Too Good to Be Real? Diagnosing and Reducing the Gap Between AI Preference and Real User Engagement

# Too Good to Be Real? Diagnosing and Reducing the Gap Between AI Preference and Real User Engagement
Source: [https://arxiv.org/html/2609.18282](https://arxiv.org/html/2609.18282)
Yuanmeng XiangYunyao ZhangZeliang ChenJunqing YuZikai Song††thanks:Corresponding author\.Affiliation:Huazhong University of Science and TechnologyAffiliation:\{normanspark, skyesong\}@hust\.edu\.cn

###### Abstract

Large language models are increasingly used to generate and evaluate online content, yet it remains unclear whether the qualities they associate with higher engagement match what real users respond to\. We study this question using 1\.17 million answers to 25,978 questions from Zhihu, Quora, and Reddit, comparing real platform answers and AI\-generated answers across four within\-question engagement levels\. We introduceOntological Preference Measurement, which represents answers along three dimensions: logic, affect, and expression\. We find a systematic gap betweenAI preferenceandreal user engagement: as target engagement increases, LLMs add more explicit logical structure, while real user engagement is more strongly associated with affective and expressive salience\. We call this tendencylogic overbinding\. Based on this diagnosis, we proposeOntology\-Masked Reasoning Autoencoding \(OMRA\), a controlled intervention that masks and reconstructs over\-explained spans while preserving stance, factual content, and coherence\. Across four LLM families, OMRA reduces the measured gap by an average of54\.4%\. In human evaluation, OMRA wins62\.4%of pairwise preference judgments against matched real platform answers, even though the real answers are more often judged to be human\-written\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2609.18282v1/Figure1_teaser.png)Figure 1:Illustrative example of the AI preference–user engagement gap\.The AI\-preferred answer is complete and explicitly reasoned but receives low engagement\. Answers with stronger affective or expressive salience receive higher engagement in this example\.Large language models are increasingly used to generate, rank, and evaluate online content, and can be prompted to write answers targeting different levels of user engagement[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib16);[Lambert et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib17)\. However, it remains unclear whether the changes they make reflect what real users actually respond to\. When targeting higher engagement, LLMs tend to make their answers more complete, more structured, and more explicitly reasoned[Rodrigues et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib30)\. Real high\-engagement answers, however, may follow different patterns\. This raises two questions:*What do LLMs change when asked to write an answer for higher user engagement?**Do these changes match the patterns found in real high\-engagement answers?*

To answer these questions, we compare real platform answers and AI\-generated answers across the same four engagement levels\. We collect 1\.17 million answers to 25,978 questions from Zhihu, Quora, and Reddit\. Because raw vote counts are affected by exposure, platform traffic, and time, we rank answers only against other answers to the same question and map them to fourwithin\-question engagement levels\. We then ask LLMs to generate an answer for eachtarget engagement level\. This setup allows us to compare how real answers vary withreal user engagementand how LLMs modify their answers when targeting the same levels\. We refer to the latter pattern asAI preference: the qualities that LLMs associate with higher engagement\.

To compare these patterns, we introduceOntological Preference Measurement, which represents each answer along three dimensions:logic,affect, andexpression\. Logic captures explicit claims and support; affect captures emotional and value\-oriented appeal; and expression captures rhetorical salience and memorability\. As target engagement increases, LLMs add more claims, evidence, causal explanations, and step\-by\-step justification, whereas real user engagement is more strongly associated with affective and expressive salience and only weakly negatively associated with explicit reasoning\. We call this tendencylogic overbinding: AI preference places disproportionate weight on explicit logical presentation relative to real user engagement\. This pattern is consistent across transfer, marginal, and geometric comparisons\.

Based on this diagnosis, we proposeOntology\-Masked Reasoning Autoencoding \(OMRA\), a controlled intervention that masks and reconstructs over\-explained spans while preserving stance, factual content, and coherence\. Across four LLM families, OMRA reduces the measured AI preference–user engagement gap by an average of54\.4%, outperforming few\-shot imitation, reasoning\-augmented generation, and alternative masking strategies\. Additional controls show that length contributes to the gap but does not explain OMRA’s improvement, while sentence\-mood changes have only minor effects; preservation checks indicate that stance, factual consistency, relevance, and usefulness are largely retained\. In human evaluation, OMRA wins62\.4%of pairwise preference judgments against matched real platform answers, even though the real answers are more often judged to be human\-written\.

In summary, our contributions are:

- •We introduceOntological Preference Measurement, a structured framework for comparing AI preference and real user engagement across logic, affect, and expression, and identifylogic overbindingusing 1\.17 million answers from three platforms\.
- •We proposeOntology\-Masked Reasoning Autoencoding \(OMRA\), which reconstructs over\-explained spans and reduces the measured gap by 54\.4% across four LLM families, with supporting evidence from control experiments and human evaluation\.

![Refer to caption](https://arxiv.org/html/2609.18282v1/Figure2_pipeline.png)Figure 2:Overview of our ontology\-guided diagnosis and correction framework\.Left: Ontological Preference Measurement\.Answers are represented along logic, affect, and expression in a shared interpretable space\.Middle: AI preference–user engagement diagnosis\.Transfer, marginal, and geometric comparisons reveal logic overbinding in AI preference\.Right: Ontology\-Masked Reasoning Autoencoding \(OMRA\)\.OMRA identifies, masks, and reconstructs over\-explained spans, moving AI\-generated answers closer to the real\-answer distribution\.
## 2Related Work

A full discussion appears in Appendix[A](https://arxiv.org/html/2609.18282#A1)and we only highlight key directions here\.

##### Ontology, discourse, and argument mining\.

Ontology\-based modeling represents text through interpretable concepts and relations[Gruber \(1993\)](https://arxiv.org/html/2609.18282#bib.bib1);[Noy et al\. \(2001\)](https://arxiv.org/html/2609.18282#bib.bib2), while argument mining identifies claims, evidence, and argumentative relations[Lippi and Torroni \(2016\)](https://arxiv.org/html/2609.18282#bib.bib3);[Lawrence and Reed \(2019\)](https://arxiv.org/html/2609.18282#bib.bib4);[Chakrabarty et al\. \(2019\)](https://arxiv.org/html/2609.18282#bib.bib26)\. Related work also studies how discourse structure, affect, framing, and rhetoric shape persuasion and user response[Feng and Hirst \(2011\)](https://arxiv.org/html/2609.18282#bib.bib33);[Tan et al\. \(2016\)](https://arxiv.org/html/2609.18282#bib.bib34);[Stab and Gurevych \(2017\)](https://arxiv.org/html/2609.18282#bib.bib5)\. We build on these directions by organizing online answers along logic, affect, and expression to compare AI preference with real user engagement in a shared interpretable space\.

##### AI preference and LLM\-as\-a\-judge\.

AI synthetic data and LLM\-as\-a\-judge evaluation are widely used as scalable substitutes for human annotation in reward modeling and alignment[Christiano et al\. \(2017\)](https://arxiv.org/html/2609.18282#bib.bib15);[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib16);[Li et al\. \(2024\)](https://arxiv.org/html/2609.18282#bib.bib14);[Lambert et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib17)\. Prior work documents biases involving verbosity, length, position, and self\-preference[Hu et al\. \(2024\)](https://arxiv.org/html/2609.18282#bib.bib18);[Ye et al\. \(2025a\)](https://arxiv.org/html/2609.18282#bib.bib19);[Saito et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib20);[Wang et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib21), but these biases are typically evaluated against benchmark labels or other models\. We instead compare the qualities associated with AI preference against naturally occurring real user engagement\.

##### Engagement modeling and preference\-controlled generation\.

Prior work predicts online engagement using user, temporal, network, and content features[Deng et al\. \(2020\)](https://arxiv.org/html/2609.18282#bib.bib12);[Rameez et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib13);[Lei et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib6);[Wang et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib8), while preference\-aligned and controllable generation steer outputs toward desired rewards or attributes[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib16);[Rafailov et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib25);[Zhang et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib22);[Song et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib44)\. These approaches do not directly diagnose which answer qualities AI preference associates with engagement or how those associations differ from real user engagement\. OMRA addresses this gap by using the ontology to identify and reconstruct over\-explained spans under stance\-, factual\-consistency\-, and coherence\-preserving constraints\.

## 3Methodology

We organize our methodology into three stages illustrated in Figure[2](https://arxiv.org/html/2609.18282#S1.F2): ontological preference measurement \(§[3\.1](https://arxiv.org/html/2609.18282#S3.SS1)\), AI preference–user engagement gap diagnosis \(§[3\.2](https://arxiv.org/html/2609.18282#S3.SS2)\), and OMRA correction \(§[3\.3](https://arxiv.org/html/2609.18282#S3.SS3)\)\.

### 3\.1Ontological Preference Measurement

To study real user engagement beyond raw vote counts, we build an ontology\-based measurement framework that represents each answer through three complementary dimensions—logic, affect, and expression\. This framework connects observable platform engagement signals with interpretable textual properties rather than treating engagement as a black\-box scalar\.

![Refer to caption](https://arxiv.org/html/2609.18282v1/Figure3_dataset.png)Figure 3:Overview of the platform engagement dataset\.The left panel summarizes corpus\-level statistics, including the number of questions, answers, domains, sub\-domains, and engagement\-related attributes\. The middle sunburst chart shows the hierarchical topic distribution across primary domains and fine\-grained sub\-domains\. The right panel visualizes the distributions of platform engagement signals and textual properties, including vote count, question title length, question view count, and answer length\.#### Real User Engagement Dataset

We collect a large\-scale dataset of question–answer pairs with platform engagement signals from three open online platforms:Zhihu,Quora, andReddit\. The corpus spans 2020–2026 and covers 10 categories and 35 subcategories, with each answer associated with naturally occurring engagement signals such as votes \(Figure[3](https://arxiv.org/html/2609.18282#S3.F3)\)\. Platform\-specific filtering, deduplication, and engagement\-quality criteria are detailed in Appendix[B](https://arxiv.org/html/2609.18282#A2)\.

Since absolute vote counts are affected by question exposure, platform traffic, and temporal dynamics, we treat votes as*platform engagement signals*rather than direct measurements of underlying human preference\. We therefore operate on*within\-question engagement*\. For each questionqqwith answer set𝒜q\\mathcal\{A\}\_\{q\}, we assign each answeraaawithin\-question engagement level

ℓ⁡\(a\)=ϕq​\(rankq⁡\(v⁡\(a\)\)\)∈\{0,1,2,3\},\\ell\(a\)=\\phi\_\{q\}\\\!\\left\(\\operatorname\{rank\}\_\{q\}\(v\(a\)\)\\right\)\\in\\\{0,1,2,3\\\},whereϕq\\phi\_\{q\}maps vote ranks withinqqto four ordinal levels, isolating relative user engagement under a shared question context\.

We compare four*answer regimes*:Real, consisting of observed platform answers labeled by within\-question engagement;Synthetic, consisting of AI\-generated answers conditioned on target engagement levels;Few\-shot, additionally conditioned on a real reference answer from the same level; andOMRA\-corrected, revised by our ontology\-guided masking and reconstruction method \(§[3\.3](https://arxiv.org/html/2609.18282#S3.SS3)\)\.

#### Ontology for Preference Decomposition

We decompose each answeraainto three preference layers,

𝒪⁡\(a\)=\{𝒪L​\(a\),𝒪A​\(a\),𝒪E​\(a\)\},\\mathcal\{O\}\(a\)=\\bigl\\\{\\mathcal\{O\}^\{L\}\(a\),\\ \\mathcal\{O\}^\{A\}\(a\),\\ \\mathcal\{O\}^\{E\}\(a\)\\bigr\\\},where𝒪L\\mathcal\{O\}^\{L\}\(logic\) captures how explicitly the answer states and supports its stance,𝒪A\\mathcal\{O\}^\{A\}\(affect\) captures emotional and value\-oriented appeal, and𝒪E\\mathcal\{O\}^\{E\}\(expression\) captures readability, memorability, and rhetorical salience\. Each layer is organized around*concepts*,*relations*, and*patterns*, allowing us to describe how answer qualities vary across engagement levels\. Full ontology definitions, subdimensions, extraction templates, and examples are provided in Appendix[C](https://arxiv.org/html/2609.18282#A3)\.

#### Ontology Extraction and Preference Scoring

Given an answer, we use a fixed LLM\-based extractor with a constrained ontology schema and output format to obtain𝒪⁡\(a\)\\mathcal\{O\}\(a\), and then convert it into a layer\-wise preference vector

𝐬⁡\(a\)=\[𝐬L​\(a\);𝐬A​\(a\);𝐬E​\(a\)\],\\mathbf\{s\}\(a\)=\\bigl\[\\mathbf\{s\}^\{L\}\(a\);\\ \\mathbf\{s\}^\{A\}\(a\);\\ \\mathbf\{s\}^\{E\}\(a\)\\bigr\],where each𝐬∗​\(a\)\\mathbf\{s\}^\{\\ast\}\(a\)aggregates the scores of its subdimensions\. The original answer text remains the primary evidence for scoring, while the extracted ontology provides structured cues\. The resulting𝐬⁡\(a\)\\mathbf\{s\}\(a\)serves as an interpretable measurement of answer qualities rather than a direct engagement score, and grounds the diagnostics and corrections in the rest of the paper\. Schemas and prompts are provided in Appendix[G](https://arxiv.org/html/2609.18282#A7)\.

