RESEARCH · GLOBAL
arXiv: Trust calibration for agentic tools via Bayesian preference learning
Formalizes the problem of deciding when an AI agent can act autonomously vs. requiring human approval as a preference-learning task. A Gaussian-process posterior over human risk tolerance learns from binary approve/deny feedback and escalates decisions where approval is most uncertain.
WHY IT MATTERS
Provides a principled framework for risk governance in autonomous agents; BFSI compliance teams can adopt this approach to calibrate when AI agents (e.g., fraud flagging, credit decisions) execute without human-in-the-loop vs. escalate to officers.
Source: arXiv · 2026-05-21