REGARD: Regional Affective Differences in Large Language Models

# 1 Introduction
Source: [https://arxiv.org/html/2607.20722](https://arxiv.org/html/2607.20722)
REGARD: Regional Affective Differences in

Large Language Models

Andrei Chetvergov1,2, Alexander Evseev1,2, Mikhail Solovev1,2, Timofei Sivoraksha1,2

Stepan Ukolov1,2, Valeriia Kuschenko2, Maria Chistyakova2, Sergey Bolovtsov1,2

1Ivannikov Institute for System Programming of the Russian Academy of Sciences, Moscow, Russia 2Russian Presidential Academy of National Economy and Public Administration, Moscow, Russia

\{chetvergov\-as,aevseev\-23\-01,sivoraksha\-ta\}@ranepa\.ru \{ukolov\-sd,mchistyakova\-25,bolovtsov\-sv\}@ranepa\.ru \{msolovev\-24,vkuschenko\-22\}@edu\.ranepa\.ru

###### Abstract

LLMs trained and aligned within different linguistic and regional ecosystems may frame the same political, cultural, and geopolitical entities in different ways\. Such differences are often evaluated through sentiment, favorability, or stance, reducing model attitudes to a single positive–negative axis\. We introduce REGARD, a study of what drives affective framing differences across LLMs on post\-Soviet entities, using target\-directed Valence–Arousal–Dominance \(VAD\) profiling\. We query 19 models on 500 CIS\-specific targets, score responses with two independent LLM judges \(GPT\-4o\-mini and Qwen3\.6\-35B\-A3B\), and validate on a 300\-item human\-annotated subset\. Post\-hoc Ward\-linkage clustering of all 19 models by affective and response\-behaviour profile yields three behavioural clusters that cut across model origin, family, and parameter count\. Generic\-answer rate is strongly associated with lower arousal \(r=−0\.81r=\-0\.81\) and with cluster placement: models that deflect evaluative prompts with templated responses cluster together at low arousal regardless of origin\. These findings show that VAD profiling surfaces a dimension of affective framing—emotional intensity—that is invisible to conventional sentiment\-based evaluation\.

Keywords:large language models, affective framing, Valence–Arousal–Dominance, Russian\-language NLP, model evaluation, regional model comparison, post\-Soviet space

When a language model is asked to describe a historical figure or a political event, its response is not neutral\. Bias benchmarks and favorability studies consistently show that model outputs carry implicit evaluative stances – shaped by training data, RLHF objectives, and the cultural context of the organizations that built them\[[1](https://arxiv.org/html/2607.20722#bib.bib1),[15](https://arxiv.org/html/2607.20722#bib.bib15),[12](https://arxiv.org/html/2607.20722#bib.bib12)\]\. Models developed in different countries have been found to diverge in how they represent political actors and cultural objects, reflecting the ideological environment of their creators\[[19](https://arxiv.org/html/2607.20722#bib.bib19),[17](https://arxiv.org/html/2607.20722#bib.bib17)\]\. At the same time, most work on LLM opinion representation evaluates general or English\-centric topic sets\[[14](https://arxiv.org/html/2607.20722#bib.bib14),[3](https://arxiv.org/html/2607.20722#bib.bib3)\], without asking whether domestically developed models frame entities*specific to their own region*differently from non\-Russian ones\.

A parallel limitation concerns measurement\. Standard sentiment analysis reduces affect to a single polarity score – positive, negative, or neutral\[[6](https://arxiv.org/html/2607.20722#bib.bib6),[2](https://arxiv.org/html/2607.20722#bib.bib2)\]\. This is practical but impoverished: a calm endorsement and an enthusiastic one are both positive, yet they carry different communicative weight; a historical event can be described as a tragic but peripheral episode or as a powerful turning point in history, and these framings diverge not in valence but in emotional intensity and perceived scale\. The Valence\-Arousal\-Dominance \(VAD\) framework from affective psychology\[[13](https://arxiv.org/html/2607.20722#bib.bib13),[16](https://arxiv.org/html/2607.20722#bib.bib16),[9](https://arxiv.org/html/2607.20722#bib.bib9)\]decomposes affect into three independent dimensions – how positive or negative a portrayal is \(valence\), how emotionally intense it is \(arousal\), and how strongly or weakly the entity comes across \(dominance\) – and provides a natural instrument for measuring affective framing in free\-text LLM output, where polarity alone cannot distinguish these dimensions\.

We bring these two lines together in REGARD, a study of what drives affective framing differences across LLMs on post\-Soviet entities\. We query 19 models on 500 CIS\-specific targets, score responses with two independent VAD judges, and ask: what best explains the arousal and dominance variation we observe — model origin, model family, or model\-level response behaviour?

#### Research questions\.

RQ1:Do models differ systematically in how they frame CIS\-related entities along VAD dimensions, and along which axes?RQ2:What structure do these differences reveal when models are clustered post\-hoc by affective and response\-behaviour profile, without origin labels?RQ3:How are the resulting clusters related to model origin, architecture, and generic\-answer behaviour?

Figure[1](https://arxiv.org/html/2607.20722#S1.F1)summarizes the complete pipeline, from target\-bank construction to generation, VAD scoring, human validation, and model\-level analysis\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_study_overview.png)Figure 1:Overview of the REGARD study\. The release contains 500 targets, 19 generator models, and three Russian prompt variants, yielding 28,500 responses and 57,000 response–judge pairs\. Human validation comprises 900 ratings of 300 items by 15 annotators\.

