ConfGuide: Conformal VLM Guidance for Hybrid Decision Making
A new AI framework called ConfGuide improves hybrid decision making by generating succinct, targeted textual guidance. Building on the learning to guide (LtG) paradigm, where humans retain final decision authority, ConfGuide uses conformal risk control to select a set of outcomes with a guaranteed cap on false negative rate. This addresses the issue of existing approaches that compound information about all possible outcomes, making guidance difficult to digest. The approach is demonstrated on a real-world multi-label medical diagnosis task, showing promise in reducing cognitive load while improving decision quality.
Key facts
- ConfGuide is a novel approach in the learning to guide (LtG) framework
- It uses conformal risk control to select outcomes
- Ensures a cap on the false negative rate
- Generates more succinct and targeted guidance
- Demonstrated on a real-world multi-label medical diagnosis task
- Aims to improve human decision quality and reduce cognitive load
- Human is always responsible for the final decision
- Addresses the limitation of existing approaches that compound information about all outcomes
Entities
—