### 3\.2AI Preference–User Engagement Gap Diagnosis

Using the ontological preference vector𝐬⁡\(a\)\\mathbf\{s\}\(a\), we diagnose the AI preference–user engagement gap\. We organize transfer, marginal, and geometric gaps as a diagnostic sequence: from cross\-regime transfer failure, to the association pattern behind it, and finally to its distributional footprints\. Letr∈\{R,S,F\}r\\in\\\{R,S,F\\\}denote an answer regime \(Real, Synthetic, or Few\-shot\) and𝒟r\\mathcal\{D\}\_\{r\}its answer set\.

##### Transfer Gap

We first test whether an engagement rule learned in one regime generalizes to another\. We train a predictorfrf\_\{r\}that maps𝐬⁡\(a\)\\mathbf\{s\}\(a\)to within\-question engagement level and evaluate it on regimer′r^\{\\prime\}:

Tr→r′=ℳ⁡\(fr,𝒟r′\),r,r′∈\{R,S,F\},T\_\{r\\to r^\{\\prime\}\}=\\mathcal\{M\}\\\!\\left\(f\_\{r\},\\mathcal\{D\}\_\{r^\{\\prime\}\}\\right\),\\qquad r,r^\{\\prime\}\\in\\\{R,S,F\\\},whereℳ\\mathcal\{M\}includes Spearmanρ\\rhoand Top\-1 accuracy\. The asymmetry betweenTR→ST\_\{R\\to S\}andTS→RT\_\{S\\to R\}distinguishes internal consistency from transfer: an AI preference rule may be internally coherent yet fail on real platform answers\.

##### Logic Overbinding

We definelogic overbindingas the tendency of AI preference to place disproportionate weight on explicit logical presentation relative to real user engagement\. For eachd∈\{L,A,E\}d\\in\\\{L,A,E\\\}, we compute the within\-regime association

ρdr=Spearman⁡\(\{sd​\(a\)\}a∈𝒟r,\{ℓ⁡\(a\)\}a∈𝒟r\),\\rho\_\{d\}^\{r\}=\\operatorname\{Spearman\}\\bigl\(\\\{s\_\{d\}\(a\)\\\}\_\{a\\in\\mathcal\{D\}\_\{r\}\},\\\{\\ell\(a\)\\\}\_\{a\\in\\mathcal\{D\}\_\{r\}\}\\bigr\),whereLL,AA, andEEdenote logic, affect, and expression\. ComparingρdR\\rho\_\{d\}^\{R\}andρdS\\rho\_\{d\}^\{S\}tests whether AI preference and real user engagement track the same answer qualities, or whether AI preference more strongly favors explicit claim–evidence bridges, causal justifications, and enumerated explanations\.

##### Marginal Gap

We compare marginal score distributions\. Ford∈\{L,A,E\}d\\in\\\{L,A,E\\\}, we compute 1\-Wasserstein distance and Jensen–Shannon divergence:

Wdr,r′\\displaystyle W\_\{d\}^\{r,r^\{\\prime\}\}=W1​\(𝒫dr,𝒫dr′\),\\displaystyle=W\_\{1\}\\bigl\(\\mathcal\{P\}\_\{d\}^\{r\},\\mathcal\{P\}\_\{d\}^\{r^\{\\prime\}\}\\bigr\),Jdr,r′\\displaystyle J\_\{d\}^\{r,r^\{\\prime\}\}=JS⁡\(𝒫dr,𝒫dr′\),\\displaystyle=\\mathrm\{JS\}\\bigl\(\\mathcal\{P\}\_\{d\}^\{r\},\\mathcal\{P\}\_\{d\}^\{r^\{\\prime\}\}\\bigr\),where𝒫dr\\mathcal\{P\}\_\{d\}^\{r\}is the empirical distribution of\{sd​\(a\)\}a∈𝒟r\\\{s\_\{d\}\(a\)\\\}\_\{a\\in\\mathcal\{D\}\_\{r\}\}\. Whileρdr\\rho\_\{d\}^\{r\}measures association with engagement,Wdr,r′W\_\{d\}^\{r,r^\{\\prime\}\}andJdr,r′J\_\{d\}^\{r,r^\{\\prime\}\}measure distributional differences across answer regimes\.

##### Geometric Gap

Marginal similarity does not imply alignment in the joint preference space: two regimes may overlap on individual dimensions while combining them differently\. We therefore compare regimes in𝐬⁡\(a\)∈ℝ\|L\|\+\|A\|\+\|E\|\\mathbf\{s\}\(a\)\\in\\mathbb\{R\}^\{\|L\|\+\|A\|\+\|E\|\}using centroid distance and maximum mean discrepancy:

Cr,r′\\displaystyle C^\{r,r^\{\\prime\}\}=‖𝔼𝒟r​\[𝐬\]−𝔼𝒟r′​\[𝐬\]‖2,\\displaystyle=\\left\\\|\\mathbb\{E\}\_\{\\mathcal\{D\}\_\{r\}\}\[\\mathbf\{s\}\]\-\\mathbb\{E\}\_\{\\mathcal\{D\}\_\{r^\{\\prime\}\}\}\[\\mathbf\{s\}\]\\right\\\|\_\{2\},MMDr,r′\\displaystyle\\operatorname\{MMD\}^\{r,r^\{\\prime\}\}=MMD⁡\(𝒟r,𝒟r′\)\.\\displaystyle=\\operatorname\{MMD\}\\bigl\(\\mathcal\{D\}\_\{r\},\\mathcal\{D\}\_\{r^\{\\prime\}\}\\bigr\)\.Low\-dimensional projections show how answer regimes occupy the shared preference space\.

Together, these probes provide complementary views of the same AI preference–user engagement gap\. Section[5](https://arxiv.org/html/2609.18282#S5)shows that their results consistently point to logic overbinding and over\-explained spans, motivating a targeted intervention rather than separate optimization of each metric\.

### 3\.3Ontology\-Masked Reasoning Autoencoding

Based on this diagnosis, we proposeOntology\-Masked Reasoning Autoencoding \(OMRA\)as both a controlled intervention and a method for reducing the measured AI preference–user engagement gap\. OMRA targets over\-explained spans that make reasoning unnecessarily explicit, repetitive, or formulaic\.

Inspired by masked autoencoding[He et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib35), OMRA identifies and masks ontology\-derived over\-explained spans, then reconstructs the answer under stance\-preserving, factual\-consistency, coherence, and preference constraints\. Given a Synthetic draftaaand its ontology representation𝒪⁡\(a\)\\mathcal\{O\}\(a\)from §[3\.1](https://arxiv.org/html/2609.18282#S3.SS1.SSSx3), OMRA produces a revised answera′a^\{\\prime\}through the five stages in Algorithm[1](https://arxiv.org/html/2609.18282#alg1)\.

Algorithm 1The OMRA Algorithm1:Synthetic draft

aa; ontology extractor

ℰ\\mathcal\{E\}; reconstructor

ℛ\\mathcal\{R\}; reviser

𝒱\\mathcal\{V\}; preference constraints

cc; over\-explanation type set

𝒯over\\mathcal\{T\}\_\{\\mathrm\{over\}\}
2:// Stage 1: Ontology Extraction

3:

𝒪⁡\(a\)←ℰ⁡\(a\)\\mathcal\{O\}\(a\)\\leftarrow\\mathcal\{E\}\(a\), with

𝒪⁡\(a\)=\{𝒪L​\(a\),𝒪A​\(a\),𝒪E​\(a\)\}\\mathcal\{O\}\(a\)=\\\{\\mathcal\{O\}^\{L\}\(a\),\\mathcal\{O\}^\{A\}\(a\),\\mathcal\{O\}^\{E\}\(a\)\\\}
4:// Stage 2: Over\-Explained Span Identification

5:

𝒮\(a\)←\{span\(u\)∣u∈𝒪L\(a\)∪𝒪E\(a\),type\(u\)∈𝒯over\}\\mathcal\{S\}\(a\)\\leftarrow\\bigl\\\{\\mathrm\{span\}\(u\)\\mid u\\in\\mathcal\{O\}^\{L\}\(a\)\\cup\\mathcal\{O\}^\{E\}\(a\),\\mathrm\{type\}\(u\)\\in\\mathcal\{T\}\_\{\\mathrm\{over\}\}\\bigr\\\}
6:// Stage 3: Ontology\-Guided Masking

7:

a~←Mask⁡\(a,𝒮⁡\(a\)\)\\tilde\{a\}\\leftarrow\\mathrm\{Mask\}\(a,\\mathcal\{S\}\(a\)\)
8:// Stage 4: Reconstruction with Preference Constraints

9:

a^←ℛ⁡\(a~,𝒪⁡\(a\),c\)\\hat\{a\}\\leftarrow\\mathcal\{R\}\(\\tilde\{a\},\\mathcal\{O\}\(a\),c\)
10:// Stage 5: Preference\-Aware Revision

11:

a′←𝒱⁡\(a^,𝒪⁡\(a\),c\)a^\{\\prime\}\\leftarrow\\mathcal\{V\}\(\\hat\{a\},\\mathcal\{O\}\(a\),c\)
12:Output:revised answer

a′\{\\color\[rgb\]\{1,0\.5,0\.0547\}a^\{\\prime\}\}

Base ModelMethodMarginal Gap↓\\downarrowGeometry Gap↓\\downarrowTransfer to Real↑\\uparrowAvg\. Shrink↑\\uparrowWass\.JSMMDCent\.𝝆\\boldsymbol\{\\rho\}\(Δ\\Delta\)Top\-1![[Uncaptioned image]](https://arxiv.org/html/2609.18282v1/Figure/deepseek-color.png)DeepSeekDirect1\.4790\.1540\.2401\.969\-0\.0510\.261—Few\-shot0\.829\(43\.9%↓\\downarrow\)0\.085\(44\.8%↓\\downarrow\)0\.102\(57\.5%↓\\downarrow\)0\.971\(50\.7%↓\\downarrow\)0\.164\(\+0\.215\)0\.316\(21\.1%↑\\uparrow\)49\.2%ToT1\.245\(15\.9%↓\\downarrow\)0\.199\(↑\\uparrow29\.0%\)0\.152\(36\.8%↓\\downarrow\)1\.570\(20\.3%↓\\downarrow\)0\.136\(\+0\.187\)0\.313\(19\.9%↑\\uparrow\)11\.0%RoT1\.343\(9\.2%↓\\downarrow\)0\.210\(↑\\uparrow36\.4%\)0\.184\(23\.1%↓\\downarrow\)1\.800\(8\.6%↓\\downarrow\)\-0\.030\(\+0\.021\)0\.253\(↓\\downarrow3\.1%\)1\.1%Ours0\.452\(69\.4%↓\\downarrow\)0\.050\(67\.5%↓\\downarrow\)0\.038\(84\.2%↓\\downarrow\)0\.422\(78\.6%↓\\downarrow\)0\.225\(\+0\.276\)0\.358\(37\.2%↑\\uparrow\)74\.9%![[Uncaptioned image]](https://arxiv.org/html/2609.18282v1/Figure/claude-color.png)ClaudeDirect1\.2960\.1570\.2521\.782\-0\.0970\.239—Few\-shot0\.987\(23\.8%↓\\downarrow\)0\.135\(14\.0%↓\\downarrow\)0\.138\(45\.2%↓\\downarrow\)1\.197\(32\.8%↓\\downarrow\)0\.074\(\+0\.171\)0\.284\(18\.8%↑\\uparrow\)29\.0%ToT1\.427\(↑\\uparrow10\.1%\)0\.227\(↑\\uparrow44\.5%\)0\.176\(30\.1%↓\\downarrow\)1\.670\(6\.3%↓\\downarrow\)\-0\.047\(\+0\.050\)0\.269\(12\.6%↑\\uparrow\)\-4\.6%RoT1\.571\(↑\\uparrow21\.2%\)0\.237\(↑\\uparrow50\.8%\)0\.212\(16\.0%↓\\downarrow\)1\.854\(↑\\uparrow4\.0%\)\-0\.023\(\+0\.074\)0\.257\(7\.5%↑\\uparrow\)\-15\.0%Ours0\.807\(37\.7%↓\\downarrow\)0\.091\(42\.0%↓\\downarrow\)0\.108\(57\.1%↓\\downarrow\)0\.786\(55\.9%↓\\downarrow\)0\.193\(\+0\.290\)0\.323\(35\.1%↑\\uparrow\)48\.2%![[Uncaptioned image]](https://arxiv.org/html/2609.18282v1/Figure/openai.png)GPTDirect1\.7380\.1970\.2772\.338\-0\.1010\.253—Few\-shot1\.272\(26\.8%↓\\downarrow\)0\.162\(17\.8%↓\\downarrow\)0\.175\(36\.8%↓\\downarrow\)1\.698\(27\.4%↓\\downarrow\)0\.124\(\+0\.225\)0\.285\(12\.6%↑\\uparrow\)27\.2%ToT1\.493\(14\.1%↓\\downarrow\)0\.223\(↑\\uparrow13\.4%\)0\.210\(24\.3%↓\\downarrow\)1\.821\(22\.1%↓\\downarrow\)0\.088\(\+0\.189\)0\.320\(26\.5%↑\\uparrow\)11\.8%RoT1\.468\(15\.5%↓\\downarrow\)0\.172\(12\.6%↓\\downarrow\)0\.183\(34\.1%↓\\downarrow\)1\.607\(31\.3%↓\\downarrow\)\-0\.100\(\+0\.001\)0\.241\(↓\\downarrow4\.7%\)23\.4%Ours0\.887\(49\.0%↓\\downarrow\)0\.154\(21\.8%↓\\downarrow\)0\.126\(54\.5%↓\\downarrow\)0\.828\(64\.6%↓\\downarrow\)0\.190\(\+0\.291\)0\.327\(29\.2%↑\\uparrow\)47\.5%![[Uncaptioned image]](https://arxiv.org/html/2609.18282v1/Figure/meta-color.png)LlamaDirect1\.3950\.2430\.4342\.657\-0\.0430\.261—Few\-shot1\.066\(23\.6%↓\\downarrow\)0\.142\(41\.6%↓\\downarrow\)0\.197\(54\.6%↓\\downarrow\)1\.815\(31\.7%↓\\downarrow\)0\.124\(\+0\.167\)0\.303\(16\.1%↑\\uparrow\)37\.9%ToT1\.746\(↑\\uparrow25\.2%\)0\.289\(↑\\uparrow18\.7%\)0\.362\(16\.5%↓\\downarrow\)2\.809\(↑\\uparrow5\.7%\)0\.032\(\+0\.075\)0\.235\(↓\\downarrow10\.0%\)\-8\.3%RoT2\.138\(↑\\uparrow53\.2%\)0\.344\(↑\\uparrow41\.4%\)0\.570\(↑\\uparrow31\.3%\)3\.878\(↑\\uparrow45\.9%\)\-0\.088\(\-0\.045\)0\.207\(↓\\downarrow20\.7%\)\-43\.0%Ours0\.916\(34\.3%↓\\downarrow\)0\.133\(45\.3%↓\\downarrow\)0\.137\(68\.4%↓\\downarrow\)1\.588\(40\.2%↓\\downarrow\)0\.149\(\+0\.192\)0\.313\(19\.9%↑\\uparrow\)47\.1%Ours \(Avg\.\)0\.7660\.1070\.1020\.9060\.1890\.33054\.4%Table 1:Main ResultsEach cell is the distance between a method’s outputs and real platform answers\.Avg\. Shrinkis the mean relative reduction over the four distributional metrics against Direct\. Parenthesized values are relative changes over Direct \(forρ\\rho: absolute point gain\), withgreenmarking improvement andredmarking degradation\. The last row averages OMRA over the four base models\.##### Over\-Explained Span Identification