## 2Related Work

Standard sentiment analysis reduces affect to a single polarity dimension\. Lexicon\- and corpus\-based affective computing scores text on continuous Valence\-Arousal\-Dominance axes\[[8](https://arxiv.org/html/2607.20722#bib.bib8),[9](https://arxiv.org/html/2607.20722#bib.bib9),[16](https://arxiv.org/html/2607.20722#bib.bib16)\]\. A parallel line of work elicits opinions from LLMs directly and compares the resulting distributions against human survey baselines\[[14](https://arxiv.org/html/2607.20722#bib.bib14),[3](https://arxiv.org/html/2607.20722#bib.bib3)\], finding systematic divergence sensitive to prompt language and framing\. LLMs encode political and cultural leanings traceable to pretraining\-corpus composition\[[4](https://arxiv.org/html/2607.20722#bib.bib4),[1](https://arxiv.org/html/2607.20722#bib.bib1)\], and models trained on English\-centric data misrepresent non\-Western cultural practices\[[11](https://arxiv.org/html/2607.20722#bib.bib11),[10](https://arxiv.org/html/2607.20722#bib.bib10)\]\. Geopolitical comparisons of US versus Chinese LLMs reveal divergent representations of the same political entities\[[19](https://arxiv.org/html/2607.20722#bib.bib19),[17](https://arxiv.org/html/2607.20722#bib.bib17)\]\. Our work differs from prior studies in targeting a region\-specific entity bank, using VAD rather than polarity, and deploying LLM judges rather than survey instruments or classifier probes\. Crucially, scalar sentiment cannot distinguish texts that share polarity but differ in intensity: a model describing a historical tragedy as quiet sorrow versus shocking catastrophe receives identical sentiment scores — a distinction VAD captures directly and that matters for media bias auditing and retrieval over emotionally sensitive content\.

## 3Data

We construct CIS\-Affective\-500, a set of 500 entities relevant to the broader CIS and post\-Soviet region, spanning six categories: persons, organizations, events, cultural symbols and places, social groups, and countries\. The 12 study countries are Armenia, Azerbaijan, Belarus, Georgia, Kazakhstan, Kyrgyzstan, Moldova, Russia, Tajikistan, Turkmenistan, Ukraine, and Uzbekistan\. Table[1](https://arxiv.org/html/2607.20722#S3.T1)gives the distribution across categories\. Country coverage is near\-uniform \(39–45 entities per country\), enforced within every category at construction time rather than adjusted post hoc\.

Table 1:Distribution of CIS\-Affective\-500 across categories\.All entities are anchored to Wikidata QIDs and drawn from two main sources\. The majority \(79%, 395 entities\) comes from Wikidata: persons, organizations, historical events, and cultural objects across all 12 study countries, with candidates ranked by Wikidata salience \(sitelink count\) and reviewed manually rather than admitted automatically\. The remaining 18\.6% \(93 entities\) comes from UNESCO Intangible Cultural Heritage and World Heritage lists, contributing epics, oral traditions, ritual practices, traditional cuisine, and architectural monuments specific to the region\. The 12 study countries form a fixed built\-in seed\. Together, the bank spans political figures, heads of state, and cultural icons underperson\(150 entities, 30%\); monuments, UNESCO\-listed sites, and cultural practices undercultural\_symbol\_or\_place\(110, 22%\); political parties, state bodies, universities, and media organizations underorganization\(80, 16%\); ethnic and demographic groups undersocial\_group\(78, 15\.6%\); wars, revolutions, protests, and national holidays underevent\(70, 14%\); and the 12 study countries themselves \(12, 2\.4%\)\.

Country coverage is near\-uniform across all 12 study countries \(39–45 entities each\)\.

#### Boundary cases\.

For 16 entities \(3\.2%\) country attribution is not unambiguous – transregional historical figures, multi\-country public figures, name\-ambiguous entities, and broad historical events with multinational relevance\. Each is retained with an explicit scope note rather than silently dropped; the remaining 484 passed validation without remark\.

#### Labeling\.

Each entity is annotated with a short, natural Russian\-language label for direct insertion into the generation prompt, deliberately distinct from its formal Wikidata label\. Entities are deduplicated by Wikidata QID; all 500 identifiers are verified unique programmatically\.

## 4Generation and Scoring

### 4\.1Models

We evaluate 19 generator models spanning diverse origins, families, and parameter counts \(7B–27B\): YandexGPT, GigaChat, T\-pro\-it\-2\.1, AVIBE, GLM\-4\.7, Gemma\-4\-26B, Qwen2\.5\-14B, Ministral\-3\-14B, Granite\-4\.1, Llama\-3\-8B, Qwen3\.6\-27B, Ministral\-8B, Mistral\-Nemo, Phi\-3\-Medium, Qwen2\.5\-Coder\-14B, SOLAR\-10\.7B, Mistral\-7B, Qwen2\.5\-7B, Granite3\.3\-8B\. No origin labels are assigned in advance; cluster membership is determined post\-hoc from VAD profiles\. All models are queried at temperature0\.70\.7, no system prompt, 512 output tokens maximum\. Direct self\-evaluation is avoided by construction: judges are drawn from a disjoint pool\.

