LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution

Adib Sakhawat; Hasan Mahmud; Md Kamrul Hasan; Syed Rifat Raiyan; Tahsin Islam; Takia Farhin

arxiv: 2606.28335 · v1 · pith:N4XRSLHZnew · submitted 2026-05-26 · 💻 cs.CY · cs.AI· cs.CL

LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution

Adib Sakhawat , Syed Rifat Raiyan , Tahsin Islam , Takia Farhin , Hasan Mahmud , Md Kamrul Hasan This is my paper

Pith reviewed 2026-06-30 10:48 UTC · model grok-4.3

classification 💻 cs.CY cs.AIcs.CL

keywords LLM political behaviorideological plasticitycontext-conditioned distributionVAA-CHES projectionOverton envelopepolitical dimensionsMTMM analysis

0 comments

The pith

LLM political ideology is a context-conditioned distribution over political space rather than a fixed point.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that an LLM's political positions cannot be captured by any single coordinate because they vary systematically with context. Using projection models on responses from nine LLMs across six contextual axes and three dimensions, it documents displacements of up to 0.57 units from persuasive framing and 0.52 units from under-represented languages. Chain-of-thought prompting tends to increase rather than reduce this instability. The models nevertheless remain inside a narrow overall range, roughly one-third the spread of major European parties. The central conclusion is that any useful summary must describe a shape, not a point.

Core claim

We argue, with systematic empirical evidence, that a large language model's political ideology is not a fixed point, but a conditional distribution P(position|context) over a real political space. We evaluate nine current LLMs using a unified measurement framework anchored by VAA-CHES projection models, which map responses onto three validated dimensions across six contextual axes. Our findings reveal high sensitivity to context: persuasive framing and under-represented languages displace coordinates by up to 0.57 and 0.52 units, respectively, while chain-of-thought reasoning often amplifies rather than dampens paraphrase instability. Despite this local plasticity, the model cohort occupies

What carries the argument

VAA-CHES projection models that map LLM responses onto the three dimensions lrgen, lrecon, and galtan across six contextual axes, yielding the distribution P(position|context) rather than a single coordinate.

If this is right

Persuasive framing displaces LLM coordinates by up to 0.57 units on the measured dimensions.
Under-represented languages displace coordinates by up to 0.52 units.
Chain-of-thought reasoning tends to increase paraphrase instability rather than reduce it.
The nine LLMs occupy a narrow Overton envelope roughly one-third the spread of major European parties.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Single-point probes of LLM politics will systematically miss the variability that appears once context is varied.
Any downstream application that treats an LLM as having a stable ideology will need to sample multiple contexts to capture the actual shape.
The narrow overall envelope suggests convergence on a limited band of positions even while local plasticity remains high.

Load-bearing premise

The VAA-CHES projection models remain valid and unbiased when applied to LLM-generated text across the six contextual axes and three dimensions.

What would settle it

Replicate the VAA-CHES mapping on the same nine models and check whether any contextual axis produces coordinate shifts at or above the reported 0.57 and 0.52 units.

Figures

Figures reproduced from arXiv: 2606.28335 by Adib Sakhawat, Hasan Mahmud, Md Kamrul Hasan, Syed Rifat Raiyan, Tahsin Islam, Takia Farhin.

**Figure 1.** Figure 1: Evidence of Algorithmic Monoculture. 3D contour plots in VAA-CHES space: Cyan for European Parliament parties, Magenta for the evaluated LLMs. The volumetric disparity visualizes the severe compression of the model’s ideological space vis-à-vis human political diversity. ditional distribution over a bounded ideological space I ⊂ Rd . Formally, the model’s stance is a random variable p ∈ I conditioned on … view at source ↗

**Figure 2.** Figure 2: Operational pipeline. The conditional distribution P(· | c) is gated by a JBS audit, decomposed into 6 contextual axes: Register (PSS), Paraphrase (PIS), Reasoning (RSS), Language (LDS), Multi-Turn (DS), Argument Role (IAS)—then projected via VAA→CHES onto ⟨lrgen, lrecon, galtan⟩, which feeds the aggregate OW and final score. instrument variation. We achieve this by projecting model stances into the thr… view at source ↗

