Calibrated deferral: teaching assistants when not to answer
S. Park · Reasoning Group
Biuron Technical Notes
An assistant that always answers is easy to build and hard to trust. We train assistants to represent their own uncertainty faithfully and to defer, whether by asking, escalating, or abstaining, when confidence is low.
Across knowledge-work tasks, calibrated deferral improves end-to-end outcomes even though it lowers raw answer coverage, because the errors it avoids are the expensive ones.
This is a plain-language summary. The full manuscript and supplementary material are available on request.