### 4\.2Prompt

All models are prompted in Russian\. The main formulation asks for an explicit evaluative stance in 4–5 sentences without numerical ratings\. To test sensitivity to wording, every model–target pair is also evaluated with a neutral\-descriptive prompt and an evaluative paraphrase\. Figure[2](https://arxiv.org/html/2607.20722#S4.F2)reproduces all three prompts and their English translations\. Main results use the evaluative\-stance prompt; robustness across formulations is discussed in Section[9](https://arxiv.org/html/2607.20722#S9)\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/prompt_variants.png)Figure 2:Russian generation prompts used in REGARD, with English translations\. The same three formulations are applied to every model–target pair\.
### 4\.3Judge Contract and VAD Scale

We introduce the notion of a*judge contract*: a fixed, versioned prompt specification that defines what each VAD axis measures, how anchor points map to real\-world framing examples, and what quality flags the judge must emit alongside the numeric scores\. Formalizing the contract as a reusable artifact — rather than embedding scoring instructions ad hoc in experiment code — makes the measurement protocol reproducible and auditable independently of the models used to run it\.

Generations are scored by two independent judge models, Qwen3\.6\-35B\-A3B and GPT\-4o\-mini, neither of which appears in the generator list above — this avoids direct self\-evaluation, a known methodological weakness of LLM\-as\-judge designs in which a model scoring its own generations can inflate agreement or mask bias in its own output\. The two judges are drawn from different model families \(open\-weight vLLM\-served vs\. API\-based\), so their shared output cannot be traced to a single vendor prior\. Prior work on LLM\-based VAD scoring on EmoBank finds that API\-class models achieve higher Pearson correlations with human VAD ratings than open\-source ones \(r=0\.67r=0\.67for valence\), and that Valence is the most reliably captured dimension across model families\[[7](https://arxiv.org/html/2607.20722#bib.bib7)\]— consistent with our own cross\-judge agreement results in Section[6](https://arxiv.org/html/2607.20722#S6)\. We use two judges rather than one because\[[7](https://arxiv.org/html/2607.20722#bib.bib7)\]show that different zero\-shot model families exhibit qualitatively distinct failure modes on affective dimensions \(API models over\-predict negative emotions; open\-source models default to conservative neutral predictions\), making single\-judge scores systematically biased in different directions; averaging two judges from distinct families reduces this directional risk\.

All three VAD axes are scored on a unified\[0,1\]\[0,1\]scale, rather than treating valence as bipolar\[−1,1\]\[\-1,1\]while arousal and dominance remain unipolar\[0,1\]\[0,1\]: valence0\.50\.5denotes a neutral framing,0a maximally negative one, and11a maximally positive one; arousal and dominance retain their standard unipolar interpretation \(calm vs\. emotionally intense; powerless/passive vs\. agentic/influential\)\. This choice keeps the three axes on a common scale for downstream distance computations \(Section[5](https://arxiv.org/html/2607.20722#S5)\)\. Table[2](https://arxiv.org/html/2607.20722#S4.T2)makes the operational anchors explicit\.

Table 2:Operational anchors shared by both VAD judges\. Intermediate values are allowed continuously on\[0,1\]\[0,1\]\.#### Use of two judges\.

The primary profile, clustering, and category\-level results use Qwen3\.6\-35B\-A3B scores after applying thetarget\_coverage≥0\.5\\texttt\{target\\\_coverage\}\\geq 0\.5filter, as stated in the corresponding figure and table captions\. GPT\-4o\-mini scores every generation under the same contract and provides an independent replication of model ordering and group patterns in Section[6](https://arxiv.org/html/2607.20722#S6)\. Human\-validation correlations and the qualitative judge\-failure analysis use the mean of the two judges unless otherwise stated\.

## 5Results: VAD Profiles Across Models

Figure[3](https://arxiv.org/html/2607.20722#S5.F3)shows mean VAD scores for all 19 models, sorted by arousal\. Arousal is the primary axis of variation: it spans from0\.340\.34\(YandexGPT\) to0\.580\.58\(GLM\-4\.7\), a range of0\.240\.24units on the\[0,1\]\[0,1\]scale\. Valence, by contrast, is compressed into a narrow band \(0\.610\.61–0\.760\.76\) with no systematic ordering across clusters\. Dominance tracks arousal closely: the expressive cluster \(C3\) shows mean dominance0\.690\.69, the moderate cluster \(C2\)0\.670\.67, and the evasive cluster \(C1\)0\.650\.65\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_permodel_vad.png)Figure 3:Mean VAD scores across all 19 generator models \(evaluative\-stance prompt, Qwen3\.6\-35B\-A3B judge, 500 targets\), sorted by arousal\. Colour indicates cluster membership \(blue = C1 Evasive, green = C2 Moderate, magenta = C3 Expressive\)\. Arousal and dominance vary substantially across models; valence stays in a narrow band across all clusters\.Table[3](https://arxiv.org/html/2607.20722#S5.T3)reports per\-model VAD means and quality\-flag rates for a representative subset of models spanning all three clusters\. YandexGPT has the lowest arousal \(0\.340\.34\) and the highest generic\-answer rate \(55%55\\%\); T\-pro\-it\-2\.1 and GigaChat both reach arousal above0\.500\.50with generic rates below12%12\\%\. GLM and Ministral\-14B show the highest arousal \(0\.580\.58and0\.570\.57\) and near\-zero generic rates \(1%1\\%and0\.2%0\.2\\%\); Qwen2\.5\-14B sits at arousal0\.350\.35with a22%22\\%generic rate, landing in the moderate cluster despite coming from a different vendor family than YandexGPT\.

Table 3:Per\-model mean VAD scores and judge\-reported quality\-flag rates \(evaluative\-stance prompt, Qwen3\.6\-35B\-A3B judge;*Generic*=generic\_answer\_flag;*Cluster*= post\-hoc Ward cluster assignment\)\.#### Qualitative illustration\.

