← Back to all news
arXiv Machine Learning By Parivesh Priye, Yufeng Wang, Haibin Ling, Michael Chaykowsky

Available Guardrails: Certifying Selective Prediction across ML Systems

Read the original on arXiv Machine Learning →

The paper introduces a method to certify selective prediction in machine learning systems by computing the availability of safety gates through exact-binomial inversion and dynamic programming. It demonstrates that a truth-informed planner can significantly improve mean coverage over naive approaches, and that reallocating error budgets further enhances coverage across diverse applications such as LLM tool‑calling, content moderation, lesion classification, and recommendation. The study highlights the importance of planning and finite‑sample estimation in ensuring reliable, granular deployment of selective predictors.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv Machine Learning.

Related stories

arXiv AI
Oct 19

RSD-Poker: Structure-Adaptive and Shift-Robust Risk-Utility Certification for Residual Policies in Imperfect-Information Games

RSD-Poker is a certification framework for residual policies in imperfect‑information games that adapts to game structure and remains robust to shifts in deployment state mixtures. It freezes a bank of residual families and a policy‑visible partition, then assigns each candidate‑group pair weighted upper and lower certificates for risk and utility. The framework guarantees that, with high probability, the selected candidate map meets its declared risk budget and utility bounds across a predeclared uncertainty set of deployment group proportions.

By Miaobo Hu, Shuhao Hu, Xiaobo Guo, Xin Wang, Bokun Wang, Peng Zhang, Daren Zha
More like this →
arXiv AI
2d ago

When Does Risk-Controlled Throughput Forecasting Actually Control Risk? A Reproduction Study on LEO Satellite Link

The study reproduces the BG-CFQS risk‑controlled throughput forecasting method on three Starlink traces and an additional dataset, finding that its risk guarantees hold only under random data splits and fail under contiguous temporal splits. It also shows the method cannot support risk budgets below 0.15, that static conformal calibration does not improve ranking, and that group‑conditional online recalibration often performs worse than no conditioning. These empirical results question the practical reliability of the technique for real‑world satellite link admission control.

By Mayan Sharma, Kriti Saini, Devansh Behl, Suganya R, S. Vijayaragavan
efficiency
More like this →