**Figure 3.** Figure 3: Cross-metric Spearman correlations across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Temporal Stability of Algorithmic Monoculture. Top panel shows the multi-dimensional ideological projections (in the form of polytopes) across three electoral cycles, while the bottom panel provides a twodimensional cross-section of this geometry mapping the Economic Left/Right (lrecon) and GAL/TAN (galtan) dimensions, rescaled to [0, 1]. Across both views, the expansive outer envelope represents the ref… view at source ↗

**Figure 5.** Figure 5: Temporal Drift of Aggregate Model Ideologies. This figure tracks the temporal trajectory of each model’s mean ideological position across the evaluated electoral years, projected onto the Economic Left/Right (lrecon) and GAL/TAN (galtan) dimensions. The models are colorcoded as follows: deepseek_deepseek-v4-flash, google_gemini-2.5-flash-lite, google_gemma-4- 26b-a4b-it, ibm-granite_granite-3.3-8b-instru… view at source ↗

**Figure 6.** Figure 6: Paraphrase-conditioned vs. reasoning-conditioned stance landscape for Grok-4.1-Fast. Each cell encodes the model’s classified stance for one of the 82 VAA statements (horizontal axis) under one of the ten semantic paraphrases p1, . . . , p10 (vertical axis). Top: the forced-choice direct-elicitation condition (PIS); Bottom: the chain-of-thought condition (RSS). The inset above illustrates the two elicitat… view at source ↗

**Figure 7.** Figure 7: Cross-lingual stance landscape for Grok-4.1-Fast. Each cell encodes the model’s classified stance for one of the 82 VAA statements (horizontal axis, grouped by wave) under one of twelve elicitation languages (vertical axis): English (EN) baseline and eleven targets—German (DE), French (FR), Spanish (ES), Russian (RU), Turkish (TR), Arabic (AR), Hindi (HI), Bengali (BN), Mandarin Simplified (ZH), Indonesian… view at source ↗

**Figure 8.** Figure 8: Raw stance variation under contextual framing (PSS). The heatmaps display the categorical responses of the nine evaluated LLMs across the 82 VAA statements when subjected to four distinct prompt registers: Personal Blog (C1), Response to a Friend (C2), Persuasive Piece (C3), and the Neutral Baseline (C4). Vertical color variation within a single statement column reveals a model’s susceptibility to rhetori… view at source ↗

**Figure 9.** Figure 9: PSS — Prompt-induced ideological displacement. Per-model CHES projections under the four PSS framing conditions (C1–C4) across the three electoral cycles (2009, 2014, 2019). Each panel shows one of the nine evaluated LLMs. Colour encodes the framing condition; marker shape encodes the year. The dimmed background indicates the four political-compass quadrants. Cross-condition separation within a panel quant… view at source ↗

**Figure 10.** Figure 10: Raw stance variation under surface-form paraphrase (PIS). The heatmaps display the categorical responses of the nine evaluated LLMs across the 82 VAA statements when subjected to ten semantically equivalent rewording templates (p1–p10). Vertical color variation within a single statement column reveals a model’s susceptibility to lexical variation while holding propositional content fixed. The visual expl… view at source ↗

**Figure 11.** Figure 11: PIS — Paraphrase-induced ideological dispersion. Per-model CHES projections of the ten paraphrase realizations (p1–p10) for each year. Each subplot shows one of the nine evaluated LLMs; persona color distinguishes the ten paraphrase variants and marker shape distinguishes the year. The spread of each year-colored cluster quantifies the model’s susceptibility to lexical reformulation while holding proposi… view at source ↗

**Figure 12.** Figure 12: Raw stance variation under chain-of-thought reasoning (RSS). The heatmaps display the categorical responses of the nine evaluated LLMs across the 82 VAA statements when subjected to ten semantically equivalent, reasoning-anchored paraphrases (p1–p10). Vertical color variation within a single statement column reveals within-paraphrase instability induced by the generated reasoning trace. Corroborating the… view at source ↗

**Figure 13.** Figure 13: RSS — Reasoning-conditioned paraphrase dispersion. Per-model CHES projections under the chain-of-thought paraphrase condition. The visual encoding mirrors [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗

**Figure 14.** Figure 14: Raw stance variation under multilingual elicitation (LDS). The heatmaps display the categorical responses of the nine evaluated LLMs across the 82 VAA statements when queried in the English baseline versus eleven target languages. The high frequency of vertical variation within individual statement columns demonstrates that interaction language systematically alters a model’s fundamental policy position p… view at source ↗