Figure[4](https://arxiv.org/html/2607.20722#S5.F4)shows responses from a C1 and a C3 model to the same target — the 1988 Spitak earthquake\. Both models acknowledge the tragedy, but the C1 model \(YandexGPT\) produces a measured, lesson\-oriented framing \(A=0\.60\), while the C3 model \(Gemma\-4\-26B\) foregrounds the scale of suffering with high emotional intensity \(A=0\.95\)\. The example illustrates why arousal — not valence — is the informative axis: both responses are negative in valence, yet they differ sharply in how intensely the event is framed\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/spitak_examples.png)Figure 4:Responses to the evaluative\-stance prompt about the 1988 Spitak earthquake\. C1 \(evasive\) and C3 \(expressive\) models frame the same target with sharply different arousal\.

## 6Judge Agreement

We check judge\-vs\-judge agreement on generations from all 19 models scored by both judges independently under all three prompt variants\. Pearson correlation between Qwen3\.6\-35B\-A3B and GPT\-4o\-mini is highest for valence \(r=0\.887r=0\.887\), followed by dominance \(r=0\.675r=0\.675\) and arousal \(r=0\.461r=0\.461\)\. Mean absolute error between the two judges per generation is0\.0940\.094for valence,0\.1780\.178for arousal, and0\.1460\.146for dominance\.

The low arousal agreement \(r=0\.461r=0\.461\) deserves direct attention, since arousal is also the axis on which the main cluster separation is observed\. Two observations mitigate but do not eliminate this concern\. First, the two judges are drawn from different model families \(open\-weight vLLM\-served vs\. API\-based\), so any shared directional bias would have to originate from a common pretraining signal rather than architectural coincidence; their disagreement on absolute scores is expected and does not imply that either judge is measuring noise\. Second, and more importantly, both judges independently reproduce the*same rank ordering*of models by mean arousal: Spearmanρ=0\.93\\rho=0\.93across all 19 models \(p<0\.001p<0\.001\)\. The absolute values diverge; the relative ordering — which is what the cluster analysis uses — does not\. Valence is the most reliably measured axis in absolute terms, but high measurement reliability and a robust directional signal are different properties: arousal has the clearer cluster separation despite lower point\-estimate agreement\.

#### Human validation protocol\.

We collected 900 human VAD ratings from 15 annotators, with exactly three ratings for each of 300 items\. Annotators saw a Russian target label and one generated response, but not the generator identity or automated judge scores\. They rated the framing on three continuous 0–10 sliders with verbal anchors for valence, arousal, and dominance \(Figure[5](https://arxiv.org/html/2607.20722#S6.F5), top\)\.

For the 300 items, the two\-judge mean correlates with mean human scores atr=0\.845r=0\.845for valence,0\.5650\.565for arousal, and0\.5130\.513for dominance \(allp<10−16p<10^\{\-16\}\); MAE is0\.1670\.167,0\.1110\.111, and0\.1640\.164, respectively\. Human Krippendorffα\\alphais0\.5270\.527,0\.2840\.284, and0\.2070\.207across the three axes\. The valence correlation exceeds ther≥0\.75r\\geq 0\.75benchmark from prior work\[[5](https://arxiv.org/html/2607.20722#bib.bib5),[7](https://arxiv.org/html/2607.20722#bib.bib7)\]; lower agreement on arousal and dominance reflects the greater subjectivity of intensity and agency judgements \(Figure[5](https://arxiv.org/html/2607.20722#S6.F5), bottom\)\.

\(a\) Human rating interface

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_annotation_interface.png)

\(b\) Judge\-human agreement

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_human_validation.png)

Figure 5:Human rating protocol and validation\. Top: annotators saw one Russian target and one model response, then used anchored 0–10 sliders for valence, arousal, and dominance; model identity and automated scores were hidden\. Bottom: two\-judge mean scores versus mean human scores for 300 items\. Dashed lines indicate perfect agreement and solid lines are least\-squares fits; panel titles report rounded Pearsonrrand human Krippendorffα\\alpha\.
#### Judge failure cases\.

Figure[6](https://arxiv.org/html/2607.20722#S6.F6)shows three items with\|Δ\|≥0\.25\|\\Delta\|\\geq 0\.25\. All three reflect the same pattern: LM judges latch onto evaluative surface markers rather than the experiential force of the content\. Importance markers inflate arousal; diplomatic framing inflates valence; a crisis\-framed political narrative suppresses dominance even when the described event was a forcible seizure of power\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/judge_failure_cases.png)Figure 6:LM\-judge versus human disagreements with\|Δ\|≥0\.25\|\\Delta\|\\geq 0\.25\. LM is the two\-judge mean; Human is the mean human rating\.
#### Score distributions and topic priors\.

Both judges produce unimodal distributions without clustering at round values \(no quantization\)\. Hallucination\-flagged generations showΔ​Val=−0\.12\\Delta\\text\{Val\}=\-0\.12,Δ​Dom=−0\.09\\Delta\\text\{Dom\}=\-0\.09, confirming sensitivity to text content\. Both judges independently recover the same group pattern, arguing against a single\-judge cultural prior\.

## 7Post\-Hoc Clustering of 19 Models

We cluster all 19 models by a four\-dimensional affective and response\-behaviour profile \(mean Valence, Arousal, Dominance, and generic\-answer rate\) using Ward\-linkage hierarchical clustering on the Euclidean distance between profile vectors\. Models span diverse families, sizes \(7B–27B\), and capability tiers; no origin labels are used at the clustering stage\.

