acceptodds
Withdrawn from ICLR 2027

Detecting Hidden Conformal Failures under Feature Shift

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Split conformal prediction provides marginal coverage under exchangeability, yet deployment-time feature shift can silently invalidate this guarantee even when point accuracy barely changes. We study label-free detection of such hidden coverage failures. Soft Conformal Boundary Sensitivity (SCBS) profiles how source-side feature interventions move examples across the conformal boundary and combines this vulnerability with null-calibrated unlabeled target drift. Across 57,600 controlled scenarios, SCBS detects five-point hidden coverage losses with dataset-macro AUROC 0.846. Direction-aware source simulation and an output-based monitor expose complementary mechanisms, so we select between them using a source-only router. Frozen on an ordered Bank Marketing audit, the router obtains AUROC/AUPRC /; across three natural hidden-failure streams it obtains / with FPR 0.146. Strong generic alternatives do not isolate harmful drift: classifier two-sample testing and linear-time MMD obtain AUROC 0.582 and 0.746 and alarm on nearly every benign shifted window. An impossibility result formalizes why unlabeled covariates cannot universally certify target coverage, and a finite-label extension restores design-based identification. The result is a mechanism-aware deployment warning, not an unconditional coverage guarantee.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal prediction, distribution shift, uncertainty quantification, deployment monitoring, coverage failure detection, unlabeled target data

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_p8mBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies label-free detection of hidden conformal failures: deployment-time feature shift that silently erodes a frozen split-conformal predictor's marginal coverage while point accuracy stays essentially unchanged (accuracy drop ≤ 0.01). It first establishes an impossibility result (Prop. 2): under arbitrary joint shift, unlabeled target covariates cannot identify target coverage, because two conditional label laws placing mass inside vs. outside the fixed prediction set yield identical observable data. This motivates a shift from estimation to a structured warning.

The proposed method, Soft Conformal Boundary Sensitivity (SCBS), implements a boundary-alignment principle: target drift signals coverage risk only when it moves along source-side directions to which the fixed conformal boundary is vulnerable. Offline feature-wise permutation interventions produce vulnerability scores vⱼ; online null-calibrated unlabeled drift scores dⱼ are computed; the risk is R = Σ vⱼdⱼ. Refinements add a direction-aware score, a source-simulated alarm threshold (Dir-SimCal), and a source-only router that selects between the boundary channel and an output-based monitor (Output-UPE) without target labels. A finite-label extension (a model-assisted estimator) restores design-based identification. Theorem 6 bounds soft miscoverage by R_SCBS up to four remainder terms under probe-alignment and drift-dominance assumptions. Validation covers 57,600 controlled scenarios (dataset-macro AUROC 0.846 for five-point failures, 0.939 for ten-point), unseen-severity and leave-one-dataset-out transfer, and three natural streams (Electricity, Airlines, Bank Marketing), where the router reaches macro AUROC/AUPRC 0.965/0.894 at FPR 0.146, versus generic detectors (C2ST AUROC 0.582, linear MMD 0.746) that alarm on ~98–100% of benign shifted windows.