**Figure 15.** Figure 15: LDS — Multilingual ideological displacement. Per-model CHES projections under the eleven target languages with arrows tracing the displacement from the English baseline to each target-language response. Color encodes the language; marker shape encodes the year. Long, parallel arrows within a panel reveal a coherent directional bias in the model’s multilingual ideological behavior; short or scattered arrow… view at source ↗

**Figure 16.** Figure 16: Raw stance variation under adversarial multi-turn debate (DS). The heatmaps display the categorical responses of the nine evaluated LLMs across the 82 VAA statements over eight continuous turns of adversarial dialogue (t1–t8). Vertical color variation within a single statement column illustrates turn-by-turn stance oscillation. Corroborating the spatial trajectories, models such as google_gemma-4-26b-a4… view at source ↗

**Figure 17.** Figure 17: DS — Adversarial multi-turn ideological trajectories. A 9×3 grid: rows correspond to the nine evaluated LLMs, columns to the three electoral cycles (2009, 2014, 2019). Each subplot traces the eight-turn (t1 → t8) ideological trajectory of the corresponding model–year configuration. Marker color encodes the turn index along a viridis ramp (dark → bright), marker size grows monotonically with the turn, and … view at source ↗

**Figure 18.** Figure 18: OW — Aggregate Overton envelope under pooled perturbations. Per-model CHES projections after pooling coordinates from PSS, PIS, RSS, LDS, and DS into a unified per-year cloud. The 90%-trimmed convex hull (shaded polygon) and its maximum-spread diameter (thick line segment) are drawn separately for each year. Hull area approximates the reachable ideological region under contextual variation; the diameter … view at source ↗

read the original abstract

We argue, with systematic empirical evidence, that a large language model's political ideology is not a fixed point, but a conditional distribution $\mathbb{P}($position$\mid$context$)$ over a real political space. We evaluate nine current LLMs using a unified measurement framework anchored by VAA-CHES projection models, which map responses onto three validated dimensions (lrgen, lrecon, galtan) across six contextual axes. Our findings reveal high sensitivity to context: persuasive framing and under-represented languages displace coordinates by up to 0.57 and 0.52 units, respectively, while chain-of-thought reasoning often amplifies rather than dampens paraphrase instability. Despite this local plasticity, the model cohort occupies a remarkably narrow Overton envelope overall, occupying roughly one-third the spread of major European parties. Supported by a multi-trait multi-method (MTMM) analysis, we conclude that a single point cannot summarize LLM political behavior; it must be characterized as a shape. Our code and data are publicly available at https://github.com/sakhadib/LLM-Ideoplasticity.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows LLMs shift political positions with context and stay in a narrow band overall, but all the displacement numbers rest on projecting LLM text through human-calibrated VAA-CHES models with no transfer check.

read the letter

The central point is that a single score misses how LLMs move on political dimensions when the prompt changes, and the authors give concrete numbers for those moves while noting the whole group stays compressed compared with real parties.

They treat ideology as P(position|context) and test nine models across six axes, projecting answers onto lrgen, lrecon, and galtan. The reported shifts reach 0.57 units for persuasive framing and 0.52 for low-resource languages, with chain-of-thought sometimes increasing rather than reducing variance. The MTMM analysis and public code are straightforward strengths; anyone can rerun the measurements.

The soft spot is the projection itself. VAA-CHES was fit on human survey language. LLM outputs differ in fluency, hedging, and lexical choices, yet the paper does not report an ablation or alignment test that would show the mapping stays unbiased or even monotonic under the new conditions. If the instrument compresses or distorts LLM text differently, both the size of the context effects and the narrow Overton envelope claim become harder to interpret.

This is a measurement study aimed at people who evaluate or deploy LLMs on political content. The reframing is useful even if the exact magnitudes need later correction. It is coherent on its own terms and the data pipeline looks reproducible, so it should go to peer review rather than a desk reject.

Referee Report

1 major / 1 minor

Summary. The paper claims that LLM political ideology is not a fixed point but a conditional distribution P(position|context) over a real political space. Using a unified framework anchored in VAA-CHES projection models, it evaluates nine LLMs across six contextual axes and three dimensions (lrgen, lrecon, galtan), reporting context-induced displacements of up to 0.57 and 0.52 units, amplified paraphrase instability under chain-of-thought, and a narrow Overton envelope occupying roughly one-third the spread of major European parties. An MTMM analysis is presented to support measurement reliability, leading to the conclusion that LLM political behavior must be characterized as a shape rather than a single point. Code and data are made publicly available.