A note on interpretation: generic\-answer rate is included as a clustering feature, so the observation that clusters differ in generic\-answer rate is partly built into the construction\. The substantive claim is different: within each cluster, and after excluding generic responses entirely, the arousal separation persists \(Δarousal=−0\.14\\Delta\_\{\\text\{arousal\}\}=\-0\.14between C1 and C3,n=500n=500targets\), showing that the cluster structure reflects genuine differences in affective framing intensity, not only differences in how often models deflect the prompt\.

#### Cluster structure\.

Figure[7](https://arxiv.org/html/2607.20722#S7.F7)shows the dendrogram\. At a cut ofk=3k=3clusters, the partition is:

- •Cluster 1 \(“Evasive”\): YandexGPT, Ministral\-8B, Qwen2\.5\-7B, Qwen2\.5\-Coder\-14B\.Mean arousal0\.360\.36; generic\-answer rate3030–55%55\\%\. The defining feature is not origin but response behaviour: all four produce high rates of generic or templated output, which the judge consistently scores at low arousal\.
- •Cluster 2 \(“Moderate”\): Qwen2\.5\-14B, AVIBE, Phi\-3\-Medium, Granite\-4\.1, Mistral\-Nemo, SOLAR\-10\.7B, Llama\-3\-8B, Granite3\.3\-8B\.Mean arousal0\.410\.41; generic\-answer rate88–22%22\\%\. The broadest cluster; models with average expressiveness and low refusal rates\.
- •Cluster 3 \(“Expressive”\): Mistral\-7B, GigaChat, Qwen3\.6\-27B, T\-pro\-it\-2\.1, Ministral\-14B, Gemma\-4\-26B, GLM\-4\.7\.Mean arousal0\.530\.53; generic\-answer rate<12%<12\\%\. High\-arousal expressive responses across all targets\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_cluster_dendrogram_19.png)Figure 7:Ward\-linkage dendrogram over 19 LLMs clustered by mean affective and response\-behaviour profile \(evaluative\-stance prompt,n=500n=500targets, Qwen3\.6\-35B\-A3B judge\)\. Leaf colours indicate cluster membership \(blue = C1, green = C2, magenta = C3\)\.![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_cluster_scatter_19.png)Figure 8:Mean valence vs\. arousal across 19 LLMs \(evaluative\-stance prompt\)\. Bubble size encodes generic\-answer rate\. Colour indicates cluster\.![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_cluster_bars_19.png)Figure 9:\(a\) Mean arousal and \(b\) generic\-answer rate across 19 LLMs, sorted by arousal\. Colour indicates cluster\.
#### Key observations\.

First, the cluster partition does not reproduce the Russian\-developed/non\-Russian split\. GigaChat and T\-pro\-it\-2\.1 \(Russian\-developed\) land in the high\-arousal Cluster 3 alongside GLM, Gemma, and Mistral families\. YandexGPT \(Russian\-developed\) is grouped with Ministral\-8B, Qwen2\.5\-7B, and Qwen2\.5\-Coder\-14B in Cluster 1 — models that, despite different origins, share a tendency to produce evasive, generic\-phrased responses under the evaluative prompt\. AVIBE \(Russian\-developed\) joins the moderate Cluster 2\.

Second, the primary axis of variation across models is*arousal*, with generic\-answer rate as a confounding moderator\. Arousal spans from0\.340\.34\(YandexGPT\) to0\.580\.58\(GLM\), while valence stays in a narrow band \(0\.610\.61–0\.760\.76\) with no systematic cluster ordering \(Figure[8](https://arxiv.org/html/2607.20722#S7.F8)\)\.

Third, YandexGPT’s low arousal is largely attributable to its high generic\-answer rate \(55%55\\%\), as the judge assigns low arousal to templated non\-committal responses\. The cluster analysis shows this is not a property specific to Russian\-developed models: Qwen2\.5\-7B and Qwen2\.5\-Coder\-14B, despite different origins, land in the same evasive cluster for the same reason\.

#### Additional findings\.

Beyond the cluster structure itself, the expanded panel reveals three further patterns\.

Arousal and generic\-answer rate are tightly coupled across all 19 models\.The Pearson correlation between mean arousal and generic\-answer rate isr=−0\.81r=\{\-0\.81\}\(p<0\.001p<0\.001,n=19n=19; 95% bootstrap CI\[−0\.91,−0\.71\]\[\-0\.91,\\,\-0\.71\]\)\. This is not a within\-cluster artifact: the relationship holds globally, spanning the full range from GLM \(A=0\.58A=0\.58, generic=1%=1\\%\) to YandexGPT \(A=0\.34A=0\.34, generic=55%=55\\%\)\. The evaluative\-stance prompt appears to trigger either committed framing \(low generic, high arousal\) or templated deflection \(high generic, low arousal\), with few models occupying a middle ground on both dimensions simultaneously\.

Dominance follows arousal, not valence\.Cluster 3 \(expressive\) shows the highest mean dominance \(0\.6940\.694\), followed by Cluster 2 \(moderate,0\.6700\.670\) and Cluster 1 \(evasive,0\.6510\.651\), a gradient of\+0\.043\+0\.043units from C1 to C3\. Valence, by contrast, is nearly flat across clusters \(C1:0\.7190\.719, C2:0\.7060\.706, C3:0\.7240\.724\) — a C3−\-C1 difference of only\+0\.005\+0\.005\. Cross\-model correlations confirm this: arousal and dominance are tightly linked \(r=0\.61r=0\.61across 19 models\), while arousal and valence are essentially unrelated \(r=0\.17r=0\.17\)\. Generic\-answer rate predicts arousal strongly \(r=−0\.81r=\-0\.81\) but dominance only moderately \(r=−0\.45r=\-0\.45\), and has no predictive relationship with valence \(r=\+0\.17r=\+0\.17\)\. The arousal–dominance co\-movement suggests that when a model avoids committing to a stance, it simultaneously reduces both the intensity and the assertiveness of its output, while the positive–negative register remains unchanged\.

