Research map / Large language models / Language, safety and interpretability
AutoElicit: Using Large Language Models for Expert Prior Elicitation in Predictive Modelling
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.
Proposes AutoElicit, which prompts LLMs (GPT-3.5-turbo, GPT-4-turbo, DeepSeek-R1-32B) to elicit Gaussian mixture priors over linear model parameters and compares them with in-context learning and uninformative priors using energy statistics and Bayes factors.
Problem
Eliciting expert priors for linear predictive models is time-consuming and costly, especially in low-resource clinical settings where labelled data is scarce; LLM in-context prediction is costly, opaque, and weak at numerical prediction.
Approach
Ask an LLM for per-feature Gaussian mean and standard deviation over many prompt variants, mix them into a prior over linear model weights, and then perform Bayesian updating with observed labels; also extract the implicit in-context prior/posterior via MLE to test whether ICL behaves Bayesianly and use the Bayes factor for model selection.
Key results
- On the synthetic task, GPT-3.5-turbo-elicited priors give an order of magnitude lower test MSE than an uninformative prior after 5 observed examples and match it after 20.
- On the private UTI task, all AutoElicit priors give higher accuracy than the uninformative prior at every training size, reaching peak accuracy about 220 days (over six months) earlier in the study.
- The energy statistic between GPT-3.5-turbo's approximated in-context prior and its elicited prior ranges from 0.15 to 0.39, indicating the in-context prior differs from the elicited one.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoElicit LLM-elicited priors instead of an uninformative prior | Linear predictive models · UTI | accuracy | Helped |
| Use AutoElicit LLM-elicited priors instead of an uninformative prior | Linear predictive models · Synthetic task: y = 2x1 - x2 + x3 | MSE: an order of magnitude lower test MSE after 5 observed examples | Helped |
| Use GPT-3.5-turbo AutoElicit prior instead of an uninformative prior | Linear predictive model · Heart Disease | mean posterior accuracy | No clear effect |
| Use GPT-4-turbo or DeepSeek-R1-32B-Q4 AutoElicit priors instead of an uninformative prior | Linear predictive model · Heart Disease | mean posterior accuracy | Helped |
| Use AutoElicit priors instead of in-context learning | Predictive models · Tasks tested | accuracy and MSE | Helped |
| Add expert-provided natural-language correlations for previous UTIs, night waking, and bathroom usage | AutoElicit · UTI | posterior accuracy | Helped |
| Use GPT-3.5-turbo in-context prior instead of its elicited prior | GPT-3.5-turbo predictive model · UTI, Heart Disease, Hypothyroid, Diabetes, and Breast Cancer | energy statistic: 0.39; 0.18; 0.15; 0.38; 0.31 | Mixed |
| Use in-context learning rather than AutoElicit priors | GPT-3.5-turbo · Diabetes | agreement with Bayesian posterior | Helped |
| Use in-context learning rather than AutoElicit priors | GPT-3.5-turbo · Heart Disease | agreement with Bayesian posterior | Hurt |
| Use GPT-3.5-turbo in-context learning to imitate a linear model | GPT-3.5-turbo · Breast Cancer and Hypothyroid | linear-model prediction approximation error | Hurt |
Surprising findings
- Approximated in-context learning often fails to imitate a linear model (e.g., hesitant to predict negative values on the synthetic task), and in-context posteriors are inconsistent with Bayesian updating on Heart Disease.
- GPT-3.5-turbo returned uninformative priors for Heart Disease, yielding the same accuracy as the uninformative baseline.
Limitations
- Priors may carry unwanted bias from LLM training; the authors suggest uninformative priors for sensitive features.
- Only linear models are studied; extension to neural networks and non-linear models is left for future work.
- Possible memorisation of public datasets by LLMs complicates evaluation on public tasks.
Experimental setup
Datasets: Heart Disease Diabetes Hypothyroid Breast Cancer UTI (private dementia-care sensor data) synthetic linear task y=2x1-x2+x3
Benchmarks: Heart Disease Diabetes Hypothyroid Breast Cancer UTI synthetic linear task
Metrics: posterior accuracy mean squared error energy statistic Bayes factor
Baselines: uninformative prior in-context learning LLM-generated synthetic data pretraining approach
Builds on: in-context learning Bayes factor
Closest papers by meaning
- LLM Processes: Numerical Predictive Distributions Conditioned on Natural Language (NeurIPS 2024 · similarity 0.87)
- Can Generative AI Solve Your In-Context Learning Problem? A Martingale Perspective (ICLR 2025 · similarity 0.85)
- Pre-trained Large Language Models Learn to Predict Hidden Markov Models In-context (NeurIPS 2025 · similarity 0.85)
- Is In-Context Learning in Large Language Models Bayesian? A Martingale Perspective (ICML 2024 · similarity 0.85)
- ELICIT: LLM Augmentation Via External In-context Capability (ICLR 2025 · similarity 0.84)
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · 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.