TMLR 2026 paper on uncertainty-aware learning to defer, using conformal prediction to make human-AI routing safer under rejector uncertainty and distribution shift.
Feb 1, 2026