\_\{ij\}^\{\\kappa\}=1\\right\)=\{\}β0\+βA​1​\{κ=actual\}\+βR​1​\{κ=random\}\\displaystyle\\beta\_\{0\}\+\\beta\_\{\\mathrm\{A\}\}\\,\\mathbf\{1\}\\\!\\left\\\{\\kappa=\\textsc\{actual\}\\right\\\}\+\\beta\_\{\\mathrm\{R\}\}\\,\\mathbf\{1\}\\\!\\left\\\{\\kappa=\\textsc\{random\}\\right\\\}\+βf​xj\+βA​f​1​\{κ=actual\}​xj\+βR​f​1​\{κ=random\}​xj\\displaystyle\+\\beta\_\{f\}x\_\{j\}\+\\beta\_\{\\mathrm\{A\}f\}\\,\\mathbf\{1\}\\\!\\left\\\{\\kappa=\\textsc\{actual\}\\right\\\}x\_\{j\}\+\\beta\_\{\\mathrm\{R\}f\}\\,\\mathbf\{1\}\\\!\\left\\\{\\kappa=\\textsc\{random\}\\right\\\}x\_\{j\}\+βq​qiκ\+bi\+cj,\\displaystyle\+\\beta\_\{q\}q\_\{i\}^\{\\kappa\}\+b\_\{i\}\+c\_\{j\},\(2\)
wherebi∼𝒩​\(0,σposition2\)b\_\{i\}\\sim\\mathcal\{N\}\(0,\\sigma\_\{\\mathrm\{position\}\}^\{2\}\)is a random intercept for matched position andcj∼𝒩​\(0,σrule2\)c\_\{j\}\\sim\\mathcal\{N\}\(0,\\sigma\_\{\\mathrm\{rule\}\}^\{2\}\)is a random intercept for rule\.

Under this parameterization,βA\\beta\_\{\\mathrm\{A\}\}is the actual\-model\-versus\-human contrast at centered mean rule frequency and prime size, andβA​f\\beta\_\{\\mathrm\{A\}f\}is the corresponding difference in the frequency slope\. The actual\-versus\-random contrast is

βA−R=βA−βR,\\beta\_\{\\mathrm\{A\-R\}\}=\\beta\_\{\\mathrm\{A\}\}\-\\beta\_\{\\mathrm\{R\}\},\(3\)
and the interaction testing whether that increment varies with rule frequency is

βA−R×f=βA​f−βR​f\.\\beta\_\{\\mathrm\{A\-R\}\\times f\}=\\beta\_\{\\mathrm\{A\}f\}\-\\beta\_\{\\mathrm\{R\}f\}\.\(4\)
For reporting, the same specification was also fit withrandomas the reference condition, which yields the actual\-versus\-random contrast and its standard error directly\. Because condition interacts with frequency, each condition coefficient is evaluated at centered mean rule frequency\.

A per\-model sensitivity fit added the centered target\-size termtiκt^\{\\kappa\}\_\{i\}as a main effect, yielding the conditional contrastβA\|t\\beta\_\{A\|t\}\. This fit asks how condition differences change when the amount of eligible syntactic structure in the target is held constant \(and since target size is produced during generation, these estimates are interpreted as descriptive conditional comparisons rather than causal adjustments\)\.

Comparisons between sixteen separately estimated coefficients do not themselves constitute a formal test of instruction tuning\. I therefore grouped the models into eight matched architecture pairs and fit two pooled GLMMs over model\-target rows\. The actual\-prime model \(N=212,544N=212\{,\}544\) included architecture pair, tuning, centered log rule frequency, centered log prime size, the tuning\-by\-frequency interaction, and random intercepts for position, pair\-by\-position, and rule\. This model estimates the instruction\-minus\-pretraining difference in total actual\-prime overlap within the evaluated model suite\.

A second pooled model used the actual\- and unrelated\-prime model rows \(N=415,008N=415\{,\}008\) and included the full tuning\-by\-prime\-type\-by\-frequency interaction\. Its tuning\-by\-actual\-prime coefficient tests whether instruction tuning changes the actual\-versus\-random increment, while the three\-way interaction tests whether that change varies with rule frequency\. As a complementary architecture\-level consistency check, I calculated the raw actual\-prime reuse difference within each matched pair and applied an exact two\-sided sign test\.