We define*over\-explained spans*as surface spans in the final answer that make its reasoning unnecessarily explicit, repetitive, or formulaic\. They include claim–evidence bridges, justification\-heavy sentences, causal explanations, enumerated steps, and explicit transitions\. These spans are not hidden reasoning traces\. OMRA reuses the ontology extractor rather than training a separate span classifier: each unitu∈𝒪L​\(a\)∪𝒪E​\(a\)u\\in\\mathcal\{O\}^\{L\}\(a\)\\cup\\mathcal\{O\}^\{E\}\(a\)is aligned with its source span, and units whose relation or pattern type belongs to𝒯over\\mathcal\{T\}\_\{\\mathrm\{over\}\}form𝒮⁡\(a\)\\mathcal\{S\}\(a\)\.

##### Ontology\-Guided Masking

OMRA replaces spans in𝒮⁡\(a\)\\mathcal\{S\}\(a\)with mask tokens while preserving the main stance and surrounding context\.

##### Reconstruction with Preference Constraints

The reconstructorℛ\\mathcal\{R\}fills the masks under four constraints: \(i\) stance preservation, \(ii\) factual consistency with non\-masked content, \(iii\) discourse coherence, and \(iv\) affective or expressive salience only when contextually supported\. The goal is not to recover the original spans, but to express the same answer with less over\-explanation\.

##### Preference\-Aware Revision

A verification pass𝒱\\mathcal\{V\}checks whethera^\\hat\{a\}still contains over\-explanation associated with logic overbinding\. If so,𝒱\\mathcal\{V\}applies a lightweight revision; otherwise,a′=a^a^\{\\prime\}=\\hat\{a\}\.

## 4Experiments

### 4\.1Experimental Setup

We evaluate on a controlled benchmark of 3,600 real platform answers sampled for matched within\-question comparison\. Each selected question provides one representative answer at each of the four within\-question engagement levels \(Appendix[B\.4](https://arxiv.org/html/2609.18282#A2.SS4)\)\. For each question–level pair, we construct three generated regimes:Direct \(Synthetic\), generated directly under the target engagement level;Few\-shot, additionally conditioned on a level\-matched real reference; andOMRA\-corrected, revised by Algorithm[1](https://arxiv.org/html/2609.18282#alg1)\. We instantiate the pipeline with four LLM families—DeepSeek, Claude, GPT, and Llama—and compare against two reasoning\-augmented baselines, Tree\-of\-Thought \(ToT\)[Yao et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib24)and Ripple\-of\-Thought \(RoT\)[Lei et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib6)\.

We use DeepSeek\-V3\.2 as the fixed ontology extractor across all four model families\. Since ontology scoring is automated, we verify its reliability against human annotators and alternative LLM judges on a randomly selected 100\-question subset, obtaining strong agreement on relative answer\-quality rankings \(Appendix[C\.6](https://arxiv.org/html/2609.18282#A3.SS6)\)\.

Figure 4:Diagnosing and reducing the AI preference–user engagement gap\.\(a\)Dimension\-wise Spearmanρ\\rhobetween ontology scores and within\-question engagement levels across answer regimes\.\(b\)Joint preference geometry in a diagnostic space of logic overbinding and affective–expressive salience\.\(c\)Ablation over masking and revision variants in OMRA on DeepSeek\.MethodAvg\.ρ\\rhoMed\.ρ\\rhoTop\-1PearsonAcc\.𝒓\\boldsymbol\{r\}Random baseline0\.0000\.0000\.2500\.000Text embedding0\.2140\.4000\.3240\.309LLM judgeDeepSeek\-0\.080\-0\.2000\.141\-0\.050ChatGPT\-0\.127\-0\.2580\.140\-0\.075Llama\-0\.0040\.2580\.082\-0\.014Claude0\.1820\.2580\.1640\.121Ontology features0\.3590\.4000\.4350\.465Table 2:Validation of ontological preference measurement on real\-user answers\. All ranking metrics are computed within questions\.
### 4\.2Evaluation Results

##### Ontology features capture signals associated with real user engagement better than dense embeddings and LLM judges\.

Before evaluating OMRA, we verify that the ontology vector𝐬⁡\(a\)\\mathbf\{s\}\(a\)carries engagement\-relevant signals in real platform answers\. Table[2](https://arxiv.org/html/2609.18282#S4.T2)compares a lightweight predictor over𝐬⁡\(a\)\\mathbf\{s\}\(a\)with dense text embeddings and four LLM\-as\-a\-judge baselines on within\-question ranking\. Ontology features outperform all baselines, reaching Avg\.ρ=0\.359\\rho=0\.359and Top\-1=0\.435=0\.435, while three of four LLM judges fall below the dense\-embedding baseline\. These results suggest that the ontology representation captures real user engagement signals that black\-box LLM judges do not reliably recover\.

##### OMRA reduces the AI preference–user engagement gap across all four LLM families\.

Table[1](https://arxiv.org/html/2609.18282#S3.T1)compares OMRA with Direct, Few\-shot, ToT, and RoT across marginal, geometric, and transfer metrics\. OMRA reduces the measured gap by an average of54\.4%and is the only method that improves all three views on every base model\. Few\-shot is the strongest baseline but remains weaker on geometry and transfer, suggesting that reference imitation does not fully recover the patterns associated with real user engagement\. ToT and RoT do not consistently reduce the gap across model families, suggesting that reasoning augmentation alone does not recover these patterns\. OMRA performs consistently across all four model families, indicating that its improvement is not specific to a single generator\.

##### Additional controls rule out simple alternative explanations\.

We conduct additional diagnostic experiments on DeepSeek \(Appendix[D\.5](https://arxiv.org/html/2609.18282#A4.SS5)\)\. Length control improves both Direct and Few\-shot generation, but does not account for OMRA’s gains\. Platform\-aware prompting provides modest improvements, while sentence\-mood constraints have only minor effects\. A separate preservation evaluation further shows that OMRA largely preserves stance, factual consistency, answer relevance, and usefulness\.

### 4\.3Ablation Study

##### Joint span identification is more effective than surface\-marker or single\-layer masking\.

We ablate OMRA with five variants in Figure[4](https://arxiv.org/html/2609.18282#S4.F4)\(c\):w/o Reconstructleaves masked spans empty;Direct Refinerewrites the answer without masking;Connectors Maskmasks only surface discourse markers; andClaim Mask/Evidence Maskmask only one component type\. Full OMRA performs best across all gap metrics\. Direct Refine and Connectors Mask yield the weakest improvements, showing that generic rewriting or surface\-marker deletion is insufficient\. Claim Mask and Evidence Mask achieve intermediate results, suggesting that over\-explanation is better captured through joint spans involving claims, evidence, and their relations\.

Training→\\toTest RegimeSpearman𝝆\\boldsymbol\{\\rho\}Top\-1Synthetic→\\toSynthetic0\.9470\.895Real→\\toSynthetic0\.7590\.800Synthetic→\\toReal−\-0\.0730\.254Table 3:Cross\-regime transfer of ontological preference predictors\.A predictor is trained on the source regime and evaluated on the target regime; metrics are averaged over four base LLM families\.

## 5Analysis

### 5\.1AI Preference–User Engagement Gap

We analyze the AI preference–user engagement gap through the diagnostic sequence introduced in §[3\.2](https://arxiv.org/html/2609.18282#S3.SS2): cross\-regime transfer, dimension\-wise associations, and joint preference geometry\.

Pairing𝑵\\boldsymbol\{N\}Human\-like Win RatePreferred Win Rate𝜿\\boldsymbol\{\\kappa\}Reversal RateDirect vs Real1342\.7% \[1\.2, 7\.4\]55\.7% \[47\.5, 64\.1\]0\.62553\.0% \[44\.6, 61\.2\]Few\-shot vs Real13421\.1% \[14\.9, 28\.5\]58\.8% \[50\.5, 66\.9\]0\.85140\.4% \[32\.4, 48\.8\]OMRA vs Real13225\.6% \[19\.1, 33\.8\]62\.4% \[53\.6, 69\.9\]0\.82340\.4% \[32\.2, 48\.7\]Table 4:Human evaluation against matched real platform answers\.Win rates include 95% confidence intervals;κ\\kappadenotes inter\-annotator agreement, and Rev\. denotes the perception–preference reversal rate\.Finding 1: AI preference does not transfer to real user engagement\.A predictor trained on Synthetic answers cannot reliably rank real platform answers, although a predictor trained on real answers transfers reasonably well to Synthetic answers\.

Averaged over four base models \(Table[3](https://arxiv.org/html/2609.18282#S4.T3)\), Synthetic→\\toSynthetic reachesρ=0\.947\\rho=0\.947and Real→\\toSynthetic reachesρ=0\.759\\rho=0\.759, whereas Synthetic→\\toReal drops toρ=−0\.073\\rho=\-0\.073\. This asymmetry shows that internal consistency in AI preference does not imply transfer to real user engagement\.

Finding 2: AI preference overweights explicit logic relative to real user engagement\.Explicit logical structure is positively associated with target engagement in Synthetic answers, whereas real user engagement is more strongly associated with affective and expressive salience\.

Figure[4](https://arxiv.org/html/2609.18282#S4.F4)\(a\) shows the dimension\-wise association between ontology scores and within\-question engagement levels\. Logic is positively associated with engagement in Synthetic answers but weakly negatively associated with real user engagement, while affect and expression carry stronger positive associations in real platform answers\. We refer to this pattern aslogic overbinding\. ToT and RoT do not consistently reduce the gap across model families, indicating that reasoning augmentation alone does not recover the patterns associated with real user engagement\.

Finding 3: AI\-generated answers occupy a displaced preference geometry\.Real and AI\-generated answers combine logic, affect, and expression differently, placing them in distinct regions of the joint preference space\.

Figure[4](https://arxiv.org/html/2609.18282#S4.F4)\(b\) shows that Direct answers occupy a more logic\-heavy and lower\-salience region than real platform answers\. Few\-shot moves partially toward the real\-answer distribution, while OMRA moves substantially closer\. The MMD and centroid results in Table[1](https://arxiv.org/html/2609.18282#S3.T1)provide corresponding quantitative evidence\.

### 5\.2Human Evaluation

Findings 1–3 characterize the gap in the ontological preference space\. We next examine whether reducing this gap requires making answers appear human\-written\.

##### Setup\.

We compare each method’s answers with matched real platform answers\. Each item is evaluated by all 30 annotators, who answer two questions: \(i\) which answer appears more human\-written, and \(ii\) which answer they would prefer to like or endorse\. Full instructions, interface details, and the reversal\-rate definition are provided in Appendix[E](https://arxiv.org/html/2609.18282#A5)\.

Finding 4: Pairwise preference judgments can diverge from perceived human\-likeness\.Annotators may prefer an AI\-generated answer even when they judge the matched real answer as more likely to be human\-written\.

Direct answers are judged less human\-like in 97\.3% of comparisons, yet win 55\.7% of pairwise preference judgments\. OMRA achieves the highest preferred win rate at 62\.4%, while its perceived human\-likeness win rate is 25\.6%\. Since the confidence intervals for OMRA and Few\-shot overlap, we treat this result as supportive rather than statistically decisive evidence\. Overall, pairwise preference judgment is not reducible to perceived human\-likeness\.

## 6Conclusion

We introducedOntological Preference Measurementto compare AI preference and real user engagement in a shared interpretable space\. Our analysis identifies an AI preference–user engagement gap across transfer, marginal, and geometric comparisons, characterized bylogic overbinding: AI preference places disproportionate weight on explicit logical presentation\. Based on this diagnosis, we proposedOntology\-Masked Reasoning Autoencoding \(OMRA\), which identifies and reconstructs over\-explained spans and reduces the measured gap by an average of54\.4%across four LLM families\. Human evaluation further shows that OMRA wins 62\.4% of pairwise preference judgments against matched real platform answers, although the real answers are more often judged to be human\-written\. These results show that pairwise preference judgment is not reducible to perceived human\-likeness\.

