Research map / Large language models / Language, safety and interpretability
Using Imperfect Surrogates for Downstream Inference: Design-based Supervised Learning for Social Science Applications of Large Language Models
Naoki Egami, Musashi Hinck, Brandon M. Stewart, Hanying Wei
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 |
|---|---|---|---|
| use DSL bias-corrected pseudo-outcome instead of surrogate-only labels | LLM surrogate (flan-ul2) · class prevalence, Toxicity (n=100) | bias x100: 34.8 -> 0.0 | Improved |
| use DSL instead of surrogate-only for coverage | LLM surrogate (flan-ul2) · class prevalence, Power | 95% CI coverage: 0.0 -> 94.8 | Improved |
| use DSL instead of gold-standard-only | gold-standard sample (n=100) · class prevalence, Misinfo. | RMSE x100: 5.0 -> 4.3 | Improved |
| use DSL instead of gold-standard-only | gold-standard sample (n=100) · class prevalence, Toxicity | RMSE x100: 5.0 -> 4.9 | No clear effect |
| use DSL instead of supervised learning (SL) | SL with cross-fitting · class prevalence, Emotion | coverage: 91.3 -> 94.0 | Improved |
| use DSL instead of SL for RMSE | SL with cross-fitting · class prevalence (17 datasets) | RMSE: 4.3 -> 4.3 (Misinfo.) | No clear effect |
| use DSL instead of GSO on congressional bills logistic regression (balanced, 0-shot) | GPT-3 surrogate · Congressional Bills Project | RMSE: 14% gain in RMSE (31% in five-shot) | Improved |
| use surrogate-only labels in downstream regression with 90% accurate surrogate | simulated logistic regression · simulation | 95% CI coverage: coverage only 40% | Hurt |
Closest papers by meaning
- Valid Inference with Imperfect Synthetic Data (NeurIPS 2025 · similarity 0.87)
- Prediction-Powered Inference Across Many Tasks for AI Evaluations and Social Science Research (NeurIPS 2026 · similarity 0.87)
- Language-Assisted Debiasing and Smoothing for Foundation Model-Based Semi-Supervised Learning (CVPR 2025 · similarity 0.85)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.84)
- Maximizing Signal in Human-Model Preference Alignment (AAAI 2025 · similarity 0.84)
- Biases Mitigation and Expressiveness Preservation in Language Models: A Comprehensive Pipeline (Student Abstract) (AAAI 2024 · similarity 0.84)
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.