To separate total natural\-output alignment from conditional reuse propensity, I fit a pooled actual\-prime model containing the tuning\-by\-target\-size interaction\. The full model used all 212,544 rule rows\. Because pretrained and instruction\-tuned target\-size distributions differed substantially, I also fit the same specification within a central common\-support sample\. For each architecture pair, positions were retained when their log target size lay in the intersection of the pretrained and instruction\-tuned 5th–95th percentile ranges, yielding 149,101 rule rows\. Contrasts at the 10th, 50th, and 90th percentiles of the supported target\-size distribution were calculated as linear combinations of the tuning and tuning\-by\-target\-size coefficients\.

All models were fit in R usinglme4::glmer\(\)\(Bateset al\.,[2015](https://arxiv.org/html/2607.26015#bib.bib24)\), with binomial errors, thebobyqaoptimizer, 200,000 maximum function evaluations, andnAGQ=1\\mathrm\{nAGQ\}=1\.

### 3\.5Lexical and semantic similarity

Beyond the primary inferential analysis, two exploratory analyses were conducted to better characterize the observed convergence behavior\. Lexical similarity was measured as TF–IDF cosine similarity using one vocabulary shared across all primes, human targets, and model targets \(6,481 terms\)\. Semantic similarity was measured as the cosine between L2\-normalizedall\-MiniLM\-L6\-v2sentence embeddings\(Reimers and Gurevych,[2019](https://arxiv.org/html/2607.26015#bib.bib23)\)\. For each matched position, I calculated human\-minus\-model similarity \(so negative values indicate that the model response is closer to the prime\)\. For each model and metric, the mean difference was estimated with an intercept\-only OLS regression and conversation\-clustered standard errors\. Holm correction\(Holm,[1979](https://arxiv.org/html/2607.26015#bib.bib25)\)was applied separately across the sixteen lexical tests and the sixteen semantic tests\. Seventy\-eight blank model outputs were excluded, leaving 30,338 valid model\-position comparisons\.

## 4Results

### 4\.1Dataset and fit diagnostics

The analysis contains 1,901 matched target positions per model, 39,222 rule\-level observations per model, and 627,552 rows across the sixteen per\-model datasets\. All forty\-eight per\-model fits were nonsingular and produced no convergence messages; diagnostics for the four pooled fits appear in Table[7](https://arxiv.org/html/2607.26015#A5.T7)\.

### 4\.2Primary result I: all sixteen models exceed the unrelated\-prime baseline \(RQ1\)

Every model reused actual\-prime rules more than rules from its unrelated prime\. The actual\-versus\-random coefficient ranged fromβ=\.706\\beta=\.706to1\.4531\.453, with allp≤6\.5×10−31p\\leq 6\.5\\times 10^\{\-31\}\(Appendix[A](https://arxiv.org/html/2607.26015#A1)\)\. Thus, all sixteen evaluated models showed greater syntactic reuse from the true preceding human turn than from an unrelated human turn\.

### 4\.3Primary result II: instruction\-tuned models show greater natural\-output alignment than matched humans, while pretrained results are mixed \(RQ2\)

Under the human\-reference primary model, every instruction\-tuned model reused actual\-prime rules more than the matched human response at centered mean frequency:β=\.398\\beta=\.398–\.796\.796, allp≤8\.3×10−14p\\leq 8\.3\\times 10^\{\-14\}\. The largest coefficient belonged to Llama\-3\.2\-1B\-Instruct \(β=\.796\\beta=\.796\)\. Scale effects were not monotonic across families, however, so these data do not warrant a more general claim that natural\-output syntactic alignment varies with parameter count \(Figure[1](https://arxiv.org/html/2607.26015#S4.F1)a\)\.

Pretrained models were closer to the human baseline as well as more heterogeneous; their coefficients ranged from−\.030\-\.030to\.240\.240, and four were significantly positive while four were not distinguishable from zero\. All eight matched architecture pairs showed greater actual\-prime reuse under instruction tuning \(Appendix[3](https://arxiv.org/html/2607.26015#A2.T3)\), with raw instruction\-minus\-pretraining differences of\.022\.022–\.094\.094and an exact two\-sided sign\-testp=\.0078p=\.0078\. The pooled rule\-level GLMM estimated a corresponding within\-suite instruction\-minus\-pretraining contrast ofβ=\.447\\beta=\.447, 95% CI\[\.412,\.481\]\[\.412,\.481\], andOR=1\.56\\mathrm\{OR\}=1\.56\.

![Refer to caption](https://arxiv.org/html/2607.26015v1/x1.png)Figure 1:Primary and pooled syntactic results\.\(a\) Per\-model actual\-prime contrasts against the matched human response at centered mean rule frequency; error bars are 95% Wald intervals\. \(b\) Pooled instruction\-minus\-pretraining coefficients under four estimands\. Positive values favor instruction tuning\. Conditional coefficients are evaluated at the target\-size centering mean\.#### Decomposing the instruction\-tuning difference\.