## Limitations

\(1\) Dependence on ontology extraction quality\.Our framework relies on LLM\-based ontology extraction\. Although the extractor shows strong agreement with human annotations, errors may still occur for implicit affect, sarcasm, culturally specific rhetoric, and highly context\-dependent expressions\.

\(2\) Inference\-time correction does not change the model’s internal preference\.OMRA revises generated answers at inference time, but does not directly modify the model’s underlying preference representation or decoding behavior\. Future work may explore more durable model\-level alignment methods\.

\(3\) Offline evaluation, data constraints, and dual\-use risks\.OMRA is evaluated offline and pairwise preference does not fully reproduce real platform dynamics\. We therefore use OMRA as a controlled intervention rather than a tool for maximizing engagement\. Our corpus contains publicly visible answer\-level data without user profiling; released resources will remove user identifiers and follow platform\-specific redistribution constraints\. Live deployment would require additional ethical safeguards against engagement manipulation and large\-scale content generation\.

## References

- Chakrabartyet al\.\(2019\)T\. Chakrabarty, C\. Hidey, S\. Muresan, K\. Mckeown, and A\. HwangAMPERSAND: argument mining for persuasive online discussions\.InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing \(EMNLP\-IJCNLP\),pp\. 2933–2943\.Cited by:[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Christianoet al\.\(2017\)P\. F\. Christiano, J\. Leike, T\. Brown, M\. Martic, S\. Legg, and D\. AmodeiDeep reinforcement learning from human preferences\.Advances in neural information processing systems30\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Cuiet al\.\(2026a\)Z\. Cui, B\. Wang, Q\. Liu, Y\. Wang, Z\. Zhou, Y\. Liang, and Y\. WangAugur: modeling covariate causal associations in time series via large language models\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 764–787\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Cuiet al\.\(2026b\)Z\. Cui, X\. Yin, Y\. Tang, X\. Zhang, Y\. Hu, S\. Zhong, W\. Tang, Y\. Liang, W\. Li, M\. Jin,et al\.EarthVerse: benchmarking scientific agents across dynamic earth systems and natural hazards\.arXiv preprint arXiv:2608\.23525\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Denget al\.\(2020\)S\. Deng, Y\. Jiang, H\. Li, and Y\. LiuWho contributes what? scrutinizing the activity data of 4\.2 million zhihu users via immersion scores\.Information Processing & Management57\(5\),pp\. 102274\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Donget al\.\(2025\)Q\. Dong, L\. Dong, X\. Zhang, Z\. Sui, and F\. WeiSelf\-boosting large language models with synthetic preference data\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 65440–65463\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1)\.
- Feng and Hirst \(2011\)V\. W\. Feng and G\. HirstClassifying arguments by scheme\.InProceedings of the 49th annual meeting of the association for computational linguistics: Human language technologies,pp\. 987–996\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Gruber \(1993\)T\. R\. GruberA translation approach to portable ontology specifications\.Knowledge acquisition5\(2\),pp\. 199–220\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Heet al\.\(2022\)K\. He, X\. Chen, S\. Xie, Y\. Li, P\. Dollár, and R\. GirshickMasked autoencoders are scalable vision learners\.InProceedings of the IEEE/CVF conference on computer vision and pattern recognition,pp\. 16000–16009\.Cited by:[§3\.3](https://arxiv.org/html/2609.18282#S3.SS3.p2.1)\.
- Huet al\.\(2024\)Z\. Hu, L\. Song, J\. Zhang, Z\. Xiao, T\. Wang, Z\. Chen, N\. J\. Yuan, J\. Lian, K\. Ding, and H\. XiongExplaining length bias in llm\-based preference evaluations\.arXiv preprint arXiv:2407\.01085\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Jinet al\.\(2026\)D\. Jin, Y\. Zhou, J\. Zhou, J\. Ma, R\. Guo, and D\. GuoSimtoken: a simple baseline for referring audio\-visual segmentation\.InICASSP 2026\-2026 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 22702–22706\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.
- Lambertet al\.\(2025\)N\. Lambert, V\. Pyatkin, J\. Morrison, L\. Miranda, B\. Y\. Lin, K\. Chandu, N\. Dziri, S\. Kumar, T\. Zick, Y\. Choi,et al\.Rewardbench: evaluating reward models for language modeling\.InFindings of the Association for Computational Linguistics: NAACL 2025,pp\. 1755–1797\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1),[§1](https://arxiv.org/html/2609.18282#S1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Lawrence and Reed \(2019\)J\. Lawrence and C\. ReedArgument mining: a survey\.Computational linguistics45\(4\),pp\. 765–818\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Leiet al\.\(2025\)Y\. Lei, C\. Zhang, Z\. Liu, H\. Leng, S\. Liu, T\. Gao, Q\. Liu, and Y\. WangGODBench: a benchmark for multimodal large language models in video comment art\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 11884–11952\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1),[§4\.1](https://arxiv.org/html/2609.18282#S4.SS1.p1.1)\.
- Liet al\.\(2025\)D\. Li, R\. Sun, Y\. Huang, M\. Zhong, B\. Jiang, J\. Han, X\. Zhang, W\. Wang, and H\. LiuPreference leakage: a contamination problem in llm\-as\-a\-judge\.arXiv preprint arXiv:2502\.01534\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p2.1)\.
- Liet al\.\(2024\)T\. Li, W\. Chiang, E\. Frick, L\. Dunlap, T\. Wu, B\. Zhu, J\. E\. Gonzalez, and I\. StoicaFrom crowdsourced data to high\-quality benchmarks: arena\-hard and benchbuilder pipeline\.arXiv preprint arXiv:2406\.11939\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Liet al\.\(2026\)W\. Li, Z\. Song, H\. Zhou, J\. Yu, Y\. Zhang, and W\. YangLora\-mixer: coordinate modular lora experts through serial attention routing\.InInternational Conference on Learning Representations,Vol\.2026,pp\. 14694–14716\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1)\.
- Lippi and Torroni \(2016\)M\. Lippi and P\. TorroniArgumentation mining: state of the art and emerging trends\.ACM Transactions on Internet Technology \(TOIT\)16\(2\),pp\. 1–25\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Liuet al\.\(2025\)A\. Liu, A\. Mei, B\. Lin, B\. Xue, B\. Wang, B\. Xu, B\. Wu, B\. Zhang, C\. Lin, C\. Dong,et al\.Deepseek\-v3\. 2: pushing the frontier of open large language models\.arXiv preprint arXiv:2512\.02556\.Cited by:[§D\.1](https://arxiv.org/html/2609.18282#A4.SS1.p1.1)\.
- Liuet al\.\(2026a\)H\. Liu, D\. Brahma, and R\. HenaoCalibrating model\-based evaluation metrics for summarization\.arXiv preprint arXiv:2604\.17200\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.
- Liuet al\.\(2026b\)H\. Liu, L\. Ding, and R\. HenaoLearning to control summaries with score ranking\.arXiv preprint arXiv:2604\.17197\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.
- Liu and Henao \(2025\)H\. Liu and R\. HenaoLearning to substitute words with model\-based score ranking\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 11551–11565\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.
- Maet al\.\(2025\)J\. Ma, Z\. Cui, B\. Wang, P\. Wang, Z\. Zhou, Z\. Zhao, and Y\. WangCausal learning meet covariates: empowering lightweight and effective nationwide air quality forecasting\.\.InIJCAI,pp\. 3171–3179\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Noyet al\.\(2001\)N\. F\. Noy D\. L\. McGuinnesset al\.Ontology development 101: a guide to creating your first ontology\.Stanford knowledge systems laboratory technical report KSL\-01\-05 and …\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Ouyanget al\.\(2022\)L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.Training language models to follow instructions with human feedback\.Advances in neural information processing systems35,pp\. 27730–27744\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1),[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1),[§1](https://arxiv.org/html/2609.18282#S1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Rafailovet al\.\(2023\)R\. Rafailov, A\. Sharma, E\. Mitchell, C\. D\. Manning, S\. Ermon, and C\. FinnDirect preference optimization: your language model is secretly a reward model\.Advances in neural information processing systems36,pp\. 53728–53741\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1),[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Rallapalliet al\.\(2026\)S\. Rallapalli, S\. Gallagher, R\. Yurko, T\. Brooks, C\. Loughin, M\. Sezgin, and V\. TurriInterpretable stylistic variation in human and llm writing across genres, models, and decoding strategies\.arXiv preprint arXiv:2604\.14111\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1)\.
- Rameezet al\.\(2022\)R\. Rameez, H\. A\. Rahmani, and E\. YilmazViralBERT: a user focused bert\-based approach to virality prediction\.InAdjunct Proceedings of the 30th ACM Conference on User Modeling, Adaptation and Personalization,pp\. 85–89\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Rodrigueset al\.\(2026\)F\. A\. Rodrigues, N\. F\. Sturm, and F\. L\. PinheiroA linguistic comparison between human\-and ai\-generated content\.Iscience29\(3\)\.Cited by:[§1](https://arxiv.org/html/2609.18282#S1.p1.1)\.
- Saitoet al\.\(2023\)K\. Saito, A\. Wachi, K\. Wataoka, and Y\. AkimotoVerbosity bias in preference labeling by large language models\.arXiv preprint arXiv:2310\.10076\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Songet al\.\(2026\)Z\. Song, X\. Li, Y\. Zhang, X\. Zhang, W\. Yang, and J\. YuSocial intelligence modeling: a comprehensive survey from social perception to social simulation\.Note:ResearchGate preprintPreprint available on ResearchGateExternal Links:[Document](https://dx.doi.org/10.13140/RG.2.2.21157.87528),[Link](https://www.researchgate.net/publication/411123257_Social_Intelligence_Modeling_A_Comprehensive_Survey_from_Social_Perception_to_Social_Simulation)Cited by:[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Stab and Gurevych \(2017\)C\. Stab and I\. GurevychParsing argumentation structures in persuasive essays\.Computational Linguistics43\(3\),pp\. 619–659\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Tanet al\.\(2016\)C\. Tan, V\. Niculae, C\. Danescu\-Niculescu\-Mizil, and L\. LeeWinning arguments: interaction dynamics and persuasion strategies in good\-faith online discussions\.InProceedings of the 25th international conference on world wide web,pp\. 613–624\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px1.p1.1)\.
- Touvronet al\.\(2023\)H\. Touvron, T\. Lavril, G\. Izacard, X\. Martinet, M\. Lachaux, T\. Lacroix, B\. Rozière, N\. Goyal, E\. Hambro, F\. Azhar, A\. Rodriguez, A\. Joulin, E\. Grave, and G\. LampleLLaMA: open and efficient foundation language models\.External Links:2302\.13971,[Link](https://arxiv.org/abs/2302.13971)Cited by:[§D\.1](https://arxiv.org/html/2609.18282#A4.SS1.p1.1)\.
- Wanget al\.\(2026\)D\. Wang, Y\. Zhang, J\. Yu, Y\. P\. Chen, C\. Xu, and Z\. SongSeeing further and wider: joint spatio\-temporal enlargement for micro\-video popularity prediction\.arXiv preprint arXiv:2604\.20311\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Wanget al\.\(2025\)Z\. Wang, H\. Zhang, X\. Li, K\. Huang, C\. Han, S\. Ji, S\. Kakade, H\. Peng, and H\. JiEliminating position bias of language models: a mechanistic approach\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 91212–91239\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Wataokaet al\.\(2024\)K\. Wataoka, T\. Takahashi, and R\. RiSelf\-preference bias in llm\-as\-a\-judge\.arXiv preprint arXiv:2410\.21819\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p1.1)\.
- William and Thompson \(1988\)M\. William and S\. ThompsonRhetorical structure theory: towards a functional theory of text organization\.Text8\(3\),pp\. 243–281\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1)\.
- Wuet al\.\(2026\)Y\. Wu, Y\. Zhang, L\. Ye, G\. Zeng, J\. Yu, C\. Xu, and Z\. SongHotComment: a benchmark for evaluating popularity of online comments\.arXiv preprint arXiv:2604\.25614\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Yaoet al\.\(2023\)S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. Griffiths, Y\. Cao, and K\. NarasimhanTree of thoughts: deliberate problem solving with large language models\.Advances in neural information processing systems36,pp\. 11809–11822\.Cited by:[§4\.1](https://arxiv.org/html/2609.18282#S4.SS1.p1.1)\.
- Yeet al\.\(2025a\)J\. Ye, Y\. Wang, Y\. Huang, D\. Chen, Q\. Zhang, N\. Moniz, T\. Gao, W\. Geyer, C\. Huang, P\. Chen,et al\.Justice or prejudice? quantifying biases in llm\-as\-a\-judge\.InInternational Conference on Learning Representations,Vol\.2025,pp\. 102351–102390\.Cited by:[§A\.3](https://arxiv.org/html/2609.18282#A1.SS3.p2.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px2.p1.1)\.
- Yeet al\.\(2025b\)L\. Ye, Y\. Zhang, Y\. Wu, Y\. P\. Chen, J\. Yu, W\. Yang, and Z\. SongMvp: winning solution to smp challenge 2025 video track\.InProceedings of the 33rd ACM International Conference on Multimedia,pp\. 14079–14085\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Zhanget al\.\(2023\)H\. Zhang, H\. Song, S\. Li, M\. Zhou, and D\. SongA survey of controllable text generation using transformer\-based pre\-trained language models\.ACM Computing Surveys56\(3\),pp\. 1–37\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1),[§2](https://arxiv.org/html/2609.18282#S2.SS0.SSS0.Px3.p1.1)\.
- Zhanget al\.\(2026a\)X\. Zhang, Y\. Zhang, Z\. Chen, J\. Yu, W\. Yang, and Z\. SongLogical phase transitions: understanding collapse in LLM logical reasoning\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 18836–18860\.External Links:[Link](https://aclanthology.org/2026.acl-long.858/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.858),ISBN 979\-8\-89176\-390\-6Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p2.1)\.
- Zhang and Gosline \(2023\)Y\. Zhang and R\. GoslineHuman favoritism, not ai aversion: people’s perceptions \(and bias\) toward generative ai, human experts, and human–gai collaboration in persuasive content generation\.Judgment and Decision Making18,pp\. e41\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.
- Zhanget al\.\(2026b\)Y\. Zhang, Y\. Ai, Z\. Ying, Q\. Mi, J\. Yu, W\. Yang, and Z\. SongCoupling macro dynamics and micro states for long\-horizon social simulation\.arXiv preprint arXiv:2604\.05516\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Zhanget al\.\(2026c\)Y\. Zhang, Z\. Ying, X\. Zhang, J\. Yu, P\. Fang, X\. Chen, W\. Yang, and Z\. SongIntervenSim: intervention\-aware social network simulation for opinion dynamics\.arXiv preprint arXiv:2604\.06600\.Cited by:[§A\.2](https://arxiv.org/html/2609.18282#A1.SS2.p1.1)\.
- Zhanget al\.\(2026d\)Y\. Zhang, X\. Zhang, J\. Sheng, W\. Li, J\. Yu, Y\. P\. Chen, W\. Yang, and Z\. SongSemantic\-aware logical reasoning via a semiotic framework\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 18349–18374\.External Links:[Link](https://aclanthology.org/2026.acl-long.835/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.835),ISBN 979\-8\-89176\-390\-6Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p2.1)\.
- Zhenget al\.\(2026\)Z\. Zheng, Y\. Zhang, J\. Sheng, J\. Yu, and Z\. SongSurfacing the unsaid: cue\-bench for affective stance in chinese discourse\.arXiv preprint arXiv:2608\.10810\.Cited by:[§A\.1](https://arxiv.org/html/2609.18282#A1.SS1.p1.1)\.
- Zhouet al\.\(2026\)Y\. Zhou, J\. Ma, Y\. Peng, Z\. Sun, Y\. Bai, and B\. F\. KarlssonExoActor: exocentric video generation as generalizable interactive humanoid control\.arXiv preprint arXiv:2604\.27711\.Cited by:[§A\.4](https://arxiv.org/html/2609.18282#A1.SS4.p1.1)\.

