Research map / Large language models / Alignment and preferences
AlignSurvey: A Comprehensive Benchmark for Human Preferences Alignment in Social Surveys
Chenxi Lin, Weikang Yuan, Zhuoren Jiang, Biao Huang, Ruitao Zhang, Jianan Ge, Yueqian Xu, Jianxing Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune with two-stage SurveyLM alignment | Qwen2.5-7B · Task 3 stance accuracy | accuracy: 36.88% -> 56.83% | Improved |
| fine-tune with two-stage SurveyLM alignment | Mistral-7B-v0.3 · Task 4 ASE | accuracy: 28.04% -> 54.59% | Improved |
| fine-tune Qwen2.5-7B instead of using Qwen2.5-72B | Qwen2.5 · Task 1 | accuracy: 37.83% -> 54.65% | Improved |
| add few-shot examples (3) | general LLMs · Task 1 / Task 3 | accuracy: GPT-4o Task1 44.08% -> 40.59%; Task3 38.12% -> 42.73% | Mixed |
| add few-shot examples | general LLMs · Task 3 group-level | Wasserstein distance: GPT-4o 1.648 -> 1.180 | Improved |
| fine-tune SurveyLM | Qwen2.5-7B · Task 4 ASE | Wasserstein distance: 1.8479 -> 0.0541 | Improved |
| add Stage I foundation adaptation (ablation) | SurveyLM Mistral · Task 4 ASE | accuracy: 0.53 -> 0.55 | Improved |
| add Stage II task-specific SFT (ablation) | SurveyLM Mistral · Task 4 ASE | accuracy: 0.36 -> 0.55 | Improved |
| add Stage I foundation adaptation (ablation) | SurveyLM Qwen · Task 4 ASE | Wasserstein distance: 0.889 -> 0.054 | Improved |
| SurveyLM on Task 2 naturalness | Mistral-7B · Task 2 | naturalness score: 3.21 -> 2.61 | Hurt |
| SurveyLM Task 3 group Wasserstein vs few-shot Mistral | Mistral-7B · Task 3 group | Wasserstein distance: 0.302 -> 0.297 | No clear effect |
Closest papers by meaning
- Prompts to Proxies: Emulating Human Preferences via a Compact LLM Ensemble (NeurIPS 2026 · similarity 0.88)
- Questioning the Survey Responses of Large Language Models (NeurIPS 2024 · similarity 0.88)
- 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)
- Position: Rethinking LLM Bias Probing Using Lessons from the Social Sciences (ICML 2025 · similarity 0.85)
- No Preference Left Behind: Group Distributional Preference Optimization (ICLR 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.