Model family does not predict cluster\.The Qwen2\.5 family splits across all three clusters: Qwen2\.5\-Coder\-14B and Qwen2\.5\-7B are evasive \(C1\), Qwen2\.5\-14B is moderate \(C2\), and Qwen3\.6\-27B is expressive \(C3\)\. Similarly, the two Granite models land in C2 despite architectural similarity\. This suggests that instruction\-tuning decisions — specifically how the model handles evaluative or politically adjacent prompts — matter more for affective output than base architecture or parameter count\.

#### Cluster differences by target category\.

Figure[10](https://arxiv.org/html/2607.20722#S7.F10)shows all three VAD axes broken down by category and cluster\. The axes tell strikingly different stories\.

*Arousal*is the only axis with a consistent, large C3−\-C1 gap across all six categories \(range\+0\.123\+0\.123to\+0\.225\+0\.225\)\. Cultural objects and places show the largest separation \(\+0\.225\+0\.225\): expressive models respond to monuments, traditions, and UNESCO\-listed sites with substantially higher emotional intensity than evasive ones\. Persons and countries follow closely \(\+0\.211\+0\.211and\+0\.192\+0\.192\), while organisations and social groups show the smallest gap \(\+0\.123\+0\.123and\+0\.127\+0\.127\) — these targets are framed institutionally rather than emotionally\.

*Valence*shows no such ordering: the C3−\-C1 gap is near zero or weakly negative for persons \(−0\.039\-0\.039\), events \(−0\.039\-0\.039\), countries \(−0\.027\-0\.027\), and organisations \(−0\.025\-0\.025\), with only cultural symbols showing a modest positive gap \(\+0\.067\+0\.067\)\. Expressive models do not assign higher positive sentiment; they assign higher intensity\.

*Dominance*follows arousal in direction but with much smaller magnitude \(range−0\.013\-0\.013to\+0\.099\+0\.099\): cultural symbols again show the largest gap \(\+0\.099\+0\.099\), while events are near zero \(−0\.013\-0\.013\)\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_category_vad.png)Figure 10:Mean Valence, Arousal, and Dominance by target category and cluster \(evaluative\-stance prompt, 19 models, Qwen3\.6\-35B\-A3B judge\)\. Left column: grouped bars per cluster\. Right column: C3−\-C1 gap\. Arousal shows a large, consistent gap across all categories; valence shows near\-zero or negative gaps; dominance follows arousal at reduced magnitude\.
#### Cluster separation by country\.

Figure[11](https://arxiv.org/html/2607.20722#S7.F11)shows the same three\-axis breakdown by CIS country\.

*Arousal*gaps are large and positive for every country \(range\+0\.156\+0\.156to\+0\.217\+0\.217\): Georgia \(\+0\.217\+0\.217\), Armenia \(\+0\.216\+0\.216\), and Ukraine \(\+0\.206\+0\.206\) show the largest separation; Turkmenistan \(\+0\.156\+0\.156\) and Uzbekistan \(\+0\.157\+0\.157\) the smallest\. The countries with the largest cluster gap are precisely those with high geopolitical salience and contested representation — Georgia \(2008 war\), Armenia \(Nagorno\-Karabakh\), Ukraine \(ongoing conflict\)\.

*Valence*gaps are small and mixed: roughly half the countries show positive gaps and half negative \(range−0\.052\-0\.052to\+0\.036\+0\.036\), with no pattern matching the arousal ranking\. Russia, Ukraine, and Georgia — the three most politically salient countries — occupy opposite ends of the valence gap ranking, confirming that valence does not track contestedness\.

*Dominance*gaps are consistently positive but modest \(range\+0\.007\+0\.007to\+0\.076\+0\.076\): Armenia \(\+0\.076\+0\.076\) and Georgia \(\+0\.072\+0\.072\) lead, echoing the arousal pattern at roughly one\-third the magnitude\. Russia stands out on absolute dominance level \(C1:0\.7770\.777, C3:0\.8040\.804\) — substantially higher than any other country — reflecting the large share of high\-agency persons and institutions among Russian targets\.

![Refer to caption](https://arxiv.org/html/2607.20722v1/figures/fig_country_vad.png)Figure 11:Mean Valence, Arousal, and Dominance by CIS country and cluster \(evaluative\-stance prompt, 19 models, Qwen3\.6\-35B\-A3B judge\), sorted by C3−\-C1 gap\. Left column: dumbbell plot \(circle = C1/C3, diamond = C2\)\. Right column: C3−\-C1 gap\. Arousal shows uniform large gaps; valence is near\-zero and unsystematic; dominance mirrors arousal at reduced scale\.
#### Interpretation\.

The full VAD breakdown converges on a single conclusion: the cluster structure is carried by arousal and, secondarily, by dominance\. Valence — the axis captured by conventional sentiment analysis — is flat across clusters, countries, and categories alike \(r​\(generic,V\)=\+0\.17r\(\\text\{generic\},V\)=\+0\.17, not significant\)\. Models that deflect evaluative prompts do not produce more negative sentiment; they produce*less intense*framing\. This distinction is invisible to sentiment\-only measurement but directly observable in the arousal and dominance axes of the VAD framework\.