Nonetheless, the actual\-versus\-random pooled model complicates a simple account in which instruction tuning straightforwardly strengthens CFG reuse from the immediately preceding turn\. Under the unrelated\-prime reference, instruction\-tuned outputs already had higher overlap than pretrained outputs \(β=\.743\\beta=\.743\)\. The pretrained actual\-versus\-random increment was strongly positive \(β=1\.177\\beta=1\.177\), but the tuning\-by\-actual\-prime interaction was negative \(β=−\.310\\beta=\-\.310, 95% CI\[−\.379,−\.240\]\[\-\.379,\-\.240\],OR=\.734\\mathrm\{OR\}=\.734,p=2\.8×10−18p=2\.8\\times 10^\{\-18\}\)\. Thus, instruction\-tuned models produced more total overlap under the actual prime while receiving a smaller incremental boost from replacing the unrelated prime with the true conversational prime\.

The three\-way tuning\-by\-prime\-type\-by\-frequency interaction was positive \(β=\.083\\beta=\.083,p=2\.8×10−6p=2\.8\\times 10^\{\-6\}\)\. Both tuning regimes retained a low\-frequency\-weighted actual\-prime effect, but that concentration was weaker under instruction tuning\. Pretrained and instruction\-tuned responses then appear to differ in*how much*syntax overlaps between the prime and its subsequent generated response, but also in how much of that overlap is specifically attributable to the true conversational prime\.

Table 1:Pooled instruction\-tuning effects under distinct estimands\. Conditional coefficients are evaluated at the target\-size centering mean; percentile\-specific contrasts appear in Appendix[C](https://arxiv.org/html/2607.26015#A3)\.
#### Conditioning on target structural opportunity\.

Additionally, target rule\-set size was a predictor of overlap\. In the full pooled sensitivity model, a one\-unit increase in centered log target size multiplied the odds of rule overlap by7\.297\.29for the pretrained reference condition \(β=1\.987\\beta=1\.987\)\. Once target size was conditioned upon, the instruction\-tuning coefficient reversed:β=−\.505\\beta=\-\.505, 95% CI\[−\.552,−\.458\]\[\-\.552,\-\.458\],OR=\.604\\mathrm\{OR\}=\.604,p=3\.0×10−98p=3\.0\\times 10^\{\-98\}\(Table[1](https://arxiv.org/html/2607.26015#S4.T1)\)\. The same direction held in the common\-support sample \(β=−\.423\\beta=\-\.423, 95% CI\[−\.477,−\.369\]\[\-\.477,\-\.369\], OR=\.655=\.655,p=1\.5×10−53p=1\.5\\times 10^\{\-53\}\), though retention in that sample ranged from 33\.2% to 95\.5% across pair\-by\-tuning cells, so it should not be read as representative of every natural instruction\-tuned response\. Moreover, instruction\-tuned variants showed lower conditional reuse at the 10th, 50th, and 90th percentiles of the supported target\-size distribution in both analyses \(Appendix[5](https://arxiv.org/html/2607.26015#A3.T5)\)\. The instruction disadvantage grew as target size increased, reflected in negative tuning\-by\-target\-size interactions in both the full \(β=−\.212\\beta=\-\.212\) and common\-support \(β=−\.196\\beta=\-\.196\) models\.

#### The human comparison under the same conditioning\.

The per\-model sensitivity fits \(Table[2](https://arxiv.org/html/2607.26015#A1.T2), columnβA−H\|t\\beta\_\{A\-H\|t\}\) apply the same conditioning to the model–human contrast, and they reorder it\. Unconditionally, pretrained models sat at or near the human baseline \(βA−H=−\.030\\beta\_\{A\-H\}=\-\.030to\.240\.240\) while every instruction\-tuned model exceeded it\. Holding target rule\-set size constant reverses this pattern, in that every pretrained model exceeded the matched human response byβA−H\|t=\.683\\beta\_\{A\-H\|t\}=\.683to\.922\.922\(allp<\.001p<\.001; OR=1\.98=1\.98to2\.512\.51\), whereas instruction\-tuned models exceeded it more modestly \(βA−H\|t=\.166\\beta\_\{A\-H\|t\}=\.166to\.525\.525in five architectures\) or not detectably at all \(Gemma\-3\-12B\-IT,\.069\.069; Gemma\-3\-27B\-IT,−\.081\-\.081; Gemma\-3\-4B\-IT,−\.023\-\.023\)\. The elevated natural\-output alignment of instruction\-tuned responses therefore rests substantially on the amount of eligible structure those responses contain, while the strongest conditional reuse propensity in the suite belongs to pretrained models\.

