acceptodds
Under review as a conference paper at ICLR 2027

Defect-Aware Confidence Routing: Black-Box Confidence Signals Have Regime-Dependent Reliability

Abstract

Reliable LLM deployment requires knowing when to abstain. In the black-box setting, practitioners combine confidence signals—verbalized confidence, self-consistency, cross-model dissent—using a single fixed rule. We show this is fundamentally miscalibrated: the reliability of each signal is regime-dependent. Cross-model dissent is an excellent error indicator on unanswerable questions (AUROC up to ) but collapses to near-chance () on answerable ones, where it no longer separates errors, so any fixed fusion is trapped in a cross-regime compromise—and indeed degrades below the best single signal. We turn this into a method, Defect-Aware Confidence Routing (DACR). Two facts make it deployable: the polarity depends on the latent defect type, and the defect type is predictable from the question text by a light probe (AUROC ) that leaks no labels. DACR uses this input-side probe to route signals—selecting which signal and with which polarity—through probe signal interactions. End-to-end, DACR significantly beats the best single signal by points in-domain and points under zero-transfer to SelfAware (paired bootstrap CIs exclude zero), recovering of the per-slice oracle gap; the same gate is diagnostic, predicting on which corpora the router transfers directly and on which its defect structure differs enough to require re-fitting. A five-seed ablation isolates the gain to routing (the same features without routing interactions degrade below behaviour-only), and leave-one-model-out confirms it generalizes to five of six held-out models.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
selective prediction, abstention, uncertainty estimation, confidence calibration, large language models

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_uiFKOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors tackle the problem of distinguishing answerable and unanswerable questions. For an answerable question, there exists a correct answer, while an unanswerable question is controversial, ambiguous, based on a false assumption, counterfactual, unsolved, or about a future unknown. They find that existing uncertainty estimates (verbalization, self-consistency, ensemble) and a fixed combination do not do well. They propose DACR, which first predicts whether the question is answerable or unanswerable, then uses this prediction to decide how to combine the uncertainty estimates. DACR leads to better results.

Critical strengths
  • They find that the answerability of a question is much more easily predicted from the question itself than from models’ behavior (i.e. uncertainty estimates based on verbalization, self-consistency, ensemble). This is a mild contribution, since it means that we can detect unanswerable questions with minimal means.
Critical weaknesses
  • The problem tackled in this paper – distinguishing answerable and unanswerable questions – has very little practical utility and thus severely limits the impact of the work. The typical and useful usage of confidence/uncertainty is to estimate the probability that a given prediction is correct (e.g. to abstain when the probability is low). The key difference from the problem tackled in this paper is that even if a question is answerable (i.e. there exists a correct answer), the model may not know the answer, which means that it is best to abstain. Thus, the basic premise that answerable questions should always be answered is flawed.
Questions for authors
  • How is the within-slice AUC defined, e.g. in Figure 2? My guess: for the answerable class, the AUC is computed on questions that are answerable or unanswerable (any type t); for each unanswerable type t, the AUC is computed on questions that are answerable or unanswerable type t.
1 comment

Thank you for the clarifications. I will maintain my score, given that the authors' response confirms my W1.

Reviewer_ajm4OpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors study selective Q&A with abstention for black-box LLMs. Specifically, they consider the known-unknowns setting where each question is either answerable (the model should answer) or unknown (the model should abstain).

They propose Defect-Aware Confidence Routing (DACR), a metod that weighs three different confidence signals (verbalized confidence, self-consistency, cross-model agreement) by training a probe on the input question. The method is based on the observation that individual confidence signals can work well on unknown questions but poorly on answerable questions, while the proposed input-conditioned/predicted weighting should be able to work well across both settings.

The proposed method is compared with the best individual confidence signal, a fixed global fusion classifier, and an oracle. They mainly evaluate on the KUQ dataset, using six recent LLMs.

Critical strengths
  • The studied problem is quite interesting, and the proposed approach (conditioned weighting of different confidence signals, predicted from the input question) makes some sense overall.
Critical weaknesses
  • The paper could definitely be more well written, the overall presentation is a bit sloppy, and some parts are somewhat convoluted and difficult to follow (especially Section 5).

  • It is somewhat unclear how well the proposed method actually works, as the main results are just cross-validation on KUQ, and the results on SelfAware, SQuAD and TruthfulQA don't seem overly impressive.

Questions for authors

Questions/suggestions:

  • Abstract/Introduction: "+2.9 points in-domain and +6.5 points under zero-transfer to SelfAware", "recovering ∼45% of the per-slice oracle gap", at least when reading the paper at this point, it's not clear to me what "SelfAware" or "the per-slice oracle gap" mean.

  • Left of Figure 1, it says "[placeholder -- export final from fig_dacr.pptx]". I guess 0.24 in the left sub-figure should be 0.54? "the per-slice oracle (green) is the routing upper bound" also seems to be missing.

  • And related to Figure 1, the 0.24 for "Cross-dissent" on " Known" in Figure 2 should be 0.54 instead as well then I suppose?

  • The main paper just says "six contemporary LLMs", should probably give a bit more detall (not just in Appendix N).

  • Section 5: "DACR recovers this slice to 0.831—a +0.295 within-slice swing that turns a near-chance, uninformative signal into a strong predictor, close to the oracle ceiling here of 0.745", this doesn't make sense, DACR is better than the oracle ceiling?