## Appendix

The appendix provides supplementary details for related work, dataset construction, ontology design, experimental protocols, human evaluation, case studies, and prompts\.

## The Usage of LLM

In accordance with ACL guidelines, we used large language models solely for writing assistance and language refinement\.

## Appendix ARelated Work

### A\.1Ontology, Discourse Structure, and Argument Mining

Ontology\-based modeling represents a domain through interpretable concepts, relations, and constraints[Gruber \(1993\)](https://arxiv.org/html/2609.18282#bib.bib1);[Noy et al\. \(2001\)](https://arxiv.org/html/2609.18282#bib.bib2), and has been widely used in NLP to organize discourse and persuasion phenomena\. Rhetorical Structure Theory[William and Thompson \(1988\)](https://arxiv.org/html/2609.18282#bib.bib32)provides a classical account of discourse relations, while later work examines how framing, emotion, and rhetorical choices shape persuasion and user response in online discussion[Feng and Hirst \(2011\)](https://arxiv.org/html/2609.18282#bib.bib33);[Tan et al\. \(2016\)](https://arxiv.org/html/2609.18282#bib.bib34);[Zheng et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib43)\.

A closely related line is argument mining, which extracts claims, premises, evidence, and argumentative relations from natural language[Lippi and Torroni \(2016\)](https://arxiv.org/html/2609.18282#bib.bib3);[Lawrence and Reed \(2019\)](https://arxiv.org/html/2609.18282#bib.bib4);[Stab and Gurevych \(2017\)](https://arxiv.org/html/2609.18282#bib.bib5)\. Prior work develops corpora and models for component identification, relation classification, and argument quality assessment, focusing primarily on how arguments are structured and supported\. Our work uses ontology for a different purpose\. Rather than analyzing argument structure alone, we organize online answers along logic[Zhang et al\. \(2026d\)](https://arxiv.org/html/2609.18282#bib.bib42);[Zhang et al\. \(2026a\)](https://arxiv.org/html/2609.18282#bib.bib41), affect, and expression, allowing Real, Synthetic, Few\-shot, and OMRA\-corrected answers to be compared in a shared interpretable preference space\.

### A\.2Online Engagement Prediction

Online popularity and engagement have been studied across social media, micro\-video, and community discussion settings using signals such as user authority, temporal dynamics[Cui et al\. \(2026b\)](https://arxiv.org/html/2609.18282#bib.bib45);[Cui et al\. \(2026a\)](https://arxiv.org/html/2609.18282#bib.bib46), propagation networks, content features, and neural representations[Deng et al\. \(2020\)](https://arxiv.org/html/2609.18282#bib.bib12);[Rameez et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib13);[Ye et al\. \(2025b\)](https://arxiv.org/html/2609.18282#bib.bib7);[Wang et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib8);[Zhang et al\. \(2026c\)](https://arxiv.org/html/2609.18282#bib.bib10);[Zhang et al\. \(2026b\)](https://arxiv.org/html/2609.18282#bib.bib11);[Lei et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib6);[Wu et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib9)\. These approaches generally treat engagement as a scalar prediction target and optimize forecasting[Ma et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib47)performance, but rarely examine how logic, affect, rhetoric, and informativeness are differently associated with engagement\.

Our setting differs in both object and goal\. We focus on long\-form open\-platform answers and compare answers only within the same question context\. Rather than predicting absolute popularity, we examine which textual qualities vary with real user engagement and whether LLMs associate the same qualities with higher target engagement\. This motivates our within\-question engagement levels and ontology\-based decomposition\.

### A\.3AI preference and LLM\-as\-a\-Judge

Preference data underpins reward modeling, RLHF, and alignment research[Christiano et al\. \(2017\)](https://arxiv.org/html/2609.18282#bib.bib15);[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib16);[Rafailov et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib25);[Rallapalli et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib29);[Wataoka et al\. \(2024\)](https://arxiv.org/html/2609.18282#bib.bib27);[Li et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib48), as well as benchmarks for reward\-model evaluation[Lambert et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib17);[Li et al\. \(2024\)](https://arxiv.org/html/2609.18282#bib.bib14)\. Because human annotation is costly, recent work increasingly relies on AI preference data and LLM\-as\-a\-judge evaluation, in which strong models score or compare outputs as proxies for human annotators[Dong et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib23)\.

A growing literature documents systematic biases in these signals, including verbosity and length bias[Hu et al\. \(2024\)](https://arxiv.org/html/2609.18282#bib.bib18);[Saito et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib20), position and order effects[Wang et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib21);[Li et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib31), and self\-preference toward outputs from the same model family[Ye et al\. \(2025a\)](https://arxiv.org/html/2609.18282#bib.bib19)\. These biases are usually evaluated against benchmark labels or other model judgments\. We instead compareAI preference—the qualities LLMs associate with higher target engagement—withreal user engagementobserved on online platforms\. Our results show that internally consistent AI preference does not necessarily transfer to real platform answers and may place disproportionate weight on explicit logical presentation\.

### A\.4Preference\-Controlled Generation

Preference\-aligned generation optimizes models toward preferred outputs through reward modeling and policy optimization[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.18282#bib.bib16);[Rafailov et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib25);[Liu and Henao \(2025\)](https://arxiv.org/html/2609.18282#bib.bib36);[Liu et al\. \(2026b\)](https://arxiv.org/html/2609.18282#bib.bib37), while controllable generation steers outputs toward attributes such as style, sentiment, topic, or quality[Zhang et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib22);[Zhang and Gosline \(2023\)](https://arxiv.org/html/2609.18282#bib.bib28);[Zhou et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib39);[Jin et al\. \(2026\)](https://arxiv.org/html/2609.18282#bib.bib40);[Liu et al\. \(2026a\)](https://arxiv.org/html/2609.18282#bib.bib38)\. Common approaches include prompting, demonstrations, latent attribute control, classifier guidance, and reward\-based fine\-tuning\.

These methods generally assume that the target preference or attribute can be specified through a reward signal, prompt, or small set of demonstrations\. However, they do not directly identify which answer qualities should be changed when AI preference diverges from real user engagement\. OMRA addresses this problem using an explicit ontology: it identifies and masks over\-explained spans, then reconstructs the answer under stance\-preservation, factual\-consistency, coherence, and preference constraints\. In this way, OMRA provides a targeted intervention derived from the diagnosed pattern rather than relying on unrestricted style imitation\.

## Appendix BDataset Construction and Benchmark Sampling

### B\.1Platforms and Collection

We construct a large\-scale corpus of question–answer pairs from three open online platforms with naturally occurring engagement signals: Zhihu, Quora, and Reddit\. Zhihu and Quora provide long\-form Q&A content, while Reddit contributes discussion threads from selected question\-answering and explanation\-oriented communities\.

For each platform, we collect publicly visible question–answer pairs and the engagement metadata available at collection time\. The corpus spans 2020–2026 and covers 10 primary domains and 35 sub\-domains, as summarized in Figure[3](https://arxiv.org/html/2609.18282#S3.F3)\. We retain multiple answers to the same question because our analysis relies on within\-question comparison\. The analysis is conducted at the answer level; we do not construct or use user profiles\.

### B\.2Filtering and Quality Control

We apply the following filters before forming the final corpus:

- •Length filtering\.We remove extremely short or unusually long answers to exclude trivial replies, link\-only responses, and low\-quality dumps\.
- •Deduplication\.We remove near\-duplicate answers within the same question using lexical similarity\.
- •Availability filtering\.We discard answers marked as deleted, unavailable, or abnormally inaccessible at collection time, and do not retain deleted content\.
- •Minimum\-answer filtering\.We keep only questions with at least four surviving answers, which is required for defining four within\-question engagement levelsℓ⁡\(a\)∈\{0,1,2,3\}\\ell\(a\)\\in\\\{0,1,2,3\\\}\.
- •Engagement\-anomaly filtering\.We remove questions with extreme engagement concentration or other anomalous feedback patterns\.
- •Privacy filtering\.We remove usernames, user identifiers, profile links, and other unnecessary personal metadata from released resources\.

We preserve the natural distribution of each platform rather than explicitly rebalancing the corpus\. Corpus\-level statistics and domain distributions are reported in Figure[3](https://arxiv.org/html/2609.18282#S3.F3)\.

### B\.3Within\-Question Engagement Levels

Absolute vote counts are affected by exposure, platform traffic, author reputation, recommendation algorithms, and time\. We therefore do not compare raw votes across questions\. Instead, for each questionqqwith answer set𝒜q\\mathcal\{A\}\_\{q\}, we rank answers by their observed platform engagement signal and map them to four ordinal levels:

ℓ⁡\(a\)=ϕq​\(rankq⁡\(v⁡\(a\)\)\)∈\{0,1,2,3\}\.\\ell\(a\)=\\phi\_\{q\}\(\\operatorname\{rank\}\_\{q\}\(v\(a\)\)\)\\in\\\{0,1,2,3\\\}\.This construction compares answers only within the same question context, reducing confounds from question exposure and topic popularity\. The resulting label is a relative measure of real user engagement rather than a direct measurement of latent human preference\.

### B\.4Controlled Benchmark Sampling

For the controlled experiments in §[4](https://arxiv.org/html/2609.18282#S4), we construct a fixed benchmark of 3,600 real platform answers\. We select questions with sufficiently many surviving answers and clear engagement separation so that all four within\-question engagement levels can be represented\.

For each of 900 questions, we retain one real answer at each engagement level\. This keeps the question fixed while varying engagement level, reducing confounds from topic, exposure, and question\-level popularity\. For each question–level pair, we generate a matched Synthetic answer conditioned on the target engagement level, a Few\-shot answer additionally conditioned on a level\-matched real reference, and an OMRA\-corrected answer\. The four answer regimes—Real, Synthetic, Few\-shot, and OMRA—are therefore compared under shared question contexts and aligned engagement levels\.

### B\.5Scope, Privacy, and Release

Platform engagement is influenced by exposure, timing, author identity, ranking algorithms, and community norms in addition to textual content\. Within\-question comparison reduces but does not eliminate these platform\-specific effects\. Our goal is therefore not to equate platform engagement with human preference, but to compare AI preference with the answer qualities associated with observed real user engagement\.

For reproducibility, we will release the processing code, prompts, ontology schema, filtering rules, benchmark splits, and all de\-identified text and metadata permitted by each platform\. Where full\-text redistribution is restricted, we will release retrieval identifiers and derived metadata instead\. All released resources will follow platform\-specific redistribution constraints and exclude usernames, user identifiers, and deleted or unavailable content\.

