Research map / Large language models / Alignment and preferences
Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Uses representativeness-heuristic and kernel-of-truth framings to show LLMs exaggerate Republican and Democratic party positions on ANES and MFQ survey items more than humans do, and tests prompt-based mitigations.
Problem
LLMs' deviations from empirical political positions, i.e., how and when they exaggerate party stances and produce political stereotypes, remain under-examined.
Approach
Applies cognitive-science representativeness heuristics and the kernel-of-truth assumption, formalized via likelihood ratios and a dual empirical/prediction question design, to quantify LLM exaggeration of partisan positions and tests awareness, feedback, and reasoning prompts as mitigations.
Key results
- LLM predictions for party positions exceed empirical mean differences more than human predictions do across evaluated models on ANES.
- Kernel-of-truth gamma is positive for most models on ANES and MFQ, with Llama3-8b a notable exception (gamma -0.33 on MFQ).
- Prompt mitigations lower kappa relative to baseline: Reasoning gave the largest reduction on ANES and Feedback on MFQ, though effects vary by model and task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| LLMs predict more polarized party positions than empirical survey means and human predictions | Evaluated language models · ANES | party-position differences | Hurt |
| Add prompt-based mitigation (awareness, reasoning, or feedback) (ablation) | LLMs · ANES | κ (stereotyping-risk proxy): Llama2-70b: 83.34 (33.26) -> 22.17 (14.79); GPT-3.5: 68.99 (27.04) -> 19.21 (8.78); GPT-4: 114.72 (40.15) -> 25.37 (5.05 | Helped |
| Add prompt-based mitigation (awareness, reasoning, or feedback) (ablation) | LLMs · MFQ | κ (stereotyping-risk proxy): Llama2-70b: 191.55 (117.72) -> 39.89 (34.22); GPT-3.5: 71.73 (45.69) -> 29.42 (17.39); GPT-4: 157.41 (115.91) -> 18.48 ( | Helped |
| Add speaker information to the statement prompt (ablation) | GPT-4 and Gemini · Misinformation detection | accuracy: GPT-4: 0.677 -> 0.706; Gemini: 0.622 -> 0.632 | Helped |
| Add party affiliation information to the statement prompt (ablation) | Llama2-70B and GPT-3.5 · Misinformation detection | accuracy: Llama2-70b Democrats: 0.482 -> 0.739; Republicans: 0.602 -> 0.571; GPT-3.5 overall: 0.645 -> 0.597 | Mixed |
| Increase temperature (ablation) | GPT-4 · ANES | coefficient of variation: 0.00 -> 0.11 | Hurt |
| Increase temperature (ablation) | GPT-4 · ANES | mean deviation from empirical means: Diff_D: -1.51, -1.46, -1.4, -1.4; Diff_R: 0.48, 0.46, 0.49, 0.47 | No clear effect |
| Use party affiliation information in misinformation prompts (ablation) | Evaluated models · Misinformation detection | accuracy | Mixed |
| LLM predictions reflect empirical party patterns, with Llama3-8B as an exception on MFQ | Evaluated language models · ANES and MFQ | γ (kernel-of-truth coefficient): Llama3-8b MFQ: -0.33 (2.18) | Mixed |
| Representative heuristics influence generated predictions | Evaluated language models · ANES and MFQ | ε (representative-heuristic coefficient) | Mixed |
| Use an attribute that is highly representative but differs from the most probable attribute | Evaluated language models · ANES: Defense Spending vs. Liberal-Conservative | predicted-minus-empirical mean difference: maximum mean difference 0.89 -> 2.01 | Helped |
Surprising findings
- Llama3-8b shows negative deviations on several moral foundations, diverging from the general exaggeration pattern.
- Including party affiliation in misinformation detection does not reliably improve accuracy and shifts accuracy disparity between Democrat and Republican statements.
Limitations
- Limited to Republicans and Democrats in the US.
- Party affiliation is assumed to indicate collective ideology, ignoring individual variation.
- Empirical data from survey samples may not represent the broader population.
Experimental setup
Datasets: ANES Moral Foundations Questionnaire dataset (N=919) fake news detection benchmark (LIAR-style, speaker party labels)
Benchmarks: ANES MFQ
Metrics: predicted mean vs empirical mean difference kernel-of-truth gamma representativeness epsilon kappa stereotyping index coefficient of variation accuracy and false positive rate on misinformation detection
Baselines: Human Pred Empirical
Builds on: representativeness heuristic (Kahneman and Tversky) kernel-of-truth hypothesis
Closest papers by meaning
- Linear Representations of Political Perspective Emerge in Large Language Models (ICLR 2025 · similarity 0.89)
- Bias and Volatility: A Statistical Framework for Evaluating Large Language Model's Stereotypes and the Associated Generation Inconsistency (NeurIPS 2024 · similarity 0.88)
- Distributive Fairness in Large Language Models: Evaluating Alignment with Human Values (NeurIPS 2025 · similarity 0.87)
- Questioning the Survey Responses of Large Language Models (NeurIPS 2024 · similarity 0.87)
- The PRISM Alignment Dataset: What Participatory, Representative and Individualised Human Feedback Reveals About the Subjective and Multicultural Alignment of Large Language Models (NeurIPS 2024 · similarity 0.87)
- StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMs (NeurIPS 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.