Evaluating probabilistic decision making by LLMs: a decision-theoretic framework and application...
Talkat Bridging Prediction and Intervention Problems in Social Systems (Simons Institute, 2026) · 2026-01-12..01-18·
summary
Proposes a revealed-preference elicitation that asks an LLM for both verbalized probabilities and its chosen action, fits a discrete-choice utility model (false positive/negative and defer costs), and simulates interventions. Applied to medical diagnosis datasets (including structural heart disease), verbalized probabilities correlate with truth but are overconfident; the implied utility rationalizes the chosen action about 80% utility-maximizing of the time. Experiments show calibration fixes minimal downstream loss, while prompting the loss function helps mainly for high-reasoning models.
AI-written summary · about
chapters (7)
- 0:00Evaluating LLM decisions
- 1:51Decision-making under uncertainty
- 4:50Decision-theoretic framework
- 14:56Empirical framework
- 20:11Diagnosing heart disease with LLMs
- 27:49Error attribution and fixes
- 31:50Summary
related talks
39mHallucination-Free LLMs: Strategies for Monitoring and Mitigation
27mExploring the Mechanisms of LLM Reasoning Capabilities
40mThe causality trap: why your LLM sounds smart but thinks shallow
01h 29mLLM Safety, Alignment, and Generalization
38mMaking LLMs reliable - A practical framework
29mLLMs that think: Demystifying Reasoning Models
29mCustomizing and Evaluating LLMs, an Ops Perspective
46mEnhancing Agent Reliability Through Evaluation-Driven Development