{"id":"ad3fa3d7-a47b-492c-a861-f18339b82880","arxiv_id":"2507.12913","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"Aleatoric uncertainty selects between counterfactual and feature-importance explanations, epistemic uncertainty rejects unreliable explanations, and correlation experiments support the rule.","lead":"An international team proposes using the model's own uncertainty to choose which explanation to show: counterfactuals when data is noisy, feature-importance when data is clean, and no explanation when the model is ignorant. The paper reports correlations across 10 tabular datasets and MNIST suggesting this selection yields more stable and attainable explanations, though code and baselines are missing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline AU-vs-counterfactual-dissimilarity correlation is dominated by AUbel, which shares the K=7 neighborhood machinery with the counterfactual definition (Eq. 3 vs Eq.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing concern: AUbel (Eq. 13) and the counterfactual distance (Eqs. 3 and 7) are both computed from the K=7 nearest-neighbor structure of the same training set, so the strong negative correlations may be partly definitional. This matters because the paper's central claim is that aleatoric uncertainty is a practical guide for choosing between counterfactual and feature-importance explanations. If the main supporting correlation is an artifact of shared machinery, the empirical foundation of the protocol is weakened, and the mixed results from the independent estimators (AUens, AUcen) reinforce that worry. The concern is concrete and testable: by varying the neighborhood definition or using an independent counterfactual generator, one can determine whether the correlation reflects a genuine predictive relationship or a tautology. I do not think the concern by itself warrants rejection: the protocol idea is plausible, the paper reports significance tests and multiple UQ methods, and the feature-importance correlations (Table 2) are less obviously coupled. But it does justify the CONDITIONAL verdict: the authors should provide the decoupling test or a protocol-level comparison before the claim is accepted as established. The code is also not public, which makes independent verification harder; this is a secondary, not primary, issue. Overall, the verdict should remain conditional, and the reader's weakest-assumption and my independent reading agree on the key risk.","tokens_in":17500,"tokens_out":6387,"duration_ms":76466,"concrete_test":"Decouple the shared KNN structure: keep the K=7 KNN classifier for counterfactual generation and for computing AUbel, but recompute AUbel on a neighborhood defined by a different distance metric (e.g., L1 or cosine) or by a different number of neighbors K' (e.g., 3 or 15), then recompute the Spearman correlations in Table 5. If the correlations remain around -0.8, the coupling is not the sole driver; if they collapse toward zero or change sign, the reported negative correlation is an artifact of the shared KNN neighborhood. A complementary check is to replace the counterfactual distance with one from an independent counterfactual generator (e.g., a gradient-based method or a separately trained model) and test whether AUbel still predicts it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest positive result in the paper is the negative Spearman correlation between aleatoric uncertainty and counterfactual dissimilarity (Table 1, Table 5). For the belief-based method AUbel, the correlations are very strong (-0.83, -0.82, -0.89). But AUbel is computed from the class conflict inside the K=7 neighborhood of the test point (Eq. 13), while the counterfactual distance d in Eq. (7) is the distance to the nearest training point with a different label (Eq. 3), found in the same training set and with the same KNN classifier (K=7, Appendix C). A small d means a different-class neighbor is inside that same neighborhood, which mechanically tends to increase the class conflict captured by AUbel. The correlation is therefore at least partly a restatement of the shared neighborhood structure, not an independent confirmation that aleatoric uncertainty predicts counterfactual attainability. The methodologically independent estimators provide weak or even inconsistent evidence: AUens is positive on Glass (+0.15) and Ionosphere (+0.30), and AUcen is near zero on Ionosphere (-0.04). Since the paper explicitly concludes that 'more local uncertainty quantification should be privileged' (i.e., AUbel), the central empirical pillar of the explanation-selection protocol rests on a measure that is entangled with the outcome it is supposed to predict. No protocol-level benchmark against a fixed explanation