### 4\.4Primary result III: frequency profiles differ across humans, pretrained models, and instruction\-tuned models \(RQ3\)

Across all sixteen models, the actual\-versus\-random interaction with frequency was negative \(β=−\.504\\beta=\-\.504to−\.231\-\.231, allp≤9\.7×10−12p\\leq 9\.7\\times 10^\{\-12\}\): the additional reuse attributable to the true conversational prime rather than an unrelated one was larger for rarer rules\.

Relative to matched human responses, all eight pretrained model\-by\-frequency interactions were negative, seven significantly so \(β=−\.190\\beta=\-\.190to−\.052\-\.052\)\. The instruction\-tuned interactions were more varied: four were significantly positive, one was significantly negative, and three were nonsignificant \(β=−\.088\\beta=\-\.088to\.121\.121\)\.

In the pooled actual\-prime model, tuning interacted positively with frequency \(β=\.153\\beta=\.153,p=3\.2×10−64p=3\.2\\times 10^\{\-64\}\), indicating that the instruction\-tuning advantage in total actual\-prime overlap was larger for more frequent rules\. The positive tuning\-by\-prime\-type\-by\-frequency interaction \(β=\.083\\beta=\.083,p=2\.8×10−6p=2\.8\\times 10^\{\-6\}\) further indicated that the low\-frequency weighting of the actual\-versus\-random increment was weaker under instruction tuning\. Humans, pretrained models, and instruction\-tuned models therefore did not exhibit a single shared frequency profile\.

### 4\.5Exploratory finding: models show greater mean lexical and semantic similarity than matched human responses \(RQ4\)

All sixteen models were more lexically similar to the prime than the matched human response \(Appendix[6](https://arxiv.org/html/2607.26015#A4.T6)\)\. Human\-minus\-model TF–IDF differences ranged from−\.083\-\.083to−\.036\-\.036, and all sixteen tests remained significant after Holm correction\. Likewise, all sixteen models were more semantically similar to the prime, with differences from−\.104\-\.104to−\.018\-\.018and all Holm\-adjusted tests significant\. The instruction\-tuned variant was more semantically similar than its matched pretrained counterpart in all eight architecture pairs\. The lexical effect was not consistent, however; four pairs moved towards greater lexical similarity under tuning and four moved in the opposite direction\.

## 5Discussion

This study measured directional turn\-adjacent syntactic convergence in human\-LLM dialogue, analyzingBlevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\)’s substitution\-paradigm corpus and drawing upon a CFG\-based measurement framework; findings are discussed below\.

First, every instruction\-tuned model showed greater natural\-output syntactic alignment than the matched human response, and all eight matched architecture pairs showed greater total actual\-prime overlap under tuning\. A plausible interpretation is that instruction tuning optimizes models for local "responsiveness" – engaging directly with the immediately preceding turn through partial reuse of its phrasing and structure – which would be expected to raise alignment with the prime across numerous levels at once\. Additionally \(and beyond the syntactic domain\), humans have been found to evaluate more positively others who are perceived to behave similarly to themselves\(Montoyaet al\.,[2008](https://arxiv.org/html/2607.26015#bib.bib7)\), which raises the possibility that preference\-based post\-training could \(either directly or indirectly\) reward certain forms of linguistic similarity\. The present analysis, however, does not identify the training mechanism responsible for the observed pattern\.

Second \(and complicating a takeaway that "instruction\-tuned models converge more than pretrained ones"\), instruction\-tuned outputs overlap more withunrelatedprimes, their actual\-versus\-random increment is smaller, and their tuning advantage reverses after target structure is held constant\. In other words, instruction\-tuned models produce responses containing more eligible syntactic structure, which affords more opportunities for overlap, while pretrained models show a stronger tendency to reuse any given available rule\. This framing also reconciles the present results withBlevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\), who find that instruction\-tuned models generally converge less than pretrained variants on several stylometric features\. They suggest that reduced convergence may in part reflect a tendency to bring novel material into exchanges during instruction tuning; such a tendency could also extend to syntax, although this interpretation does remain speculative\.

Third, at the lexical and semantic levels, all sixteen models exceeded human similarity to the preceding turn\. In every pair, the instruction\-tuned model was more semantically similar than its pretrained counterpart, while the lexical direction split\. Notably, convergence in human dialogue is a graded and socially modulated signal\(Gileset al\.,[1991](https://arxiv.org/html/2607.26015#bib.bib12)\)\. In these models, similarity appears to be biased towards the immediately preceding turn, though future analyses could examine differential model convergence \(to various speakers and across registers\) in more depth\. To be fair, I would resist reading the similarity measures as convergence proper: prime\-to\-response similarity conflates accommodation with topical echoing, particularly under a 40\-token generation cap\.

