AI safety & alignment · Statistical learning theory · Transfer & distribution shift / Open question
Knowing When Not to Answer
High average accuracy is not enough when mistakes concentrate on the people or situations least visible to the evaluator.
Seed catalog · Showing the bundled record.
Problem statement
For binary prediction on a finite context/action space, fix G observable subgroups, all contiguous windows of length at least W, a total-variation drift budget V and a minimum logging propensity p_min. Labels are revealed only by paid audits or a specified selection rule. Determine the attainable answer-rate and audit-cost trade-off for simultaneously bounding selective error in every sufficiently sampled subgroup/window with probability 1-delta, including adaptive actions that change future observations.
Assumptions and scope
- Define the loss, abstention action, subgroup complexity, minimum group mass and time-window lengths; arbitrary exact individual-conditional coverage is excluded by known impossibility results.
- Specify when labels arrive, which propensities are known or estimated, and how decisions affect logging. Unobserved outcomes require an access or identification assumption.
- Require an answering-rate or set-size comparison to avoid a vacuous always-abstain or all-labels solution; bound query cost and deployment horizon explicitly.
What would settle it?
A joint high-probability, subgroup/window error and usefulness guarantee with minimax query/overlap dependence, or a lower bound showing which combinations cannot be achieved.
Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.
Status and evidence
Conditional-coverage impossibility, adaptive long-run calibration, counterfactual coverage under adaptive logging, adversarial conformal abstention and label-efficient prediction already have theory. The residual target is their stronger joint subgroup/window and decision-feedback resource frontier, not basic marginal coverage or selective-risk control. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.
Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.
Sources and further reading
- Barber et al. — The Limits of Distribution-Free Conditional Predictive Inference ↗partial theory · source checked 2026-11-19
Primary abstract checked: exact conditional coverage is impossible without distributional assumptions; the paper studies achievable relaxations. Marginal coverage cannot be substituted for individual- or subgroup-conditional reliability.
- Gibbs and Candès — Adaptive Conformal Inference Under Distribution Shift ↗partial theory · source checked 2026-11-19
Primary abstract checked: adaptive long-run coverage-frequency guarantees under changing data distributions. These do not by themselves bound every short-window or subgroup error.
- Qiao et al. — Counterfactual Online Conformal Prediction Under Adaptive Logging ↗partial theory · source checked 2026-11-19
Primary abstract checked: propensity-weighted and doubly robust recursions; minimax expected coverage rates under positivity and known logging probabilities. The extra target must go beyond these adaptive-logging guarantees.
- Lee, Jung and Park — Online Conformal Abstention Under Adversarial Bandit Feedback ↗partial theory · source checked 2026-11-19
Primary abstract checked: ExAUL converts bandit regret to selective-risk violation control and achieves square-root-T guarantees in its feedback model. Selective-risk control itself is already available.
- Skalse et al. — Online Conformal Prediction Beyond Feedback ↗partial theory · source checked 2026-11-19
Primary abstract checked: separate prediction/query actions give O(T^(2/3)) expected regret and coverage error O(T^(-1/3)) on oblivious streams. Costed label acquisition already has a formal model.