Research map / Large language models / Alignment and preferences
Questioning the Survey Responses of Large Language Models
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.
Surveys 43 LLMs with 25 ACS questions and shows that apparent survey responses are dominated by A-label/ordering bias; after randomizing answer order, models trend toward uniform responses, making survey-based alignment largely reflect subgroup entropy.
Problem
Survey-based evaluations of LLMs assume model answer distributions to multiple-choice survey questions faithfully represent human populations; it is unclear whether such survey responses are valid and robust.
Approach
Shows that LLM survey response distributions are confounded by A-label and choice-ordering biases; after averaging over all answer-choice permutations, responses become near-uniform regardless of size or data, so survey alignment metrics mainly track subgroup response entropy.
Key results
- Across 43 models, after randomized choice ordering, base model survey responses have normalized entropy approximately 1 irrespective of size or survey question.
- Models exhibit substantial A-bias; small models (a few billion parameters or fewer) tend toward mono-answers.
- For GPT-4, the best-aligned model, only 6 of 25 questions (24%) are closer to the U.S. census than to a uniform baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Randomize answer-choice ordering (ablation) | LLMs answering ACS questions · ACS | response entropy: approximately~$1$ | Helped |
| Use randomized choice ordering instead of fixed ACS ordering (ablation) | LLMs answering ACS questions · ACS | similarity to human populations | Hurt |
| Instruction-tune the model (ablation) | LLMs answering ACS questions · ACS | KL divergence to U.S. census | Mixed |
| Replace the B label with I (ablation) | Small language models · ACS | labeling bias | Mixed |
| Use labels R, S, N, L, O, T, M, P, W, U, Y, V instead of A, B, etc. (ablation) | Language models · ACS | labeling bias | Mixed |
| Add ACS or U.S.-household context to the prompt (ablation) | Publicly available language models · ACS individual questions | response patterns | No clear effect |
| Rephrase questions in the second person (ablation) | Publicly available language models · ACS individual questions | response patterns | No clear effect |
| Add multiple-choice instructions or an example to the prompt (ablation) | Publicly available language models · ACS individual questions | response patterns | No clear effect |
| Use chat-style or interview-style prompts (ablation) | Publicly available language models · ACS individual questions | response patterns | No clear effect |
| Keep previous questions and answers in context when asking subsequent survey questions (ablation) | Language models generating survey responses sequentially · ACS | similarity to census data: >90% | Hurt |
| Generate responses sequentially without conditioning on human respondents | Language models · 2016 ANES | similarity to survey data: > 99~\% | Hurt |
| Use randomized ordering to adjust model responses (ablation) | Base models · ACS | response entropy variation across questions | Helped |
Surprising findings
- Models exhibit substantial A-bias even when the position of the 'A' choice is randomized, unlike primacy bias in humans.
- Synthetic questionnaires generated sequentially by models are distinguishable from census data with over 90% discriminator accuracy (over 99% for ANES).
Limitations
- GPT-family models only accessible via top-5 logprobs and limited permutations (up to 50), reducing adjustment precision.
- Focus is on a single prompting methodology; conclusions may not extend to other elicitation methods.
- Survey questions binned or reformatted (e.g., HISP reduced to Yes/No), altering the original ACS.
Experimental setup
Datasets: 2019 ACS PUMS ATP (OpinionsQA) GAS/WVS (GlobalOpinionsQA) 2016 ANES
Benchmarks: 2019 ACS ATP GAS/WVS 2016 ANES
Metrics: normalized entropy KL divergence A-bias discriminator accuracy
Baselines: OpinionQA silicon sampling GPT-4
Builds on: OpinionQA silicon sampling Folktables
Closest papers by meaning
- 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 0.88)
- Examining Alignment of Large Language Models through Representative Heuristics: the case of political stereotypes (ICLR 2025 · similarity 0.87)
- Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble (NeurIPS 2026 · similarity 0.87)
- Peering Through Preferences: Unraveling Feedback Acquisition for Aligning Large Language Models (ICLR 2024 · similarity 0.86)
- Prediction-Powered Ranking of Large Language Models (NeurIPS 2024 · similarity 0.86)
- Uncertainty Quantification for LLM-Based Survey Simulations (ICML 2025 · similarity 0.85)
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.