Political space, VAA-CHES, and LLM ideology

Political competition is often described using words such as left, right, liberal, conservative, green, nationalist, pro-EU, or Eurosceptic. Those labels are useful, but they compress many issue positions into a few categories. A political coordinate system makes the compression explicit. Each actor receives numeric positions on dimensions that summarize recurring patterns in policy conflict.

In European party politics, three dimensions are especially useful for this project. lrgen captures general left-right orientation. lrecon captures economic left-right conflict, such as redistribution, public spending, market regulation, and welfare. galtan captures the cultural dimension between Green/Alternative/Libertarian and Traditional/Authoritarian/Nationalist politics.

A coordinate is not a moral judgment and not an essence. It is an instrument reading. If two actors land near each other, they are similar with respect to the instrument's dimensions and data sources. If they are far apart, they differ in the patterns the instrument has learned to represent.

A Voting Advice Application, or VAA, is an online tool that asks users to respond to political issue statements. The application compares those answers with party or candidate positions on the same statements, then presents proximity scores, rankings, or spatial maps. VAAs became widespread in Europe because multiparty systems and coalition politics make party choice harder to reduce to a binary left-right decision.

EU Profiler and euandi are flagship transnational VAAs for European Parliament elections. They use policy statements that are meant to be salient, divisive, and comparable across countries. Users usually answer on a Likert-style agreement scale: completely disagree, disagree, neutral, agree, completely agree, often with a no-opinion option in the public application.

For this project, the VAA side matters because it gives a structured prompt interface. A human, a party, or an LLM can answer the same statements. Those answers become numeric features, using a five-point coding:

• CD = 0.00, completely disagree
• D = 0.25, disagree
• N = 0.50, neutral
• A = 0.75, agree
• CA = 1.00, completely agree

The Chapel Hill Expert Surveys estimate party positions in Europe using expert judgments. Across repeated waves, CHES supplies comparable measurements for party families, countries, and policy dimensions. This makes it a research infrastructure: it is not a one-off poll, but a repeated map of elite-level party positioning.

CHES matters here because it provides the target coordinate system. The transformation models learn how VAA statement answers correspond to CHES party coordinates. The output coordinates are rescaled to the unit interval [0,1] for visual and computational consistency.

The paper's central claim is that an LLM's political position should be modeled as a context-conditioned distribution, not as one fixed point. The VAA-CHES instrument makes that claim measurable. A model can answer VAA statements under different prompt registers, paraphrases, languages, reasoning conditions, debates, or argumentative roles. Each answer set becomes a coordinate.

Once many coordinates are collected, the question changes. Instead of asking "where is this model politically?", the study asks: how far does it move under context, how stable is its projection, and how large is its aggregate envelope compared with European party diversity?

This is why the project can report both local plasticity and global compression. Local context can move model coordinates substantially. Yet the pooled model cohort can still occupy a small ideological volume relative to the European party reference set.

The repository releases three serialized transformation models: ideology_model_2009.pkl, ideology_model_2014.pkl, and ideology_model_2019.pkl. Each is a complete scikit-learn pipeline: preprocessing and prediction are stored together, so users should not manually normalize inputs before inference.

The training data are cleaned by dropping observations where any of the three CHES targets are missing. The feature matrix contains the remaining numeric VAA features after removing identifiers such as CHESS, YEAR, and the target variables. The target vector is three-dimensional: lrgen, lrecon, and galtan.

From a computer-science perspective, the VAA-CHES instrument is a representation pipeline. Natural-language political statements are converted into categorical stances. Categorical stances become numeric features. Numeric features are transformed by a learned model into coordinates. Coordinates are then used to study geometry: displacement, dispersion, hull volume, centroid distance, and temporal drift.

LLMs sit at the unusual middle of this pipeline. They are language systems, so they can produce natural-language justifications. But they can also be forced into structured stance choices, making their outputs comparable to survey responses. This means political behavior can be measured as a distribution over elicitation contexts rather than as a single static survey answer.

The bridge is therefore bidirectional. Political science supplies theory, dimensions, and reference data. Computer science supplies model instrumentation, prompt perturbation, representation learning, and geometric analysis. LLMs become both subjects of measurement and interfaces for translating language into structured political signals.