Significance. If the central measurement assumptions hold, the work supplies a systematic, reproducible framework for quantifying ideological plasticity in LLMs and demonstrates that context can produce substantial, measurable shifts while the overall cohort remains narrowly distributed. The public release of code and data is a clear strength that enables direct verification and extension.

major comments (1)

[Methods (VAA-CHES projection)] Methods section (VAA-CHES projection application): the paper applies VAA-CHES models calibrated exclusively on human survey responses directly to LLM-generated text without any reported ablation, human-LLM alignment check, or domain-transfer validation. LLM outputs differ systematically from human language in fluency, hedging, and lexical distribution; if these differences distort or compress the projections, both the reported displacement magnitudes (0.57/0.52 units) and the narrow Overton envelope claim become unreliable. This assumption is load-bearing for every quantitative result and the 'shape vs point' conclusion.

minor comments (1)

[Abstract] Abstract: the phrase 'systematic empirical evidence' could be strengthened by briefly stating the number of models (nine) and contextual axes (six) to give readers an immediate sense of scale.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive review and for identifying a key methodological assumption. We respond to the major comment below and indicate planned revisions.

read point-by-point responses

Referee: Methods section (VAA-CHES projection application): the paper applies VAA-CHES models calibrated exclusively on human survey responses directly to LLM-generated text without any reported ablation, human-LLM alignment check, or domain-transfer validation. LLM outputs differ systematically from human language in fluency, hedging, and lexical distribution; if these differences distort or compress the projections, both the reported displacement magnitudes (0.57/0.52 units) and the narrow Overton envelope claim become unreliable. This assumption is load-bearing for every quantitative result and the 'shape vs point' conclusion.

Authors: We agree that the manuscript applies the VAA-CHES models, originally calibrated on human responses, to LLM-generated text without reporting dedicated ablation, human-LLM alignment, or domain-transfer validation studies. This constitutes a substantive and load-bearing assumption. The MTMM analysis provides evidence of internal measurement consistency across elicitation methods, but does not directly test cross-domain validity. We will revise the Methods and Limitations sections to explicitly discuss potential effects of linguistic differences on projection accuracy, the reported displacement magnitudes, and the Overton envelope width. We will also note that relative within-model displacements may remain informative under uniform projection bias, while absolute positioning claims require caution. Revision will be made. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims rest on direct empirical measurements

full rationale

The paper derives its central claim (LLM ideology as context-conditioned distribution rather than fixed point) from prompting nine LLMs across six contextual axes, obtaining responses, and projecting them onto lrgen/lrecon/galtan dimensions via pre-existing VAA-CHES models. The observed displacements (0.57/0.52 units), paraphrase instability, and narrow Overton envelope (~1/3 party spread) are reported outcomes of these measurements, not quantities fitted or defined in terms of themselves. No equations reduce a prediction to a fitted input by construction, no load-bearing self-citation chain is invoked, and the MTMM analysis is presented as supporting reliability of the method rather than substituting for external validation. The validity of VAA-CHES transfer to LLM text is an untested assumption (correctness risk), but does not create circularity in the reported derivation. Steps array left empty per rules for non-circular empirical work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the assumption that existing VAA-CHES models can be applied unchanged to LLM text and that the six contextual axes capture the relevant sources of plasticity.

axioms (1)

domain assumption The VAA-CHES projection models accurately map LLM responses onto the lrgen, lrecon, and galtan dimensions without systematic distortion from model-generated text.
The entire measurement framework is anchored by these models, which are treated as validated instruments.

pith-pipeline@v0.9.1-grok · 5757 in / 1330 out tokens · 43558 ms · 2026-06-30T10:48:25.217333+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

20 extracted references · 8 canonical work pages · 2 internal anchors

[1]

InFindings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 24767– 24773, Suzhou, China

POW: Political overton windows of large language mod- els. InFindings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 24767– 24773, Suzhou, China. Association for Computa- tional Linguistics. Ryan Bakker, Seth Jolly, and Jonathan Polk

2025
[2]

InProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), pages 14272–14284, Torino, Italia

SaGE: Evaluating moral consistency in large language models. InProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), pages 14272–14284, Torino, Italia. ELRA and ICCL. Tanise Ceron, Neele Falk, Ana Bari´c, Dmitry Nikolaev, and Sebastian Padó

2024
[3]