## Appendix COntology Design and Scoring Reliability

### C\.1Ontology Schema and Conceptualization

Our Ontological Preference Measurement framework represents each answer through a fixed schema ofconcepts,relations, andpatterns\. Rather than replacing the answer with a fully formal graph, the ontology makes preference\-relevant textual properties explicit\. Given an answeraa, the extractor produces

𝒪⁡\(a\)=\{𝒪L​\(a\),𝒪A​\(a\),𝒪E​\(a\)\},\\mathcal\{O\}\(a\)=\\\{\\mathcal\{O\}^\{L\}\(a\),\\mathcal\{O\}^\{A\}\(a\),\\mathcal\{O\}^\{E\}\(a\)\\\},where𝒪L\\mathcal\{O\}^\{L\},𝒪A\\mathcal\{O\}^\{A\}, and𝒪E\\mathcal\{O\}^\{E\}denote the logical, affective, and expressive layers\.Conceptsare textual units such as claims, evidence, emotions, values, and expressions;relationsdescribe their interactions; andpatternsare recurring local structures used as interpretable cues for scoring\.

### C\.2Ontology Taxonomy

Table[5](https://arxiv.org/html/2609.18282#A3.T5)summarizes the core ontology classes, following the three\-layer decomposition used throughout the paper\. Figure[5](https://arxiv.org/html/2609.18282#A3.F5)illustrates how the logic, affect, and expression layers organize preference\-relevant cues before they are converted into the preference vector𝐬⁡\(a\)\\mathbf\{s\}\(a\)\.

Table 5:Ontology taxonomy for preference decomposition\.LayerClassSym\.Description / Preference RoleLogicMajorClaimM​CMCThe central stance, conclusion, or main judgment of the answer\.ClaimCCA supporting or contrasting argumentative unit that develops the stance\.EvidenceEEConcrete backing such as facts, examples, statistics, historical references, or personal experience\.CounterpointC​PCPAn alternative viewpoint, objection, or contrastive claim acknowledged by the answer\.AffectEmotionE​mEmAffective cues such as empathy, anger, admiration, nostalgia, concern, or disappointment\.ValueVVNormative or moral concerns such as fairness, dignity, responsibility, freedom, or security\.GroupValueG​VGVCollective identities or group\-level concerns that make an answer socially resonant\.ExpressionFramingF​rFrThe way an answer packages its stance through perspective, contrast, emphasis, or narrative framing\.RhetoricalDeviceR​DRDStylistic devices such as analogy, metaphor, contrast, irony, or compression\.QuoteQQConcise, memorable, or reusable expressions with potential platform salience\.![Refer to caption](https://arxiv.org/html/2609.18282v1/ontology_example.png)Figure 5:Illustration of the ontology schema\.The ontology organizes preference\-relevant textual cues into logic, affect, and expression layers\. Logic captures claims, evidence, and support relations; affect captures emotions and value appeals; expression captures framing, readability, and quotable forms\. The figure is intended as a schema illustration rather than a complete graph representation of every extracted answer\.
### C\.3Representative Relations and Patterns

The ontology records representative relations among extracted concepts\. These relations are not a rigid formal grammar, but structured evidence for layer\-wise scoring\. Table[6](https://arxiv.org/html/2609.18282#A3.T6)lists the main relation and pattern types\.

Table 6:Representative ontology relations and preference patterns\.LayerRelation / PatternInterpretationLogicE→supportsCE\\xrightarrow\{\\textit\{supports\}\}CEvidence provides factual, experiential, or illustrative support for a claim\.C→supportsM​CC\\xrightarrow\{\\textit\{supports\}\}MCA claim reinforces the answer’s central stance\.C​P→contrastsCCP\\xrightarrow\{\\textit\{contrasts\}\}CA counterpoint or alternative view is acknowledged and contrasted with the answer’s position\.Serial supportA multi\-step support chain where evidence leads to intermediate claims and then to the main stance\.Convergent supportMultiple pieces of evidence or claims independently support the same conclusion\.AffectC→evokesE​mC\\xrightarrow\{\\textit\{evokes\}\}EmA claim or example evokes a recognizable emotional response\.E​m→activatesV/G​VEm\\xrightarrow\{\\textit\{activates\}\}V/GVAn emotion makes a value, identity, or group concern salient\.C→appeals\_toV/G​VC\\xrightarrow\{\\textit\{appeals\\\_to\}\}V/GVA claim directly invokes a value or collective concern\.Value activationA recurring pattern in which stance, emotion, and value appeal reinforce each other\.ExpressionF​r→framesCFr\\xrightarrow\{\\textit\{frames\}\}CA framing device shapes how a claim is interpreted\.R​D→highlightsC/E​mRD\\xrightarrow\{\\textit\{highlights\}\}C/EmA rhetorical device makes a claim or emotion more salient\.R​D→producesQRD\\xrightarrow\{\\textit\{produces\}\}QA rhetorical device compresses an idea into a memorable expression\.Quotable framingA pattern where style and compression make an answer easier to remember or share\.
### C\.4Preference Vector Construction

The extracted ontology is mapped into a layer\-wise preference vector

𝐬⁡\(a\)=\[𝐬L​\(a\);𝐬A​\(a\);𝐬E​\(a\)\]\.\\mathbf\{s\}\(a\)=\[\\mathbf\{s\}^\{L\}\(a\);\\mathbf\{s\}^\{A\}\(a\);\\mathbf\{s\}^\{E\}\(a\)\]\.The original answer text remains the primary evidence, while ontology units serve as structured cues\. Table[7](https://arxiv.org/html/2609.18282#A3.T7)summarizes the scoring dimensions\.

Table 7:Preference vector dimensions and ontology\-guided scoring cues\.LayerPreference AxisOntological Cues and IndicatorsLogicStructureClarity of stance, organization of claims, and coherence of claim–evidence relations\.Evidence StrengthPresence and relevance of facts, examples, experience, or other concrete backing\.Reasoning DepthDegree of multi\-step explanation, causal development, and layered justification\.CounterargumentPresence of contrast, qualification, or engagement with alternative viewpoints\.AffectEmotion IntensityStrength and clarity of affective cues such as empathy, anger, admiration, nostalgia, or concern\.Value ActivationAppeals to shared values, group concerns, identity, fairness, dignity, or responsibility\.Affective CoherenceWhether the affective tone supports the stance rather than appearing detached or inconsistent\.ExpressionReadabilityFluency, pacing, paragraph organization, and ease of comprehension\.Rhetorical SalienceUse of framing, contrast, analogy, metaphor, emphasis, or stylistic compression\.QuoteabilityPresence of concise, memorable, or reusable expressions likely to travel on the platform\.
### C\.5Scoring Protocol

We implement ontology extraction and scoring with structured prompts: the extractor receives the answer, fixed ontology schema, and constrained output format, identifies ontology units and relations, and assigns layer\-wise scores according to the rubric\. Unlike ordinary LLM\-as\-a\-judge scoring, the output space is schema\-constrained and decomposed into logical, affective, and expressive dimensions rather than collapsed into a single quality score\. This decomposition enables the marginal, geometric, and transfer gap analyses in the main paper\.

Table 8:Agreement between the primary ontology scorer and human consensus scores on a 100\-question subset\. Pearson and Spearman correlations measure ranking consistency; MAD is reported on the original score scale\.ScorePearsonSpearmanMADLogic sum0\.9420\.9320\.768Affect sum0\.8930\.8730\.675Expression sum0\.9340\.9350\.258Overall sum0\.9540\.9531\.311
### C\.6Scoring Reliability

We use DeepSeek\-V3\.2 as the primary ontology scorer for cost reasons\. To assess whether the scores are robust beyond a single extractor, we re\-score a randomly sampled 100\-question benchmark subset with two human annotators under the same rubric\. Table[8](https://arxiv.org/html/2609.18282#A3.T8)reports agreement between the primary extractor and human consensus scores using Pearson correlation, Spearman correlation, and mean absolute difference on the original score scale\.

Agreement is high across all reported scores: Pearson correlations exceed0\.890\.89and Spearman correlations exceed0\.870\.87, with the overall sum reaching Pearson0\.9540\.954and Spearman0\.9530\.953\. This suggests that the ontology\-based scores are reliable enough for our ranking\-based diagnostics\. Since our analyses rely mainly on relative comparisons, such as Spearman correlation and within\-question ranking, consistency in relative ordering is more important than exact absolute calibration\. We therefore use the primary extractor for full\-corpus scoring without further adjustment\.

## Appendix DExperiment Details

### D\.1Model Families and Decoding Settings

We instantiate the pipeline with four LLM families:DeepSeek\-V3\.2,Claude\-4\.5\-Haiku,GPT\-5\.4, andLlama\-3\.1\-70B[Liu et al\. \(2025\)](https://arxiv.org/html/2609.18282#bib.bib49);[Touvron et al\. \(2023\)](https://arxiv.org/html/2609.18282#bib.bib50)\. Across all generated regimes, we use temperature=1\.0=1\.0, maximum generation length=8192=8192tokens, and top\-p=1\.0p=1\.0\.

### D\.2Generation Regimes and Baselines

For each question–level pair in the controlled benchmark, we construct matched answers under four regimes:Real,Synthetic,Few\-shot, andOMRA\.Realuses the observed platform answer at the corresponding within\-question engagement level\.Syntheticdirectly prompts the base model under the target engagement level\.Few\-shotadditionally provides a level\-matched real reference answer from the same question context, enabling style and content imitation at that level\.OMRAtakes the Synthetic draft as input and applies ontology\-guided masking and reconstruction of over\-explained spans\.

We also compare with two reasoning\-augmented baselines\.Tree\-of\-Thought \(ToT\)encourages the model to generate and evaluate multiple intermediate reasoning paths before producing its answer\.Ripple\-of\-Thought \(RoT\)expands reasoning through broader associative and multi\-step elaboration\. These baselines assess whether reasoning\-augmented generation reduces the measured AI preference–user engagement gap\. All generated regimes are matched by question and target engagement level, so differences can be attributed to the generation or correction strategy rather than the underlying question\.

### D\.3OMRA Implementation

OMRA operates on a Synthetic draft and its ontology representation\. It first applies the fixed ontology extractor to identify units in the logic, affect, and expression layers\. Over\-explained spans are selected from ontology units associated with unnecessarily explicit reasoning, including claim–evidence bridges, causal justifications, enumerated steps, and explicit discourse transitions\. These spans are surface spans in the final answer rather than hidden reasoning traces\.

After span identification, OMRA masks the selected spans while preserving the surrounding context\. The reconstruction stage fills the masks under stance\-preservation, factual\-consistency, discourse\-coherence, and contextually supported salience constraints\. It preserves the main stance and non\-masked factual content, avoids unsupported additions, and introduces affective or expressive salience only when supported by the original draft\. A final verification pass checks for residual over\-explained spans associated with logic overbinding\. If such spans remain, a lightweight revision is applied; otherwise, the reconstructed answer is retained as the final OMRA\-corrected answer\.

### D\.4Evaluation Metrics

We evaluate the AI preference–user engagement gap from three complementary perspectives: marginal gap, geometric gap, and transfer gap\. These metrics correspond to the diagnostic framework in §[3\.2](https://arxiv.org/html/2609.18282#S3.SS2)and the main results in Table[1](https://arxiv.org/html/2609.18282#S3.T1)\.

##### Marginal gap\.

For each ontology dimensiond∈\{L,A,E\}d\\in\\\{L,A,E\\\}and regimesr,r′r,r^\{\\prime\}, we compare the empirical score distributions𝒫dr\\mathcal\{P\}\_\{d\}^\{r\}and𝒫dr′\\mathcal\{P\}\_\{d\}^\{r^\{\\prime\}\}using 1\-Wasserstein distance and Jensen–Shannon divergence:

Wdr,r′=W1​\(𝒫dr,𝒫dr′\),Jdr,r′=JS⁡\(𝒫dr,𝒫dr′\)\.W\_\{d\}^\{r,r^\{\\prime\}\}=W\_\{1\}\(\\mathcal\{P\}\_\{d\}^\{r\},\\mathcal\{P\}\_\{d\}^\{r^\{\\prime\}\}\),\\qquad J\_\{d\}^\{r,r^\{\\prime\}\}=\\mathrm\{JS\}\(\\mathcal\{P\}\_\{d\}^\{r\},\\mathcal\{P\}\_\{d\}^\{r^\{\\prime\}\}\)\.Lower values indicate that the generated regime is closer to the Real regime along the corresponding ontology dimension\.

##### Geometric gap\.

To compare regimes in the joint preference space, we compute centroid distance and maximum mean discrepancy \(MMD\) over the full ontology vector𝐬⁡\(a\)\\mathbf\{s\}\(a\):

Cr,r′\\displaystyle C^\{r,r^\{\\prime\}\}=‖𝔼𝒟r​\[𝐬\]−𝔼𝒟r′​\[𝐬\]‖2,\\displaystyle=\\bigl\\\|\\mathbb\{E\}\_\{\\mathcal\{D\}\_\{r\}\}\[\\mathbf\{s\}\]\-\\mathbb\{E\}\_\{\\mathcal\{D\}\_\{r^\{\\prime\}\}\}\[\\mathbf\{s\}\]\\bigr\\\|\_\{2\},MMDr,r′\\displaystyle\\operatorname\{MMD\}^\{r,r^\{\\prime\}\}=MMD⁡\(𝒟r,𝒟r′\)\.\\displaystyle=\\operatorname\{MMD\}\\bigl\(\\mathcal\{D\}\_\{r\},\\mathcal\{D\}\_\{r^\{\\prime\}\}\\bigr\)\.These metrics capture whether two regimes combine logic, affect, and expression in similar ways, beyond agreement on individual marginal dimensions\.

##### Transfer gap\.

To evaluate whether an engagement rule learned in one regime generalizes to another, we train a predictorfrf\_\{r\}on ontology vectors from regimerrand evaluate it on regimer′r^\{\\prime\}:

Tr→r′=ℳ⁡\(fr,𝒟r′\)\.T\_\{r\\to r^\{\\prime\}\}=\\mathcal\{M\}\(f\_\{r\},\\mathcal\{D\}\_\{r^\{\\prime\}\}\)\.We report Spearman correlation and Top\-1 accuracy\. Spearman correlation measures whether the predictor preserves within\-question engagement ordering, while Top\-1 accuracy measures whether it identifies the highest\-engagement answer\.