Finally, for every model, the incremental CFG reuse was larger among the lower\-frequency rules\. Interestingly, human structural priming effects are also strongest for structures that are rare or dispreferred\(Scheepers,[2003](https://arxiv.org/html/2607.26015#bib.bib29); Bernolet and Hartsuiker,[2010](https://arxiv.org/html/2607.26015#bib.bib28); Jaeger and Snider,[2013](https://arxiv.org/html/2607.26015#bib.bib10)\)\. That this pattern appears in every model, including the smallest pretrained variants, suggests that sensitivity to the local syntactic environment is present before dialogue\-specific instruction tuning\. In humans, this frequency profile has been taken as evidence for error\-driven implicit learning\(Changet al\.,[2006](https://arxiv.org/html/2607.26015#bib.bib4); Jaeger and Snider,[2013](https://arxiv.org/html/2607.26015#bib.bib10)\); that account arguably cannot transfer directly to this paradigm, since no weights are updated during generation and any analogue must operate within the context window\. A prediction\-based account is nonetheless available in principle, in that predicting text in context plausibly rewards representations in which recently encountered structures are more available, though the present data cannot adjudicate between mechanistic accounts\.

Two implications follow from the analyses\. First, with respect to understanding LLM psycholinguistics, these findings supply a measure of how closely a generated turn tracks the human turn it answers, while suggesting that alignment measured at one level \(e\.g\., the syntactic\) does not straightforwardly license inference about another \(e\.g\., the lexical\)\.

Second, these findings contribute to emerging discussions of linguistic alignment in deployed systems and its potential risks\(Boyd and Markowitz,[2026](https://arxiv.org/html/2607.26015#bib.bib11)\)\.Augustinet al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib32)\)propose an "amplification spiral" in which linguistic alignment combines with hyperpersonalized generation and sycophancy to co\-construct maladaptive belief; they identify the degree of such alignment as an open empirical question\. The present paradigm operationalizes one local linguistic component that may be relevant to broader accounts of this phenomenon\.

## 6Scope, Limitations, and Extensions

Certain limitations qualify the present study\. First, all human\-human results are derived from DailyDialog, a corpus of short dialogues crawled from English\-learning websites\. Obtaining data and conducting analyses on other registers would be instructive to determine if similar patterns generalize\. The 40\-token generation cap together with the brevity of DailyDialog turns bounds the amount of structure available for reuse\. Human responses are not subject to the same generation cap, making response structure an especially important part of the estimand\. Relatedly, the present study is restricted to English; future researchers are encouraged to apply the substitution paradigm to multilingual corpora to address whether these convergence profiles are stable across typologically diverse languages\.

Furthermore, regarding timescale dynamics: the present analysis measures turn\-by\-turn CFG rule reuse between adjacent turns, which is most directly interpretable as short\-term structural priming\. Temporal analyses could test whether CFG\-based syntactic similarity accumulates over the course of a conversation beyond the local effects captured here\. Lastly, intermediate checkpoints \(pretraining, supervised fine\-tuning, and preference optimization\) could localize where in the tuning pipeline the observed changes in total alignment, response structure, and conditional reuse arise\.

## 7Conclusion

This study combinedBlevinset al\.\([2026](https://arxiv.org/html/2607.26015#bib.bib15)\)’s substitution\-paradigm data from sixteen open\-weight models with a CFG\-based measurement pipeline\. Every model showed greater turn\-adjacent CFG\-rule overlap with the true preceding human turn than with a sampled unrelated human turn\. Instruction\-tuned variants also produced greater total actual\-prime overlap than paired pretrained variants, but exhibited lower rule\-reuse odds than pretrained outputs when target structural opportunity was held constant\. Against the matched human baseline, all sixteen exceeded human lexical and semantic similarity to the preceding turn, and every instruction\-tuned model additionally exceeded human syntactic alignment in its natural output\.

Whether the patterns observed here reflect the tuning distribution, preference optimization, or a broader property of next\-token prediction under dialogue formatting is an empirical inquiry left to future work\. More generally speaking, these results raise questions related to human\-centered computing that extend beyond the present study; it, in any event, remains to be seen whether and how LLM syntactic convergence affects task efficacy and user wellbeing\.