Junhyuk Choi, Sohhyung Park, Chanhee Cho, Hyeonchu Park, and Bugeun Kim

Uncovering political bias in large language models using parliamentary voting records.arXiv preprint arXiv:2601.08785. Junhyuk Choi, Sohhyung Park, Chanhee Cho, Hyeonchu Park, and Bugeun Kim

work page arXiv
[4]

Diagnosing the Reliability of LLM-as-a-Judge via Item Response Theory

Diag- nosing the reliability of llm-as-a-judge via item re- sponse theory.Preprint, arXiv:2602.00521. Mariella Faulborn, Dirk Hovy, and Indira Sen

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Shangbin Feng, Chan Young Park, Yuhan Liu, and Yu- lia Tsvetkov

A detailed factor analysis for the political compass test: Navigating ideologies of large language mod- els.arXiv preprint arXiv:2506.22493. Shangbin Feng, Chan Young Park, Yuhan Liu, and Yu- lia Tsvetkov

work page arXiv
[6]

Sasuke Fujimoto and Kazuhiro Takemoto

Personas with attitudes: Controlling LLMs for diverse data annotation.arXiv preprint arXiv:2410.11745. Sasuke Fujimoto and Kazuhiro Takemoto

work page arXiv
[7]

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.Preprint, arXiv:2507.06261. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Zhouchi Lin, Bowen Zhang, Lionel Ni, Wen Gao, Yuanzh...

work page internal anchor Pith review Pith/arXiv arXiv
[8]

The political ideology of conver- sational AI: Converging evidence on ChatGPT’s pro-environmental, left-libertarian orientation

The political ideology of conversa- tional AI: Converging evidence on ChatGPT’s pro- environmental, left-libertarian orientation.arXiv preprint, arXiv:2301.01768. Seth Jolly, Ryan Bakker, Liesbet Hooghe, Gary Marks, Jonathan Polk, Jan Rovny, Marco Steenbergen, and Milada Anna Vachudova

work page arXiv
[9]

Joseph G

Chapel Hill ex- pert survey trend file, 1999–2019.Electoral Studies, 75:102420. Joseph G. Lehman

1999
[10]

InFindings of the Association for Computational Linguistics: NAACL 2024, pages 2006–2017, Mexico City, Mex- ico

Large language models sensitivity to the order of options in multiple-choice questions. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 2006–2017, Mexico City, Mex- ico. Association for Computational Linguistics. Andres Reiljan, Frederico Ferreira da Silva, Lorenzo Cicchi, Diego Garzia, and Alexander H. Trech- sel

2024
[11]

Paul Röttger, Valentin Hofmann, Valentina Py- atkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy

Longitudinal dataset of political issue- positions of 411 parties across 28 european coun- tries (2009–2019) from voting advice applications EU Profiler and euandi.Data in Brief, 31:105968. Paul Röttger, Valentin Hofmann, Valentina Py- atkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy

2009
[12]

David Rozado

The 2024 Chapel Hill expert survey on political party position- ing in Europe: Twenty-five years of party positional data.Electoral Studies, 97:102981. David Rozado

2024
[13]

Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, We- icheng Ma, and Soroush V osoughi

Political alignment in large language models: A multidimensional audit of psychome- tric identity and behavioral bias.arXiv preprint arXiv:2601.06194. Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, We- icheng Ma, and Soroush V osoughi

work page arXiv
[14]

arXiv preprint arXiv:2406.07791 (2024)

Judging the judges: A systematic study of position bias in llm-as-a-judge.Preprint, arXiv:2406.07791. Frances Stewart

work page arXiv
[15]

Let c0 denote the neutral baseline (C4) andC= {C1, C2, C3}the perturbed conditions

Expanded component-wise: d(u,v) = h (ulrgen −v lrgen)2 + (ulrecon −v lrecon)2 + (ugaltan −v galtan)2 i1/2 (2) A.1 Prompt Sensitivity Score (PSS) PSS quantifies the ideological displacement in- duced by alternative prompt framings. Let c0 denote the neutral baseline (C4) andC= {C1, C2, C3}the perturbed conditions. For modelm, statements, and conditionc∈ C:...

1996
[16]

The pipeline enforces a strict complete-case strategy, dropping observations with missing targets (lr- gen, lrecon, galtan)

to preserve temporal variation in political issue salience. The pipeline enforces a strict complete-case strategy, dropping observations with missing targets (lr- gen, lrecon, galtan). After removing metadata (CHESS, YEAR) and the three targets, the effec- tive dimensions are: •2009:153parties,30features (from153× 35). •2014:141parties,30features (from141...