##### Transfer predictor\.

For transfer\-gap evaluation, we instantiatefrf\_\{r\}as a lightweight XGBoost predictor over ontology\-based preference vectors𝐬⁡\(a\)\\mathbf\{s\}\(a\)\. For each source regimerr, the predictor is trained to estimate the within\-question engagement levelℓ⁡\(a\)\\ell\(a\)and is then evaluated on a target regimer′r^\{\\prime\}without target\-regime fine\-tuning\. We use the same feature representation, training protocol, and fixed hyperparameter configuration across all source regimes, so that differences inTr→r′T\_\{r\\to r^\{\\prime\}\}reflect cross\-regime transfer rather than changes in model capacity or tuning\.

Spearmanρ\\rhomeasures whether the predictor preserves the within\-question engagement ordering in the target regime, and Top\-1 accuracy measures whether it identifies the highest\-engagement answer for each question\. Thus, the aggregation functionℳ\\mathcal\{M\}in §[3\.2](https://arxiv.org/html/2609.18282#S3.SS2)denotes this fixed evaluation protocol rather than a separately learned scalar objective\.

##### Average gap reduction\.

For each method, we summarize improvement over Direct generation using the average relative reduction over the marginal and geometric distance metrics:

Shrink=1\|𝒦\|​∑k∈𝒦GkDirect−GkMethodGkDirect,\\mathrm\{Shrink\}=\\frac\{1\}\{\|\\mathcal\{K\}\|\}\\sum\_\{k\\in\\mathcal\{K\}\}\\frac\{G\_\{k\}^\{\\mathrm\{Direct\}\}\-G\_\{k\}^\{\\mathrm\{Method\}\}\}\{G\_\{k\}^\{\\mathrm\{Direct\}\}\},where𝒦\\mathcal\{K\}contains the four distance metrics reported in Table[1](https://arxiv.org/html/2609.18282#S3.T1)\. Higher values indicate a larger reduction in the measured AI preference–user engagement gap\. For the transfer metrics, improvements are reported as absolute gains in Spearman correlation and relative gains in Top\-1 accuracy\.

MethodMarginal Gap↓\\downarrowGeometry Gap↓\\downarrowTransfer to Real↑\\uparrowLen\. RatioAvg\. Shrink↑\\uparrowWass\.JSMMDCent\.𝝆\\boldsymbol\{\\rho\}\(Δ\\Delta\)Top\-1Direct1\.4790\.1540\.2401\.969\-0\.0510\.2611\.00—Few\-shot0\.829\(43\.9%↓\\downarrow\)0\.085\(44\.8%↓\\downarrow\)0\.102\(57\.5%↓\\downarrow\)0\.971\(50\.7%↓\\downarrow\)0\.164\(\+0\.215\)0\.316\(21\.1%↑\\uparrow\)0\.9149\.2%Length\-ControlledDirect1\.020\(31\.0%↓\\downarrow\)0\.108\(29\.9%↓\\downarrow\)0\.145\(39\.6%↓\\downarrow\)1\.180\(40\.1%↓\\downarrow\)0\.082\(\+0\.133\)0\.294\(12\.6%↑\\uparrow\)0\.6633\.8%Length\-ControlledFew\-shot0\.650\(56\.1%↓\\downarrow\)0\.070\(54\.5%↓\\downarrow\)0\.076\(68\.3%↓\\downarrow\)0\.955\(51\.5%↓\\downarrow\)0\.188\(\+0\.239\)0\.331\(26\.8%↑\\uparrow\)0\.6957\.3%OMRA0\.452\(69\.4%↓\\downarrow\)0\.050\(67\.5%↓\\downarrow\)0\.038\(84\.2%↓\\downarrow\)0\.422\(78\.6%↓\\downarrow\)0\.225\(\+0\.276\)0\.358\(37\.2%↑\\uparrow\)0\.7174\.9%Table 9:Length\-controlled baselines on DeepSeek\.Len\. Ratio denotes mean answer length relative to Direct\. Length control reduces the measured gap, but OMRA achieves substantially greater reduction while producing answers of similar length\.Direct VariantMarginal Gap↓\\downarrowGeometry Gap↓\\downarrowTransfer to Real↑\\uparrowAvg\. Shrink↑\\uparrowWass\.JSMMDCent\.𝝆\\boldsymbol\{\\rho\}\(Δ\\Delta\)Top\-1Direct1\.4790\.1540\.2401\.969\-0\.0510\.261—Declarative1\.462\(1\.1%↓\\downarrow\)0\.152\(1\.3%↓\\downarrow\)0\.237\(1\.3%↓\\downarrow\)1\.945\(1\.2%↓\\downarrow\)\-0\.047\(\+0\.004\)0\.262\(0\.4%↑\\uparrow\)1\.2%Interrogative1\.402\(5\.2%↓\\downarrow\)0\.147\(4\.5%↓\\downarrow\)0\.226\(5\.8%↓\\downarrow\)1\.860\(5\.5%↓\\downarrow\)\-0\.028\(\+0\.023\)0\.268\(2\.7%↑\\uparrow\)5\.4%Imperative1\.445\(2\.3%↓\\downarrow\)0\.151\(1\.9%↓\\downarrow\)0\.234\(2\.5%↓\\downarrow\)1\.915\(2\.7%↓\\downarrow\)\-0\.041\(\+0\.010\)0\.264\(1\.1%↑\\uparrow\)2\.3%Exclamatory1\.395\(5\.7%↓\\downarrow\)0\.146\(5\.2%↓\\downarrow\)0\.224\(6\.7%↓\\downarrow\)1\.850\(6\.0%↓\\downarrow\)\-0\.025\(\+0\.026\)0\.269\(3\.1%↑\\uparrow\)5\.8%OMRA0\.452\(69\.4%↓\\downarrow\)0\.050\(67\.5%↓\\downarrow\)0\.038\(84\.2%↓\\downarrow\)0\.422\(78\.6%↓\\downarrow\)0\.225\(\+0\.276\)0\.358\(37\.2%↑\\uparrow\)74\.9%Table 10:Sentence\-Mood\-Controlled Direct Baselines on DeepSeek\.We constrain Direct generation to declarative, interrogative, imperative, and exclamatory tones\. Tone control yields only minor improvements; even the best variant reduces the gap by less than 6%, far below OMRA\. This suggests that OMRA’s gains are not explained by simple sentence mood or surface\-tone changes\.

### D\.5Additional Controls

##### Length control\.

We first examine whether OMRA’s improvement can be explained by shorter outputs\. As shown in Table[9](https://arxiv.org/html/2609.18282#A4.T9), length control improves both Direct and Few\-shot generation on DeepSeek\. Length\-Controlled Few\-shot reaches 57\.3% average gap reduction, compared with 49\.2% for Few\-shot\. However, OMRA achieves 74\.9% while producing answers of similar length\. This shows that response length contributes to the gap but does not explain OMRA’s gains\.

##### Sentence\-mood control\.

We next constrain Direct generation to declarative, interrogative, imperative, or exclamatory forms\. Table[10](https://arxiv.org/html/2609.18282#A4.T10)shows that these constraints yield only minor improvements\. Interrogative and exclamatory variants perform slightly better, but even the strongest variant reduces the gap by only 5\.8%, far below OMRA’s 74\.9%\. Thus, OMRA’s improvement is not explained by simple sentence\-mood changes\.

##### Platform\-aware prompting\.

We also test whether the original generation prompt is under\-specified by adding explicit platform context on DeepSeek\. As shown in Table[11](https://arxiv.org/html/2609.18282#A4.T11), platform\-aware prompting improves Direct from 0\.0% to 7\.6% average gap reduction and Few\-shot from 49\.2% to 52\.0%\. These gains remain substantially below OMRA’s 74\.9%, suggesting that missing platform context explains only a limited part of the AI preference–user engagement gap\.

##### Content preservation\.

Finally, we examine whether OMRA reduces the gap by substantially changing the original answer\. Using DeepSeek\-V3\.2 as an LLM judge, we compare 200 OMRA outputs with their original Synthetic drafts\. As shown in Table[12](https://arxiv.org/html/2609.18282#A4.T12), OMRA preserves stance in 93\.5% of examples, factual consistency in 90\.0%, answer relevance in 96\.0%, and usefulness in 91\.5%\. Unsupported additions occur in 6\.5% of examples, and major semantic changes occur in 4\.0%\. These results provide supporting evidence that OMRA usually preserves the original answer rather than relying on unrestricted rewriting\.

MethodMarginal Gap↓\\downarrowGeometry Gap↓\\downarrowTransfer to Real↑\\uparrowAvg\. Shrink↑\\uparrowWass\.JSMMDCent\.𝝆\\boldsymbol\{\\rho\}\(Δ\\Delta\)Top\-1Direct1\.4790\.1540\.2401\.969\-0\.0510\.261—Platform\-aware Direct1\.365\(7\.7%↓\\downarrow\)0\.145\(5\.8%↓\\downarrow\)0\.222\(7\.5%↓\\downarrow\)1\.805\(8\.3%↓\\downarrow\)\-0\.020\(\+0\.031\)0\.270\(3\.4%↑\\uparrow\)7\.6%Few\-shot0\.829\(43\.9%↓\\downarrow\)0\.085\(44\.8%↓\\downarrow\)0\.102\(57\.5%↓\\downarrow\)0\.971\(50\.7%↓\\downarrow\)0\.164\(\+0\.215\)0\.316\(21\.1%↑\\uparrow\)49\.2%Platform\-aware Few\-shot0\.805\(45\.6%↓\\downarrow\)0\.082\(46\.8%↓\\downarrow\)0\.097\(59\.6%↓\\downarrow\)0\.930\(52\.8%↓\\downarrow\)0\.171\(\+0\.222\)0\.319\(22\.2%↑\\uparrow\)52\.0%OMRA0\.452\(69\.4%↓\\downarrow\)0\.050\(67\.5%↓\\downarrow\)0\.038\(84\.2%↓\\downarrow\)0\.422\(78\.6%↓\\downarrow\)0\.225\(\+0\.276\)0\.358\(37\.2%↑\\uparrow\)74\.9%Table 11:Platform\-Aware Baselines on DeepSeek\.We add platform context to Direct and Few\-shot generation to test whether the AI preference–user engagement gap is mainly caused by an under\-specified generation prompt\. Platform\-aware prompting slightly improves both Direct and Few\-shot, but the gains remain small compared with OMRA\. This suggests that missing platform context is not the main driver of the gap\.Stance PreservedFactual ConsistencyAnswers QuestionUsefulness PreservedUnsupported AdditionsMajor Semantic Change93\.5%90\.0%96\.0%91\.5%6\.5%4\.0%Table 12:LLM\-as\-a\-Judge Preservation Check\.We use an LLM judge to compare OMRA outputs with their original synthetic drafts\. The results indicate that OMRA generally preserves stance, factual content, relevance, and usefulness, with low rates of unsupported additions and major semantic changes\.

## Appendix EHuman Evaluation

### E\.1Annotators

We recruit 30 annotators from a university, ranging from undergraduate to graduate students\. Annotators are regular users of open Q&A or discussion platforms and have sufficient language proficiency to judge the sampled answers\. They were compensated at a fair hourly rate consistent with local norms\.

### E\.2Annotation Protocol

Each annotation item presents one question and two candidate answers shown side by side\. One answer is produced by the target method \(Direct,Few\-shot, orOMRA\), and the other is a real platform answer from the same question context\. The presentation order is randomized at the item level to reduce position bias\. Annotators are not told which answer is AI\-generated\.

For each pair, annotators answer two questions:

- •Q1 \(Perceived human\-likeness\)\.*Which answer looks more like it was written by a real platform user?*
- •Q2 \(Pairwise preference\)\.*Which answer would you prefer to like or endorse?*

The two questions are shown on the same page, but annotators are instructed to treat them as separate judgments: Q1 concerns perceived human\-likeness, while Q2 concerns pairwise preference\.

![Refer to caption](https://arxiv.org/html/2609.18282v1/Figure/Figure_annotate.png)Figure 6:Annotation interface used in the human evaluation\.Annotators see the original question, two answer candidates in randomized order, and two questions on perceived human\-likeness and pairwise preference\.
### E\.3Annotation Interface

The annotation interface is shown in Figure[6](https://arxiv.org/html/2609.18282#A5.F6)\. Each item displays the original question at the top, two answer candidates in randomized left–right order, and the two judgment questions below\. Annotators may scroll within each answer panel if an answer exceeds the display height\. The interface hides platform metadata such as vote counts, timestamps, and author identity, preventing leakage of platform engagement signals\.

### E\.4Sample Size and Coverage

We evaluate three pairings against real platform answers:Direct vs Real,Few\-shot vs Real, andOMRA vs Real\. We sample 400 paired items in total: 134 for Direct vs Real, 134 for Few\-shot vs Real, and 132 for OMRA vs Real\. Each item is judged independently by all 30 annotators on both Q1 and Q2, yielding 12,000 item–annotator assignments and 24,000 binary judgments\. This full\-overlap design allows us to compute agreement statistics over the same item set\.

### E\.5Metrics

##### Win rates\.

For each pairingMMvs Real, we report the win rate of methodMMon Q1 \(perceived human\-likeness\) and Q2 \(pairwise preference\)\. Win rates are computed from majority judgments at the item level and then aggregated within each pairing\.

