SEU Sensitivity

Modeling Sensitivity to Subjective Expected Utility Maximization

0.1 Introduction

This project provides a Bayesian framework for modeling and analyzing decision-making behavior through the lens of Subjective Expected Utility (SEU) theory. We develop computational tools for measuring an agent’s sensitivity to SEU maximization—captured by a parameter α that governs how consistently agents maximize expected utility.

ImportantStart here: the working paper

The core methodological results in this project have been consolidated, revised, and reviewed in a working paper:

Sensitivity to Subjective Expected Utility Maximization: A Methodological Study, with an Illustrative Application to LLM Decision-Making (PDF) — Jeff Helzner

The working paper is the canonical reference for this work. It consolidates and supersedes the foundational reports on m_0 and m_1 (Reports 1–6) and the application reports below; where a technical report and the paper differ, the paper reflects our current, more considered view.

The technical reports are retained as a more detailed, work-in-progress record of the analyses that fed into the paper — useful for implementation specifics, intermediate results, and provenance, but rougher and not always current. The reports on generalization (m_2/m_3, Report 7) and the hierarchical model (Reports 8–14) explore directions not yet incorporated into the paper.

0.2 Report Series

0.2.1 Foundational Reports

These reports establish the theoretical and methodological foundations:

1. Abstract Formulation — Mathematical specification and key theoretical properties
2. Concrete Implementation — Stan model implementation details
3. Prior Analysis — Prior predictive analysis and prior selection
4. Parameter Recovery — Validation that parameters can be recovered from data
5. Adding Risky Choices — Extension to model m_1 for utility identification
6. SBC Validation — Simulation-based calibration results
7. Generalizing Sensitivity — Generalized models m_2 and m_3
8. Hierarchical Formulation — Hierarchical extension for population-level inference
9. Hierarchical Implementation — Stan implementation and validation of hierarchical models
10. Hierarchical Prior Analysis — Prior predictive analysis for h_m01
11. Hierarchical Parameter Recovery — Recovery validation for the hierarchical model
12. Hierarchical SBC Validation — Simulation-based calibration for h_m01
13. Concentrated δ Prior — Evaluating a more concentrated Dirichlet prior on δ as a route to tighter β–δ recovery
14. Does m_1 Identify δ? — Matched-design recovery testing whether adding risky choices delivers the δ-identification gain predicted by Report 5

0.2.2 Application Reports

0.2.2.1 Temperature Study

1. Initial Results — How LLM temperature affects estimated SEU sensitivity

0.2.2.2 Temperature Study: EU Prompt

1. EU Prompt Study — Effect of explicit EU-maximization framing on sensitivity

0.2.2.3 Temperature Study: Risky Alternatives

1. Risky Alternatives Extension — Risky choice data for utility identification across temperatures

0.2.2.4 Ellsberg Study

1. Ellsberg Study — Claude 3.5 Sonnet on Ellsberg urn gambles

0.2.2.5 Factorial Cells

• Claude × Insurance — Claude 3.5 Sonnet on insurance claims triage
• GPT-4o × Ellsberg — GPT-4o on Ellsberg urn gambles

0.2.2.6 Factorial Synthesis

1. 2×2 Factorial Analysis — Cross-LLM × cross-task synthesis

0.3 Getting Started

See the GitHub repository for installation instructions and code.

0.4 Citation

If you use this work, please cite:

Helzner, J. (2026). SEU Sensitivity: A Bayesian Framework for Modeling 
Decision-Making Sensitivity. https://github.com/jeffhelzner/seu-sensitivity