  • Table 1 to Table 4, still a substantial drop for DACR though, from 0.85 to 0.65 AUROC. Should it not be able to generalize better?

  • I struggle to understand "The defect structure persists across further external distributions" in Section 5 and the corresponding results in Table 5, these results don't seem particularly impressive/convincing?

Minor things:

  • Abstract: "(paired bootstrap CIs exclude zero)", I don't think you need to include this detail in the abstract.
  • Figure 3 and Figure 4 could probably be made single-column, to save space (they're a bit unnecessarily big). Table 5 could also be made a bit smaller.
  • Could perhaps point to Figure 1 earlier, also in the Introduction. Same with Figure 2. Aslo, Algorithm 1 is not referenced in the text.
  • Figure 4 caption: "DACR (red)", but the curve is blue.
  • Table 4 caption (and elsewhere inte the text), "Zero-transfer" --> "Zero-shot transfer"?
Reviewer_rJXKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies black-box selective QA, where an LLM may decline to answer and a confidence signal flags wrong decisions. It uses verbalized confidence, self-inconsistency, and cross-model dissent, three common signals. Dissent is the share of other models that decide differently. On KUQ, the paper reports that dissent catches errors well on unanswerable questions but is near chance on answerable ones, and argues that any fixed combination of signals has to compromise.

It then proposes DACR. Three LLMs read only the question and estimate how likely it is to be unanswerable. A logistic regression then combines the three signals, and how much each signal counts depends on that estimate, which the paper calls routing. For example, dissent can count a lot when a question looks unanswerable and little when it looks answerable. The evaluation uses six API models and computes AUROC within each question type, which the paper calls within-slice AUROC. The main claim is that DACR beats the best single signal by 0.029 within-slice AUROC on KUQ and by 0.065 on SelfAware without retraining, and that the gain comes from routing.

Critical strengths

The paper studies a useful question about whether black-box confidence signals keep the same meaning across question types. This question matters because selective QA systems often combine such signals with one global rule. The most useful finding is that dissent is the best signal on every unanswerable type, while self-inconsistency is best on answerable questions. The stated evaluation design is also careful. Computing AUROC within each question type stops a signal from looking good just by tracking which types are harder. The headline gains are tested as paired differences with a bootstrap over questions. I also appreciate that the paper reports where DACR falls short, on SQuAD 2.0 and TruthfulQA.

On AI usage, I used it mainly to ask questions about the paper to check its empirical validity, for example, whether the results are statistically significant, and to search the literature to make sure the baselines are appropriate. I also used it to check parts of the paper against the appendix, so I didn’t have to read the whole appendix myself. Finally, I used it to reformat my comments into a proper structure and to paraphrase the text so it flows well.

Critical weaknesses

My main concern is what within-slice AUROC measures under this label. An error is a majority decision that disagrees with the gold decision, which is fixed within a question type. So, on an unanswerable type an error means the model answered, and on answerable questions it means the model declined. Once the decision is known, the error depends only on whether the question is answerable, which the probe estimates. If the peer-abstention rate means the share of other models that decline, combining it with dissent reveals the decision unless the peers split evenly. The simplest baseline is therefore missing. If the model answered, the baseline would use the probe's probability that the question is unanswerable. Otherwise, it would use one minus that probability. Without this baseline, it is unclear whether the gain comes from routing or from combining the probe with the decision. The ablation shows the interactions matter once the probe features are added. Still, the behaviour-only model already gives about two-thirds of DACR's gain, and the rest has no paired interval. As written, the paper shows that DACR beats the best single signal, but does not convincingly demonstrate that routing is responsible for the gain.

My second concern is that the headline table does not follow the paper's own protocol, which computes AUROC within each question type and never pools across types. Table 1's 0.8228 for dissent cannot be a sample-weighted average of Table 12's per-type values, since such an average is at most about 0.80 for any split of records the setup allows. Instead, Table 1 matches Table 2's two rows weighted by the answerable share, pooling the six unanswerable types. This matters because pooling is the confound the paper says it avoids, so it is unclear whether DACR's gain stays significant under the stated protocol. Figures 1 and 2 also put dissent at 0.24 on answerable questions, far below chance, while the tables and text say about 0.54, close to chance. Figure 1 still carries a placeholder label. Recomputing a headline table under the stated protocol, with figures regenerated from the same results, would fix this. Other headline claims the tables contradict are listed at the end.

Overall, I think the paper asks a useful question, and the per-type diagnostic is worth building on. If DACR keeps a significant paired gain under the stated protocol, and beats the probe-plus-decision baseline and the behaviour-only model with paired intervals, I would raise my score. These checks should be cheap, since the paper says it can regenerate all reported numbers offline. Unless answer correctness is judged, claims about wrong answers should also be narrowed to claims about wrong decisions.

Questions for authors
  1. Can the authors clarify how the error label is computed on KUQ and SelfAware? Appendix D also counts wrong answers, but Appendix N lists a correctness judge only for TruthfulQA. If you judge and count wrong answers, that would address part of my first concern. Please also define the router's peer-abstention rate.

  2. Could the authors report the probe-plus-decision baseline on KUQ and SelfAware with the same paired bootstrap, and a paired interval for DACR against the behaviour-only model? If DACR clearly beats both, my main concern would be largely resolved.

  3. Can the authors clarify how Table 1 combines the question types, and which figures and tables are final?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.