2009
[17]

Persua- sive Piece

Statistic 2D PSS 3D PSS Observed Displacement 0.5721 0.7272 95% CI Lower 0.5721 0.7272 95% CI Upper 0.5721 0.7272 Table 6: Non-parametric bootstrap (N= 10,000) con- fidence intervals for the maximum observed displace- ment (Gemma-4-26Bunder C1 in 2019). The zero- width intervals reflect strict within-condition determin- ism. The zero-width confidence inte...

2019
[18]

for”vs.“against

designates this as an exploratory fac- tor analysis rather than a confirmatory measure- ment model, the resulting eigendecomposition re- veals a highly structured and theoretically coher- ent latent topology. We extracted three princi- pal components (PCs) which together account for 83.30%of the total variance across the measure- ment framework. As detail...

2009
[19]

Each V AA round con- sists of approximately 22–30 policy statements to which respondents (and parties) express agree- ment on a five-point Likert scale, ranging from completely disagreetocompletely agree. To facilitate cross-year comparison and to make the thematic structure of the issue space transparent, we group statements into seven recurring policy d...

2009
[20]

ideological trajectory of the corresponding model–year configuration. Marker color encodes the turn index along a viridis ramp (dark→bright), marker size grows monotonically with the turn, and arrows connect consecutive turns to reveal directional structure. Trajectory length, tortuosity, and net drift are visually distinct. 36 0.00 0.25 0.50 0.75 1.00gal...

2009

[1] [1]

InFindings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 24767– 24773, Suzhou, China

POW: Political overton windows of large language mod- els. InFindings of the Association for Compu- tational Linguistics: EMNLP 2025, pages 24767– 24773, Suzhou, China. Association for Computa- tional Linguistics. Ryan Bakker, Seth Jolly, and Jonathan Polk

2025

[2] [2]

InProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), pages 14272–14284, Torino, Italia

SaGE: Evaluating moral consistency in large language models. InProceedings of the 2024 Joint International Conference on Computa- tional Linguistics, Language Resources and Evalu- ation (LREC-COLING 2024), pages 14272–14284, Torino, Italia. ELRA and ICCL. Tanise Ceron, Neele Falk, Ana Bari´c, Dmitry Nikolaev, and Sebastian Padó

2024

[3] [3]

Junhyuk Choi, Sohhyung Park, Chanhee Cho, Hyeonchu Park, and Bugeun Kim

Uncovering political bias in large language models using parliamentary voting records.arXiv preprint arXiv:2601.08785. Junhyuk Choi, Sohhyung Park, Chanhee Cho, Hyeonchu Park, and Bugeun Kim

work page arXiv

[4] [4]

Diagnosing the Reliability of LLM-as-a-Judge via Item Response Theory

Diag- nosing the reliability of llm-as-a-judge via item re- sponse theory.Preprint, arXiv:2602.00521. Mariella Faulborn, Dirk Hovy, and Indira Sen

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Shangbin Feng, Chan Young Park, Yuhan Liu, and Yu- lia Tsvetkov

A detailed factor analysis for the political compass test: Navigating ideologies of large language mod- els.arXiv preprint arXiv:2506.22493. Shangbin Feng, Chan Young Park, Yuhan Liu, and Yu- lia Tsvetkov

work page arXiv

[6] [6]

Sasuke Fujimoto and Kazuhiro Takemoto

Personas with attitudes: Controlling LLMs for diverse data annotation.arXiv preprint arXiv:2410.11745. Sasuke Fujimoto and Kazuhiro Takemoto

work page arXiv

[7] [7]

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.Preprint, arXiv:2507.06261. Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Zhouchi Lin, Bowen Zhang, Lionel Ni, Wen Gao, Yuanzh...

work page internal anchor Pith review Pith/arXiv arXiv

[8] [8]

The political ideology of conver- sational AI: Converging evidence on ChatGPT’s pro-environmental, left-libertarian orientation

The political ideology of conversa- tional AI: Converging evidence on ChatGPT’s pro- environmental, left-libertarian orientation.arXiv preprint, arXiv:2301.01768. Seth Jolly, Ryan Bakker, Liesbet Hooghe, Gary Marks, Jonathan Polk, Jan Rovny, Marco Steenbergen, and Milada Anna Vachudova

work page arXiv

[9] [9]