Critical strengths
  1. A correct and consequential problem reframing, backed by a clean impossibility result. Proposition 2 formalizes why unlabeled covariate monitoring cannot certify coverage — a point most deployment-monitoring work simply ignores. This is consequential because it correctly lowers the goal from an unconditional guarantee to a mechanism-aware warning and gives the entire inductive bias a coherent motivation, making the paper's internal logic unusually sound for this literature.
  2. An unusually sharp and honest negative result on strong baselines. The C2ST/MMD comparison (FPR 1.000 and 0.980 on benign windows, Table 9/Table 2) crisply demonstrates that shift magnitude does not identify conformal harm, and the authors consistently separate frozen (Bank Marketing) from developmental (Electricity/Airlines) evidence and retain accuracy-visible streams as negative controls. This restraint is consequential because it makes the remaining positive claims credible rather than overfit to a favorable framing.
Critical weaknesses
  1. The headline empirical claims rest on thin truly-independent evidence, and the abstract's framing outruns it. The "three natural hidden-failure streams" aggregate (0.965/0.894) uses Electricity and Airlines to select the router channel; by the authors' own account (§C.2, §B.9) only Bank Marketing is genuinely frozen — and Bank is a single stream with roughly 30–37 failures over ~75 windows, so the AUPRC 0.997 headline (Table 8) rides on high prevalence and small n. Separately, the abstract's "across 57,600 scenarios, AUROC 0.846" is actually evaluated on the ~25,183 accuracy-stable subset (981 positives). These caveats are disclosed, but the abstract and introduction read substantially more prospective and general than the evidence supports — which matters because the paper's value proposition is a deployable monitor.
  2. The central deliverable is a heuristic whose theoretical backing is conditional on unverifiable assumptions. Theorem 6 bounds soft miscoverage under Assumptions A1 (probe alignment) and A2 (drift dominance) up to four remainder terms (drift-proxy, probe-alignment, unrepresented-shift, soft-boundary) that are never empirically bounded, and by the paper's own impossibility result the method cannot certify coverage. This is an honest position, but it means the contribution is a correlational warning with no false-alarm guarantee on an arbitrary new stream, and the generalization evidence (LODO AUROC 0.807–0.833; decay at severity multiplier 2.0 to 0.849) leaves real doubt that the boundary-alignment assumption transfers out of distribution.
Questions for authors
  1. How many streams does the routed monitor hold on with no target information used at any stage, channel selection included? If Bank Marketing is the only one, please either add a second fully-frozen stream or report the three-stream aggregate split into "frozen" vs. "used for channel selection," so the 0.965/0.894 number is not read as prospective when two of three streams are design data.
  2. Please report Bank Marketing's positive base rate per seed and the AUPRC of a prevalence-matched random ranker next to your 0.997. If the gap over that baseline is small, state it.
  3. Will you report the 57,600-scenario result on the actual evaluation subset (25,183 / 981 positives) in the abstract, or correct the denominator? A one-line fix either way.
  4. Give one concrete source-side quantity a deployed user could compute to know whether A1/A2 hold on a target window — or state plainly in the main text that no such check exists and R_SCBS is a correlate, not a certificate.
  5. Can you separate your explanation of the LODO / severity-2.0 decline ("boundary locality weakens far from source support") from the alternative that A1 fails off-distribution? A measurement of the alignment slack ξⱼ (or the remainder proxies) across severities would distinguish the two; if it cannot, the method's claim at large shift is unsupported either way.
Reviewer_45GCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes Soft Conformal Boundary Sensitivity (SCBS) to detect hidden coverage failures in a label-free way.

Critical strengths

None

Critical weaknesses
  • The paper is nowhere readable. It's impossible for me to understand the arguments because it jumps from one argument to another, randomly pops out some results, and lacks a clear narrative. Please see the proof of Proposition 2. What kind of proof is that? Where is the formality, clearness? Scientific communication is a part of doing science (more important these days!), and this paper severely fails. That's why I recommend a strong reject.
Questions for authors

None

Reviewer_tdX9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at a frozen classifier and conformal set after the input distribution shifts. Coverage can drop while accuracy stays flat, and the authors want to detect this without target labels. Their score, SCBS, multiplies two things per feature: how much scrambling that feature on source data pushes points across the conformal threshold, and how much the feature drifted in the target window. They add a signed version, a threshold set from simulated source corruptions, and a router that switches to an output-based miss predictor when that works better. They also show that coverage is not identifiable from unlabeled covariates alone. Experiments cover synthetic shifts on tabular data, one external benchmark, and three natural streams.

Critical strengths

The problem of monitoring coverage of a deployed conformal set is a good one. The paper shows with real streams that coverage can fall several points while accuracy barely moves. The link between the direction of feature shift and sensitivity at the conformal boundary is an interesting object. The authors are careful with their claims. The score is not a certificate, and the paper reports where it fails.

Critical weaknesses

The natural baseline is missing: fit a classifier on source calibration data that predicts whether the set missed a point, then average it over the target window. This uses the same inputs as SCBS (and already appears as the first term of equation (4)). SCBS approximates this average feature by feature. Without the comparison, we cannot tell whether boundary alignment adds anything.

The method is a heuristic warning with no guarantee, so by-and-large the empirical results carry the paper. They seem to rest on routing between two channels that fail in different regimes. The feature channel wins on Electricity and Bank, the output channel on Airlines, neither on ACS, and on Bank the router picks by model class on the same data. The paper gives no way to know the regime in advance. Also several controlled datasets are too small for a five-point coverage drop to be distinguishable from sampling noise.

Questions for authors
  1. Please report the direct estimate (average predicted miss rate over the window, with the miss predictor fit on raw features) as a label-free baseline. Where does SCBS or Dir-SimCal beat it?

  2. Is there a source-side diagnostic that tells a deployer which channel to trust? The router picks by model class on Bank, so the regime is not just a property of the data.

  3. Under what condition on the target distribution are the error terms in Theorem 6 small?

  4. How common is hidden conformal failure in real data, and can it be an artifact of a poorly calibrated model? If the model were calibrated on the target, the miss probability would equal the mass left out of the set, computable from outputs alone.

Clarifications: please define the dependency drift statistic, and report plain SCBS in Table 2 alongside the refinements. Why 0.05 and 0.10 for coverage and 0.01 for accuracy?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.