##### Inter\-annotator agreement\.

We report inter\-annotator agreement on Q2\. Since each item is labeled by 30 annotators, we compute pairwise Cohen’sκ\\kappafor all annotator pairs and report the average value within each pairing\.

##### Reversal rate\.

The reversal rate measures how often perceived human\-likeness and pairwise preference disagree on the same item\. For each itemii, letHi∈\{M,R\}H\_\{i\}\\in\\\{M,R\\\}be the majority answer to Q1 andPi∈\{M,R\}P\_\{i\}\\in\\\{M,R\\\}be the majority answer to Q2\. We define

Rev\.=1N∑i=1N𝟙\[Hi≠Pi\]\.\\mathrm\{Rev\.\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbb\{1\}\[H\_\{i\}\\neq P\_\{i\}\]\.A higher reversal rate indicates stronger divergence between perceived human\-likeness and pairwise preference judgment\. This metric tests whether pairwise preference judgment is reducible to perceived human\-likeness, or whether annotators may prefer an AI\-generated answer even when judging it as less human\-like\.

## Appendix FCase Study

Figure[7](https://arxiv.org/html/2609.18282#A6.F7)presents qualitative examples of the AI preference–user engagement gap\. These cases illustrate the diagnostic pattern in the main results rather than provide additional quantitative evidence: AI preference often favors explicit logical presentation, whereas real user engagement may be associated with affective and expressive salience\.

### F\.1Case A: Logic Is Not Sufficient for Popularity

Case A contrasts a highly voted real answer with a more formally structured answer to the same question\. The real answer is short, compressed, and rhetorically sharp, while the logic\-heavy answer contains explicit claim–evidence bridges, legal references, and extended justification\. Although such structure increases formal completeness, it does not necessarily match the patterns associated with real user engagement\.

This example illustrates logic overbinding: over\-explained spans may make an answer more explicit without increasing its engagement\. Within the same question, affective and expressive signals may be more salient than logical density alone\.

### F\.2Case B: Few\-Shot Imitation Can Drift Toward an AI High\-Score Style

Case B shows that few\-shot imitation does not always preserve the style of a high\-engagement real answer\. The real answer is brief, informal, and memorable, whereas the few\-shot outputs expand it into longer explanations with clearer structure, explicit transitions, and rubric\-like completeness\. They preserve parts of the topic and stance but move toward a standardized AI high\-score style\.

This helps explain why Few\-shot improves over Direct while remaining displaced in the joint preference space: it captures surface cues from the reference but may reintroduce over\-explained spans associated with logic overbinding\.

### F\.3Case C: AI Preference Mapping Fails on Real Answers

Case C illustrates a transfer error\. The real top\-voted answer gains engagement through concise expression and cultural resonance, but a predictor trained on Synthetic answers assigns it a relatively low score because it lacks formal evidence and explicit reasoning\( \)\. Conversely, the Synthetic answer receives a higher predicted score despite being less consistent with the observed platform ranking\.

This example reflects the transfer gap: an internally consistent AI preference rule may still fail on real platform answers\. OMRA targets this gap by reducing over\-explained spans while preserving stance and coherence\.

![Refer to caption](https://arxiv.org/html/2609.18282v1/Case_Study_figure.png)Figure 7:Qualitative case study of synthetic\-real preference mismatch\.Case A shows that formal logical structure is not sufficient for real platform popularity\. Case B shows that few\-shot imitation can expand short human\-viral answers into longer AI high\-score style outputs\. Case C shows that synthetic\-trained preference mapping can overvalue explicit reasoning while undervaluing expressive or socially resonant real answers\.

## Appendix GFull Prompts

Below are detailed prompts used in our evaluation\.

Prompt 1: Concept ExtractionSystem:You are a structured annotation assistant for online Q&A answers\. Your task is to extract ontology concepts from the given answer under a fixed three\-layer schema\. Extract only concepts that are explicitly grounded in the answer text\. Do not infer unsupported content\. If a concept type is absent, return an empty list\. Return valid JSON only\.Ontology Layers:•Logic: concepts related to stance, claims, evidence, examples, counterpoints, and reasoning units\.•Affect: concepts related to emotions, attitudes, values, group concerns, and social resonance\.•Expression: concepts related to framing, wording, rhetorical form, readability, compression, and memorable expressions\.Concept Types:Use the predefined concept inventory:𝒞L\\displaystyle\\mathcal\{C\}^\{L\}=\{MajorClaim,Claim,Evidence,…\},\\displaystyle=\\\{\\texttt\{MajorClaim\},\\texttt\{Claim\},\\texttt\{Evidence\},\\ldots\\\},𝒞A\\displaystyle\\mathcal\{C\}^\{A\}=\{Emotion,Value,GroupValue,…\},\\displaystyle=\\\{\\texttt\{Emotion\},\\texttt\{Value\},\\texttt\{GroupValue\},\\ldots\\\},𝒞E\\displaystyle\\mathcal\{C\}^\{E\}=\{Framing,RhetoricalDevice,Quote,…\}\.\\displaystyle=\\\{\\texttt\{Framing\},\\texttt\{RhetoricalDevice\},\\texttt\{Quote\},\\ldots\\\}\.For each concept, provide anid,type,text, and the shortest supportingspanfrom the answer\.Input:Question: \{question\}Answer: \{answer\}Output JSON format:``` { "logic_concepts": [ { "id": "L1", "type": "", "text": "", "span": "" } ], "affect_concepts": [ { "id": "A1", "type": "", "text": "", "span": "" } ], "expression_concepts": [ { "id": "E1", "type": "", "text": "", "span": "" } ] } ```

Prompt 2: Relation ExtractionSystem:You are a structured annotation assistant for ontology relation extraction\. Given the original answer and the extracted ontology concepts, your task is to identify relations among concept units\. Extract only relations that are directly grounded in the answer text\. Do not create relations based on external knowledge, plausibility, or general reasoning\. If no relation is supported, return an empty list\. Return valid JSON only\.Relation Types:Use the predefined relation inventory:ℛL=\{supports,justifies,contrasts,…\},\\mathcal\{R\}^\{L\}=\\\{\\texttt\{supports\},\\texttt\{justifies\},\\texttt\{contrasts\},\\ldots\\\},ℛA=\{evokes,activates,appeals\_to,…\},\\mathcal\{R\}^\{A\}=\\\{\\texttt\{evokes\},\\texttt\{activates\},\\texttt\{appeals\\\_to\},\\ldots\\\},ℛE=\{frames,highlights,compresses,…\}\.\\mathcal\{R\}^\{E\}=\\\{\\texttt\{frames\},\\texttt\{highlights\},\\texttt\{compresses\},\\ldots\\\}\.Relations may occur within a layer or across layers\. Use only relation types from the predefined inventory\. For each relation, provide the source concept id, target concept id, relation type, confidence score, and the shortest supporting span from the answer\.Input:Question: \{question\}Answer: \{answer\}Extracted concepts: \{concepts\_json\}Output JSON format:``` { "relations": [ { "source": "ID", "target": "ID", "relation": "", "confidence": 0.0, "span": "" } ] } ```

Prompt 3: Ontology\-Guided ScoringSystem:You are a structured scoring assistant\. Score the answer according to the provided ontology\-guided rubric\. The original answer text is the primary evidence; extracted ontology concepts and relations are structured cues\. Do not score general quality, politeness, factual correctness, or helpfulness unless they are explicitly part of the specified scoring dimension\. Return valid JSON only\.Scoring Principle:Each dimension should be scored independently on a 0–5 scale\.•0: absent or irrelevant\.•1: very weak\.•2: weak but present\.•3: moderate\.•4: strong\.•5: very strong and central to the answer\.

Scoring Dimensions:Use the predefined scoring dimensions:𝒮L=\{logic\_structure,logic\_evidence\_strength,…\},\\mathcal\{S\}^\{L\}=\\\{\\texttt\{logic\\\_structure\},\\texttt\{logic\\\_evidence\\\_strength\},\\ldots\\\},𝒮A=\{emotion\_intensity,value\_activation,…\},\\mathcal\{S\}^\{A\}=\\\{\\texttt\{emotion\\\_intensity\},\\texttt\{value\\\_activation\},\\ldots\\\},𝒮E=\{expression\_readability,expression\_quoteability,…\}\.\\mathcal\{S\}^\{E\}=\\\{\\texttt\{expression\\\_readability\},\\texttt\{expression\\\_quoteability\},\\ldots\\\}\.Score each dimension according to its own rubric\. Do not assume that higher logic, stronger emotion, or more elaborate expression automatically implies higher overall preference\.Input:Question: \{question\}Answer: \{answer\}Ontology concepts: \{concepts\_json\}Ontology relations: \{relations\_json\}Output JSON format:``` { "logic_scores": { "logic_structure": { "score": 0.0, "evidence": "" }, "logic_evidence_strength": { "score": 0.0, "evidence": "" } }, "affect_scores": { "emotion_intensity": { "score": 0.0, "evidence": "" }, "value_activation": { "score": 0.0, "evidence": "" } }, "expression_scores": { "expression_readability": { "score": 0.0, "evidence": "" }, "expression_quoteability": { "score": 0.0, "evidence": "" } } } ```

Prompt 4: Target\-Level Synthetic GenerationSystem:You are writing an answer for an open Q&A platform\. Generate one answer to the given question under the specified target virality level\. Match the language of the question\. Do not mention the target level, scoring rubric, ontology, or that the answer is generated\. Do not use Markdown headings or bullet lists unless they are natural for the answer\.

Target Virality Level:The target level represents the expected relative engagement of the answer within the same question context:•Level 0: bottom\-ranked answer\.•Level 1: low\-engagement answer\.•Level 2: popular answer\.•Level 3: highly viral answer\.Use your own understanding of platform preference to write an answer matching the requested level\. The answer should be plausible for the question context and should not explicitly explain why it matches the level\.Input:Question: \{question\}Target level: \{target\_level\}Output JSON format:``` { "answer": "" } ```

Prompt 5: Few\-Shot GenerationSystem:You are writing an answer for an open Q&A platform\. Generate one new answer to the given question under the specified target virality level\. A reference answer from the same question context and target level is provided\. Your task is to imitate its general style, tone, level of detail, and engagement pattern, but not its exact wording or content\.Imitation Requirement:Use the reference answer as a style and preference\-level example\. You may imitate its general writing style, such as its length, rhythm, directness, tone, and degree of emotional or expressive salience\. However, you must write a new answer\.•Do not copy the reference answer verbatim\.•Do not paraphrase the reference answer sentence by sentence\.•Do not reuse distinctive phrases, metaphors, jokes, or memorable lines from the reference\.•Do not mention the reference answer, target level, rubric, ontology, or generation process\.•Match the language of the input question\.•Do not use Markdown headings or bullet lists unless they are natural for the answer\.Target Virality Level:The target level represents the expected relative engagement of the answer within the same question context:•Level 0: buried answer\.•Level 1: low\-engagement answer\.•Level 2: hot answer\.•Level 3: viral answer\.Input:Question: \{question\}

Target level: \{target\_level\}Reference answer: \{reference\_answer\}Output JSON format:``` { "answer": "" } ```

Prompt 6: OMRA ReconstructionSystem:You are reconstructing an answer for an open Q&A platform\. The input contains an original synthetic draft and a list of ontology\-identified masked spans\. These spans correspond to surface reasoning scaffolds that may over\-externalize the answer’s reasoning structure\. Your task is to rewrite the draft by reconstructing or compressing these spans while preserving the answer’s main stance, factual content, and discourse coherence\.The goal is not to recover the original wording of the masked spans\. The goal is to produce a more natural answer with reduced over\-externalized reasoning\.Reconstruction Requirements:•Preserve the main stance of the original draft\.•Preserve factual content that is not part of the masked spans\.•Do not introduce unsupported claims or new factual details\.•Reconstruct or compress the masked spans rather than simply restoring them\.•Reduce unnecessary explicit reasoning markers, repeated justifications, and rubric\-like completeness\.•Maintain local and global coherence\.•Match the language of the input answer\.•Do not mention the ontology, masked spans, target level, rubric, or reconstruction process\.Preference Orientation:The reconstructed answer should remain aligned with the target level while avoiding synthetic\-style over\-explanation\. Use affective or expressive salience only when it is supported by the original draft and the surrounding context\. Do not make the answer artificially emotional, exaggerated, or stylistically over\-decorated\.Input:Question: \{question\}Original draft: \{original\_draft\}Masked spans: \{masked\_spans\_json\}Ontology concepts: \{concepts\_json\}Ontology relations: \{relations\_json\}Target level: \{target\_level\}Output JSON format:``` { "reconstructed_answer": "", "edit_summary": "" } ```

Prompt 7: OMRA RevisionSystem:You are the final verification and revision pass for OMRA\. Given a reconstructed answer, check whether it still contains synthetic\-style over\-explanation, excessive explicit reasoning scaffolds, or unnatural rubric\-like completeness\. If such patterns remain, apply a lightweight revision\. If no revision is needed, keep the answer unchanged\.Revision Requirements:•Preserve the main stance of the reconstructed answer\.•Do not introduce unsupported claims or new factual details\.•Reduce unnecessary explicit reasoning markers, repeated justifications, and overly complete step\-by\-step explanation\.•Keep the answer natural for the question context and target level\.•Do not make the answer artificially emotional, exaggerated, or stylistically over\-decorated\.•Match the language of the input answer\.•Do not mention OMRA, the rubric, the target level, or the revision process\.Input:Question: \{question\}Reconstructed answer: \{reconstructed\_answer\}Target level: \{target\_level\}Output JSON format:``` { "needs_revision": false, "final_answer": "", "revision_summary": "" } ```