## References

- M\. Augustin, T\. A\. Pollak, and H\. Morrin \(2026\)Characterizing the spiral: potential mechanisms in ai\-associated delusions\.NPP—Digital Psychiatry and Neuroscience4\(1\),pp\. 14\.Cited by:[§2](https://arxiv.org/html/2607.26015#S2.p2.1),[§5](https://arxiv.org/html/2607.26015#S5.p7.1)\.
- D\. Bates, M\. Mächler, B\. Bolker, and S\. Walker \(2015\)Fitting linear mixed\-effects models using lme4\.Journal of Statistical Software67\(1\),pp\. 1–48\.External Links:[Document](https://dx.doi.org/10.18637/jss.v067.i01)Cited by:[§3\.4](https://arxiv.org/html/2607.26015#S3.SS4.p14.1)\.
- S\. Bernolet and R\. J\. Hartsuiker \(2010\)Does verb bias modulate syntactic priming?\.Cognition114\(3\),pp\. 455–461\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p5.1)\.
- T\. Blevins, S\. Schmalwieser, and B\. Roth \(2026\)Do language models accommodate their users? a study of linguistic convergence\.InProceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\),Rabat, Morocco,pp\. 791–807\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.eacl-long.34),[Link](https://aclanthology.org/2026.eacl-long.34/)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1),[§1](https://arxiv.org/html/2607.26015#S1.p3.1),[§2](https://arxiv.org/html/2607.26015#S2.p2.1),[§3\.1](https://arxiv.org/html/2607.26015#S3.SS1.p1.1),[§5](https://arxiv.org/html/2607.26015#S5.p1.1),[§5](https://arxiv.org/html/2607.26015#S5.p3.1),[§7](https://arxiv.org/html/2607.26015#S7.p1.1)\.
- T\. Blevins \(2026\)Accommodation goes both ways: studying linguistic convergence between humans and language models\.arXiv preprint arXiv:2605\.29278\.External Links:[Document](https://dx.doi.org/10.48550/arXiv.2605.29278)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.
- J\. K\. Bock \(1986\)Syntactic persistence in language production\.Cognitive psychology18\(3\),pp\. 355–387\.Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p1.1),[§2](https://arxiv.org/html/2607.26015#S2.p1.1)\.
- R\. L\. Boyd and D\. M\. Markowitz \(2026\)Artificial intelligence and the psychology of human connection\.Perspectives on Psychological Science21\(2\),pp\. 192–220\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p7.1)\.
- H\. P\. Branigan, M\. J\. Pickering, and A\. A\. Cleland \(2000\)Syntactic co\-ordination in dialogue\.Cognition75\(2\),pp\. B13–B25\.External Links:[Document](https://dx.doi.org/10.1016/S0010-0277%2899%2900081-5)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p1.1),[§2](https://arxiv.org/html/2607.26015#S2.p1.1)\.
- Z\. Cai, X\. Duan, D\. Haslett, S\. Wang, and M\. Pickering \(2024\)Do large language models resemble humans in language use?\.InProceedings of the workshop on cognitive modeling and computational linguistics,pp\. 37–56\.Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.
- F\. Chang, G\. S\. Dell, and K\. Bock \(2006\)Becoming syntactic\.\.Psychological review113\(2\),pp\. 234\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p5.1)\.
- P\. Chen, H\. Guan, and E\. J\. Jeong \(2026\)Who accommodates whom? bidirectional linguistic accommodation and progressive interpersonal convergence in human–ai conversations\.Behavioral Sciences16\(5\),pp\. 720\.External Links:[Document](https://dx.doi.org/10.3390/bs16050720)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.
- Gemma Team, A\. Kamath, J\. Ferret, S\. Pathak, N\. Vieillard, R\. Merhej, S\. Perrin, T\. Matejovicova, A\. Ramé, M\. Rivière,et al\.\(2025\)Gemma 3 technical report\.External Links:2503\.19786,[Link](https://arxiv.org/abs/2503.19786)Cited by:[§3\.1](https://arxiv.org/html/2607.26015#S3.SS1.p2.1)\.
- H\. Giles, N\. Coupland, and J\. Coupland \(1991\)Accommodation theory: communication, context, and consequence\.InContexts of Accommodation: Developments in Applied Sociolinguistics,H\. Giles, J\. Coupland, and N\. Coupland \(Eds\.\),pp\. 1–68\.External Links:[Document](https://dx.doi.org/10.1017/CBO9780511663673.001)Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p4.1)\.
- A\. Grattafiori, A\. Dubey, A\. Jauhri, A\. Pandey, A\. Kadian, A\. Al\-Dahle, A\. Letman, A\. Mathur, A\. Schelten, A\. Vaughan,et al\.\(2024\)The llama 3 herd of models\.arXiv preprint arXiv:2407\.21783\.Cited by:[§3\.1](https://arxiv.org/html/2607.26015#S3.SS1.p2.1)\.
- S\. Holm \(1979\)A simple sequentially rejective multiple test procedure\.Scandinavian Journal of Statistics6\(2\),pp\. 65–70\.Cited by:[§3\.5](https://arxiv.org/html/2607.26015#S3.SS5.p1.1)\.
- T\. F\. Jaeger and N\. E\. Snider \(2013\)Alignment as a consequence of expectation adaptation: syntactic priming is affected by the prime’s prediction error given both prior and recent experience\.Cognition127\(1\),pp\. 57–83\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p5.1)\.