Joseph G

Chapel Hill ex- pert survey trend file, 1999–2019.Electoral Studies, 75:102420. Joseph G. Lehman

1999

[10] [10]

InFindings of the Association for Computational Linguistics: NAACL 2024, pages 2006–2017, Mexico City, Mex- ico

Large language models sensitivity to the order of options in multiple-choice questions. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 2006–2017, Mexico City, Mex- ico. Association for Computational Linguistics. Andres Reiljan, Frederico Ferreira da Silva, Lorenzo Cicchi, Diego Garzia, and Alexander H. Trech- sel

2024

[11] [11]

Paul Röttger, Valentin Hofmann, Valentina Py- atkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy

Longitudinal dataset of political issue- positions of 411 parties across 28 european coun- tries (2009–2019) from voting advice applications EU Profiler and euandi.Data in Brief, 31:105968. Paul Röttger, Valentin Hofmann, Valentina Py- atkin, Musashi Hinck, Hannah Rose Kirk, Hinrich Schütze, and Dirk Hovy

2009

[12] [12]

David Rozado

The 2024 Chapel Hill expert survey on political party position- ing in Europe: Twenty-five years of party positional data.Electoral Studies, 97:102981. David Rozado

2024

[13] [13]

Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, We- icheng Ma, and Soroush V osoughi

Political alignment in large language models: A multidimensional audit of psychome- tric identity and behavioral bias.arXiv preprint arXiv:2601.06194. Lin Shi, Chiyu Ma, Wenhua Liang, Xingjian Diao, We- icheng Ma, and Soroush V osoughi

work page arXiv

[14] [14]

arXiv preprint arXiv:2406.07791 (2024)

Judging the judges: A systematic study of position bias in llm-as-a-judge.Preprint, arXiv:2406.07791. Frances Stewart

work page arXiv

[15] [15]

Let c0 denote the neutral baseline (C4) andC= {C1, C2, C3}the perturbed conditions

Expanded component-wise: d(u,v) = h (ulrgen −v lrgen)2 + (ulrecon −v lrecon)2 + (ugaltan −v galtan)2 i1/2 (2) A.1 Prompt Sensitivity Score (PSS) PSS quantifies the ideological displacement in- duced by alternative prompt framings. Let c0 denote the neutral baseline (C4) andC= {C1, C2, C3}the perturbed conditions. For modelm, statements, and conditionc∈ C:...

1996

[16] [16]

The pipeline enforces a strict complete-case strategy, dropping observations with missing targets (lr- gen, lrecon, galtan)

to preserve temporal variation in political issue salience. The pipeline enforces a strict complete-case strategy, dropping observations with missing targets (lr- gen, lrecon, galtan). After removing metadata (CHESS, YEAR) and the three targets, the effec- tive dimensions are: •2009:153parties,30features (from153× 35). •2014:141parties,30features (from141...

2009

[17] [17]

Persua- sive Piece

Statistic 2D PSS 3D PSS Observed Displacement 0.5721 0.7272 95% CI Lower 0.5721 0.7272 95% CI Upper 0.5721 0.7272 Table 6: Non-parametric bootstrap (N= 10,000) con- fidence intervals for the maximum observed displace- ment (Gemma-4-26Bunder C1 in 2019). The zero- width intervals reflect strict within-condition determin- ism. The zero-width confidence inte...

2019

[18] [18]

for”vs.“against

designates this as an exploratory fac- tor analysis rather than a confirmatory measure- ment model, the resulting eigendecomposition re- veals a highly structured and theoretically coher- ent latent topology. We extracted three princi- pal components (PCs) which together account for 83.30%of the total variance across the measure- ment framework. As detail...

2009

[19] [19]

Each V AA round con- sists of approximately 22–30 policy statements to which respondents (and parties) express agree- ment on a five-point Likert scale, ranging from completely disagreetocompletely agree. To facilitate cross-year comparison and to make the thematic structure of the issue space transparent, we group statements into seven recurring policy d...

2009

[20] [20]

ideological trajectory of the corresponding model–year configuration. Marker color encodes the turn index along a viridis ramp (dark→bright), marker size grows monotonically with the turn, and arrows connect consecutive turns to reveal directional structure. Trajectory length, tortuosity, and net drift are visually distinct. 36 0.00 0.25 0.50 0.75 1.00gal...

2009