Research map / Probabilistic and scientific modeling / Molecules and proteins
2-Step Agent: How a Bayesian Decision Maker Learns from AI-Decision Support
Otto Nyberg, Fausto Carcassi, Davide Tugnoli, Giovanni Cinà
Read the paperCheck your idea against 137,217 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.
Introduces the 2-Step Agent, a Bayesian computational framework modeling how human decision makers learn from ML-based decision support.
Problem
Lack of a clear understanding of how decision makers learn from predictions of ML-based decision support in high-stakes settings such as healthcare and the judiciary.
Approach
A general computational framework, the 2-Step Agent, in which a Bayesian decision maker updates beliefs from ML model predictions to capture how the prediction's information is incorporated into human decisions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Provide ML decision support when agent's prior on treatment effect N_E is misaligned | 2-Step Bayesian agent · Univariate linear-Gaussian simulation | CATE belief and outcome Y | Improved |
| Provide ML decision support when agent's prior on historical treatment protocol (alpha_A mu) is misaligned | 2-Step Bayesian agent · Univariate simulation, well-specified model | downstream outcome Y | Hurt |
| Provide ML decision support when prior on outcome mean (alpha_Y mu) is misaligned | 2-Step Bayesian agent · Univariate simulation | CATE belief and outcome Y | No clear effect |
| Provide ML decision support when prior on covariate mean (alpha_X mu) is misaligned and model is well-specified | 2-Step Bayesian agent · Univariate simulation with intercept | downstream outcome Y | Mixed |
| Use misspecified ML-DS model (no intercept) with misaligned alpha_X mu (ablation) | 2-Step Bayesian agent · Univariate simulation without intercept | downstream outcome Y | Hurt |
| Use misspecified ML-DS model (no intercept) with misaligned N_E (ablation) | 2-Step Bayesian agent · Univariate simulation without intercept | CATE belief | Improved |
| Increase number of confounders k from 2 to 32 (ablation) | 2-Step Bayesian agent · Multivariate simulation with intercept | learning from ML-DS | No clear effect |
Closest papers by meaning
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.85)
- Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary (NeurIPS 2024 · similarity 0.85)
- Decision Theoretic Foundations for Conformal Prediction: Optimal Uncertainty Quantification for Risk-Averse Agents (ICML 2025 · similarity 0.84)
- Bayesian Strategic Classification (NeurIPS 2024 · similarity 0.84)
- Exploiting Human-AI Dependence for Learning to Defer (ICML 2024 · similarity 0.84)
- Bayesian Decision Making around Experts (NeurIPS 2026 · 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.