- J\. Jumelet, W\. Zuidema, and A\. Sinclair \(2024\)Do language models exhibit human\-like structural priming effects?\.InFindings of the Association for Computational Linguistics: ACL 2024,Bangkok, Thailand,pp\. 14727–14742\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.findings-acl.877),[Link](https://aclanthology.org/2024.findings-acl.877/)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1),[§2](https://arxiv.org/html/2607.26015#S2.p2.1)\.
- F\. Kandra, V\. Demberg, and A\. Koller \(2025\)LLMs syntactically adapt their language use to their conversational partner\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 2: Short Papers\),Vienna, Austria,pp\. 873–886\.External Links:[Document](https://dx.doi.org/10.18653/v1/2025.acl-short.68),[Link](https://aclanthology.org/2025.acl-short.68/)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1),[§2](https://arxiv.org/html/2607.26015#S2.p2.1)\.
- N\. Kitaev, S\. Cao, and D\. Klein \(2019\)Multilingual constituency parsing with self\-attention and pre\-training\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,Florence, Italy,pp\. 3499–3505\.External Links:[Document](https://dx.doi.org/10.18653/v1/P19-1340)Cited by:[§3\.2](https://arxiv.org/html/2607.26015#S3.SS2.p1.1)\.
- N\. Kitaev and D\. Klein \(2018\)Constituency parsing with a self\-attentive encoder\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),Melbourne, Australia,pp\. 2676–2686\.External Links:[Document](https://dx.doi.org/10.18653/v1/P18-1249)Cited by:[§3\.2](https://arxiv.org/html/2607.26015#S3.SS2.p1.1)\.
- Y\. Li, H\. Su, X\. Shen, W\. Li, Z\. Cao, and S\. Niu \(2017\)DailyDialog: a manually labelled multi\-turn dialogue dataset\.InProceedings of the Eighth International Joint Conference on Natural Language Processing \(Volume 1: Long Papers\),Taipei, Taiwan,pp\. 986–995\.External Links:[Link](https://aclanthology.org/I17-1099/)Cited by:[§3\.1](https://arxiv.org/html/2607.26015#S3.SS1.p1.1)\.
- E\. Mayor, L\. M\. Bietti, and A\. Bangerter \(2025\)Can large language models simulate spoken human conversations?\.Cognitive Science49\(9\),pp\. e70106\.External Links:[Link](https://onlinelibrary.wiley.com/doi/abs/10.1111/cogs.70106),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1111/cogs\.70106Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.
- J\. A\. Michaelov, C\. Arnett, T\. A\. Chang, and B\. K\. Bergen \(2023\)Structural priming demonstrates abstract grammatical representations in multilingual language models\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 3703–3720\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.227/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.227)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.
- I\. Montani, M\. Honnibal, A\. Boyd, S\. Van Landeghem, and H\. Peters \(2023\)Explosion/spacy: v3\. 7\.2: fixes for apis and requirements\.Zenodo\.Cited by:[§3\.2](https://arxiv.org/html/2607.26015#S3.SS2.p1.1)\.
- R\. M\. Montoya, R\. S\. Horton, and J\. Kirchner \(2008\)Is actual similarity necessary for attraction? a meta\-analysis of actual and perceived similarity\.Journal of social and personal relationships25\(6\),pp\. 889–922\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p2.1)\.
- M\. J\. Pickering and S\. Garrod \(2004\)Toward a mechanistic psychology of dialogue\.Behavioral and Brain Sciences27\(2\),pp\. 169–226\.External Links:[Document](https://dx.doi.org/10.1017/S0140525X04000056)Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p1.1)\.
- N\. Reimers and I\. Gurevych \(2019\)Sentence\-BERT: sentence embeddings using siamese BERT\-networks\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing,Hong Kong, China,pp\. 3982–3992\.External Links:[Document](https://dx.doi.org/10.18653/v1/D19-1410)Cited by:[§3\.5](https://arxiv.org/html/2607.26015#S3.SS5.p1.1)\.
- D\. Reitter and J\. D\. Moore \(2014\)Alignment and task success in spoken dialogue\.Journal of Memory and Language76,pp\. 29–46\.External Links:[Document](https://dx.doi.org/10.1016/j.jml.2014.05.008)Cited by:[§2](https://arxiv.org/html/2607.26015#S2.p1.1)\.
- C\. Scheepers \(2003\)Syntactic priming of relative clause attachments: persistence of structural configuration in sentence production\.Cognition89\(3\),pp\. 179–205\.Cited by:[§5](https://arxiv.org/html/2607.26015#S5.p5.1)\.
- A\. Sinclair, J\. Jumelet, W\. Zuidema, and R\. Fernández \(2022\)Structural persistence in language models: priming as a window into abstract language representations\.Transactions of the Association for Computational Linguistics10,pp\. 1031–1050\.Cited by:[§1](https://arxiv.org/html/2607.26015#S1.p2.1)\.