## 8Discussion

#### Generic\-answer rate as the strongest correlate\.

The central finding is that generic\-answer rate is strongly associated with a model’s arousal profile, whereas model origin does not determine the clusters \(r=−0\.81r=\-0\.81across 19 models\)\. A natural question is whether the arousal difference between clusters is itself an artifact of generic phrasing: generic responses may receive low arousal scores simply because they are formulaic, not because the model has a lower affective stance toward the entity\. To check this, we recompute cluster\-level arousal after excluding all generations flagged as generic: the C1–C3 gap persists \(Δarousal=−0\.14\\Delta\_\{\\text\{arousal\}\}=\-0\.14,n=500n=500targets\), confirming that the cluster separation is not driven by generic responses alone\. The arousal difference is present even within non\-generic outputs\.

#### Knowledge gap alternative\.

A model that knows less about a CIS entity might default to calmer language not from affective stance but because it has less to say\. Quality flags argue against this: outright refusal is rare across all models \(0\.2%0\.2\\%–0\.5%0\.5\\%\), and target coverage is high and comparable across clusters\. The Spitak earthquake example \(Figure[4](https://arxiv.org/html/2607.20722#S5.F4)\) is illustrative: YandexGPT \(C1\) produces a substantive, on\-topic response — it is not failing to recognise the target — yet frames the same event at arousal0\.600\.60while Gemma\-4\-26B \(C3\) reaches arousal0\.950\.95\. The difference is in framing intensity, not in knowledge coverage\.

#### What VAD surfaces that sentiment does not\.

Both responses in Figure[4](https://arxiv.org/html/2607.20722#S5.F4)are negative in valence\. A sentiment classifier would treat them as equivalent\. VAD profiling separates them on arousal — calm acknowledgement vs\. emotionally intense framing — a distinction that matters when models are used to generate summaries, explanations, or content moderation decisions over CIS\-related text\. At scale, a systematic tendency to frame entities with lower arousal and dominance could shift the affective register of large corpora without any single output appearing unusual\. Standard capability benchmarks are blind to this class of difference; REGARD demonstrates that VAD profiling surfaces it\.

#### Instruction tuning, not origin\.

The cluster analysis shows that the Qwen2\.5 family splits across all three clusters, and the two Russian\-developed models with the lowest generic rates \(GigaChat, T\-pro\-it\-2\.1\) land in the expressive cluster alongside GLM and Gemma\. This points to instruction\-tuning decisions — specifically, how a model is trained to respond to evaluative prompts about politically adjacent entities — as the proximate cause of the arousal variation, rather than any shared property of Russian\-developed models as a class\.

## 9Limitations

#### Cluster count\.

The number of clusters was set tok=3k=3by visual inspection of the Ward\-linkage dendrogram \(Figure[7](https://arxiv.org/html/2607.20722#S7.F7)\): the tree shows three clearly separated subtrees at a linkage distance of approximately0\.300\.30, with a large gap before the next merge\. All 19 models are analysed; no origin labels are used at the clustering stage\.

#### Prompt sensitivity\.

Results are reported on the evaluative\-stance prompt\. Under a neutral\-descriptive formulation the cluster separation in arousal narrows but remains present; under an evaluative paraphrase it weakens further\. The finding that arousal is the primary axis of variation is specific to prompts that invite an evaluative stance; it may not generalise to purely descriptive or factual generation settings\.

#### Arousal measurement\.

Judge\-vs\-judge agreement on arousal is lower than on valence \(r=0\.461r=0\.461vs\.r=0\.887r=0\.887\), reflecting the inherent subjectivity of intensity judgements\. Both judges preserve the same rank ordering of models \(Spearmanρ=0\.93\\rho=0\.93\), and human validation yieldsr=0\.565r=0\.565, above the noise floor but below the valence benchmark\. Absolute arousal values should be interpreted with this measurement uncertainty in mind; the cluster structure and rank ordering are more reliable than point estimates\.

#### Language confound\.

All generations are elicited in Russian\. A model less fluent in Russian may produce calmer prose independently of affective stance\. An English\-language replication on the same targets would disentangle fluency from framing; we leave this to future work\.

#### Scope of “Russian\-developed\.”

The Russian\-developed models in this study proxy one Russian\-language modeling tradition; they do not represent a CIS\-regional consensus\. Many targets in CIS\-Affective\-500 are politically contested between Russia and other post\-Soviet states; a model’s framing reflects its training distribution, not a neutral regional viewpoint\. Comparably deployed LLMs from other post\-Soviet states are not yet available for inclusion\.

#### Judge priors\.

Both judges are non\-CIS models; if they share a common affective framing of post\-Soviet entities, their consensus may systematically diverge from CIS\-audience perception\. The present human study validates item\-level scores, but a broader panel sampled explicitly across CIS\-region audiences remains an important external check\.

## 10Conclusion

We presented REGARD, an empirical study of affective framing in LLMs on CIS\-related entities, using a curated bank of 500 targets, 19 generator models, and a VAD scoring framework with two independent LLM judges\.