## Appendix APer\-model syntactic estimates

Table[2](https://arxiv.org/html/2607.26015#A1.T2)reports the key per\-model coefficients\. TheβA−H\|t\\beta\_\{A\-H\|t\}column is a per\-model sensitivity coefficient and should not be interpreted as the pooled tuning contrast reported in Table[1](https://arxiv.org/html/2607.26015#S4.T1)\.

Table 2:Per\-model GLMM coefficients\.βA−H\\beta\_\{A\-H\}is the primary model\-versus\-human contrast andβA−R\\beta\_\{A\-R\}the actual\-versus\-random contrast, both at centered mean rule frequency and prime size\.β\(A−H\)×f\\beta\_\{\(A\-H\)\\times f\}is the model–human frequency interaction andβ\(A−R\)×f\\beta\_\{\(A\-R\)\\times f\}the model–random frequency interaction; the two differ because the former varies the respondent at a fixed prime while the latter varies the prime at a fixed target\. Negativeβ\(A−R\)×f\\beta\_\{\(A\-R\)\\times f\}indicates that the actual\-prime increment is larger for lower\-frequency rules\.βA−H\|t\\beta\_\{A\-H\|t\}is the model–human contrast from the sensitivity fit that additionally conditions on centered target rule\-set sizett\(distinct from prime sizeqiκq\_\{i\}^\{\\kappa\}in Equation[2](https://arxiv.org/html/2607.26015#S3.E2)\)\. Stars denote two\-sided Wald tests\.p∗<\.05\{\}^\{\*\}p<\.05,p∗∗<\.01\{\}^\{\*\*\}p<\.01,p∗⁣∗∗<\.001\{\}^\{\*\*\*\}p<\.001\.

## Appendix BMatched\-pair raw rates

Table 3:Raw actual\-prime reuse rates for the eight matched architecture pairs\. All differences favor instruction tuning; exact two\-sided sign\-testp=\.0078p=\.0078\. Differences are computed from unrounded rates and may differ from the displayed subtraction by \.001\.
## Appendix CTarget\-size contrasts and common support

Table 4:Instruction\-versus\-pretraining contrasts at supported target sizes\. Negative values indicate lower conditional reuse under instruction tuning\. The median contrasts differ slightly from Table[1](https://arxiv.org/html/2607.26015#S4.T1)because the empirical median is close to, but not exactly, the centering mean\.Table 5:Position retention in the central common\-support analysis\. The 22,049 retained positions are 72\.5% of 30,416 position\-by\-model observations; they contribute 149,101 rule rows, 70\.2% of the full 212,544 rows, because retained positions contain slightly fewer eligible prime rules on average\.
## Appendix DLexical and semantic similarity

![Refer to caption](https://arxiv.org/html/2607.26015v1/x2.png)Figure 2:Human\-minus\-model similarity differences\. Negative values mean that the model response is closer to the prime\. Error bars are 95% conversation\-clustered intervals\.Table 6:Human\-minus\-model similarity differences and Holm\-adjustedpp\-values\.
## Appendix EDiagnostics

Table 7:Diagnostics for the pooled and target\-size models\. All fits were nonsingular and produced no convergence messages\.The analysis used a fixed random\-prime seed of 42\. Preliminary exploratory fits usednAGQ=0\\mathrm\{nAGQ\}=0; all coefficients reported in this manuscript were refit withnAGQ=1\\mathrm\{nAGQ\}=1\.