Post\-hoc Ward\-linkage clustering of all 19 models by affective and response\-behaviour profile reveals three behavioural clusters — Evasive, Moderate, and Expressive — that cut across model origin, family, and parameter count\. Generic\-answer rate is strongly associated with lower arousal \(r=−0\.81r=\-0\.81\) and with cluster placement, whereas regional origin does not determine the clusters: models that deflect evaluative prompts with templated responses cluster together at low arousal regardless of where they were developed\. The arousal gap is largest for culturally and politically salient targets \(persons, cultural objects, countries\) and smallest for institutional ones \(organizations, social groups\)\. Both judges independently preserve the same model rank ordering on arousal \(Spearmanρ=0\.93\\rho=0\.93\), and the cluster separation survives exclusion of generic responses\. VAD profiling surfaces a dimension of model behaviour — emotional intensity of framing — that is invisible to sentiment\-only measurement and not captured by standard capability benchmarks\.

## Disclosure of Interests

The authors declare no competing interests relevant to this work\.

## References

- \[1\]Buyl, M\., Rogiers, A\., Noels, S\., Bied, G\., Heiter, E\., Johary, I\., Mara, A\.C\., Romero, R\., Lijffijt, J\., De Bie, T\., Dominguez\-Catena, I\.: Large language models reflect the ideology of their creators\. npj Artificial Intelligence2, 7 \(2026\)
- \[2\]Dufraisse, E\., Popescu, A\., Tourille, J\., Brun, A\., Deshayes, J\.: MADTSC: A multilingual aligned news dataset for target\-dependent sentiment classification\. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)\. pp\. 8286–8305 \(2023\)
- \[3\]Durmus, E\., Lovitt, L\., Ganguli, D\., Schiefer, N\., Askell, A\., Bowman, S\.R\., Saunders, W\., Joseph, N\., Chen, T\., et al\.: Towards measuring the representation of subjective global opinions in language models \(2023\), arXiv preprint arXiv:2306\.16388
- \[4\]Feng, S\., Park, C\.Y\., Liu, Y\., Tsvetkov, Y\.: From pretraining data to language models to downstream tasks: Tracking the trails of political biases leading to unfair nlp models\. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(ACL 2023\) \(2023\)
- \[5\]Gilardi, F\., Alizadeh, M\., Kubli, M\.: ChatGPT outperforms crowd workers for text\-annotation tasks\. In: Proceedings of the National Academy of Sciences of the United States of America\. vol\. 120, p\. e2305016120 \(2023\)
- \[6\]Hamborg, F\., Donnay, K\.: NewsMTSC: A dataset for \(multi\-\)target\-dependent sentiment classification in political news articles\. In: Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics\. pp\. 1663–1675 \(2021\)
- \[7\]Inoshita, K\., Zhou, X\., Kawai, A\., Yada, K\.: LLMs capture emotion labels, not emotion uncertainty: Distributional analysis and calibration of human–LLM judgment gaps \(2026\), arXiv preprint arXiv:2604\.27345
- \[8\]Mehrabian, A\., Russell, J\.A\.: An Approach to Environmental Psychology\. MIT Press, Cambridge, MA \(1974\)
- \[9\]Mohammad, S\.M\.: Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words\. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(ACL 2018\)\. pp\. 174–184 \(2018\)
- \[10\]Myung, J\., Lee, N\., Zhou, Y\., Jin, J\., Putri, R\.A\., Antypas, D\., Borkan, H\., et al\.: Blend: A benchmark for llms on everyday knowledge in diverse cultures and languages\. In: Advances in Neural Information Processing Systems \(NeurIPS 2024\) \(2024\)
- \[11\]Naous, T\., Ryan, M\.J\., Ritter, A\., Xu, W\.: Having beer after prayer? measuring cultural bias in large language models\. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(ACL 2024\) \(2024\)
- \[12\]Röttger, P\., Hofmann, V\., Pyatkin, V\., Hinck, M\., Kirk, H\., Schuetze, H\., Hovy, D\.: Political compass or spinning arrow? towards more meaningful evaluations for values and opinions in large language models\. In: Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)\. pp\. 15295–15311 \(2024\)
- \[13\]Russell, J\.A\.: A circumplex model of affect\. Journal of Personality and Social Psychology39\(6\), 1161–1178 \(1980\)
- \[14\]Santurkar, S\., Durmus, E\., Ladhak, F\., Lee, C\., Liang, P\., Hashimoto, T\.: Whose opinions do language models reflect? In: Proceedings of the 40th International Conference on Machine Learning \(ICML 2023\) \(2023\)
- \[15\]Tao, Y\., Viberg, O\., Baker, R\.S\., Kizilcec, R\.F\.: Cultural bias and cultural alignment of large language models\. PNAS Nexus3\(9\), pgae346 \(2024\)
- \[16\]Warriner, A\.B\., Kuperman, V\., Brysbaert, M\.: Norms of valence, arousal, and dominance for 13,915 English lemmas\. Behavior Research Methods45\(4\), 1191–1207 \(2013\)
- \[17\]Yu, J\., et al\.: Made\-in china, thinking in america: U\.S\. values persist in chinese LLMs \(2025\), arXiv preprint arXiv:2512\.13723
- \[18\]Zheng, L\., Chiang, W\.L\., Sheng, Y\., Zhuang, S\., Wu, Z\., Zhuang, Y\., Lin, Z\., Li, Z\., Li, D\., Xing, E\.P\., Zhang, H\., Gonzalez, J\.E\., Stoica, I\.: Judging LLM\-as\-a\-judge with MT\-Bench and chatbot arena\. In: Advances in Neural Information Processing Systems \(NeurIPS 2023\) \(2023\)
- \[19\]Zheng, X\., et al\.: Echoes of power: Investigating geopolitical bias in US and China large language models\. Humanities and Social Sciences Communications \(2026\), nature