policy is provided, so these correlations carry the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-part protocol for combining uncertainty quantification (UQ) with local explainability: (i) use aleatoric uncertainty (AU) to choose between counterfactual explanations (high AU) and feature-importance explanations (low AU), and (ii) use epistemic uncertainty (EU) as a reject option for explanations of predictions made in regions of missing training data. The empirical strategy is to measure Spearman correlations between AU estimates from three UQ methods (belief-based AUbel, centroid-based AUcen, ensemble-based AUens) and two explanation robustness metrics: the distance to the nearest counterfactual (Eq. 7) and the discrete local Lipschitz constant of SHAP explanations (Eq. 4). The authors report a significantly negative AU-vs-dissimilarity correlation (Table 1), a mostly positive AU-vs-un-robustness correlation (Table 2), and illustrative rejection curves for a Glass dataset and MNIST. They argue that uncertainty decomposition can serve as a practical guide to selecting more robust and attainable explanations.","tokens_in":17866,"tokens_out":6265,"duration_ms":67649,"significance":"The intersection of UQ and XAI is timely, and the proposed protocol is intuitive and potentially useful. The paper contributes a broad empirical study across 10 UCI datasets and MNIST, with three UQ families and a reproducible experimental setup (code is promised as supplementary material). The idea that epistemic uncertainty itself can be an explanation—'the model lacks training'—is a sensible and refreshing framing. However, the central empirical pillar—the negative AU-vs-counterfactual-dissimilarity correlation—is weakened by the fact that the belief-based estimator AUbel is constructed from the same K-nearest-neighbor geometry as the counterfactual definition. The independent estimators give mixed and sometimes sign-inconsistent results. Moreover, the paper never evaluates the full selection protocol against a fixed explanation policy, so the evidence as presented supports correlations but not the claimed benefit of the decision rule. If the tautology concern can be resolved and a protocol-level benchmark is added, the contribution would be valuable to the XAI community.","major_comments":[{"comment":"The headline negative correlation between AUbel and counterfactual dissimilarity (Table 1: -0.83, -0.89, -0.82 for several datasets) is at least partly a tautology arising from shared neighborhood structure. AUbel (Eq. 13) is computed from the belief function over the K=7 nearest neighbors of xt, and the counterfactual distance d (Eq. 7) is the Euclidean distance to the nearest training point with a different label (Eq. 3), which is exactly the kind of class conflict that increases the discord measure in Eq. (13). When a different-class neighbor lies inside the K=7 neighborhood, d is small and the local class conflict is high; when all K neighbors share a label, the nearest different-class point is farther away and the discord term is typically zero. Consequently, a strong negative Spearman correlation between AUbel and d is expected under a null model that has no genuine relationship between aleatoric uncertainty and counterfactual attainability. The independent estimators provide inconsistent evidence: AUens is positive on Glass (+0.15) and Ionosphere (+0.30), and AUcen is near zero on Ionosphere (-0.04). Since the paper explicitly concludes that 'more local uncertainty quantification should be privileged' (Section 3.1.1), the central empirical claim rests on the entangled measure. Please provide a control analysis that removes the shared-neighbor effect, e.g., by computing the AUbel vs. d correlation separately for test points whose nearest different-class neighbor is inside vs. outside the K=7 ball, or by a permutation test that randomizes labels within neighborhoods. Without such an analysis, the paper's main conclusion is not empirically supported.","section":"§3.1.2, Table 2"},{"comment":"Table 2's heading says 'aleatoric uncertainty ... vs. robustness of feature-importance explanations', but the metric L(xt) in Eq. (4) is a Lipschitz constant, i.e., a measure of un-robustness (larger L means less robust). The text in §3.1.2 correctly speaks of 'un-robustness' and states that 'more aleatoric uncertainty leads to less robust SHAP explanations'. The positive correlations in Table 2 therefore mean high AU is associated with high un-robustness, which is the intended claim. However, the table and figure captions conflate the two directions, and a reader who takes the table at face value would draw the opposite conclusion. Please correct the labels in Table 2 and any related axes/captions to explicitly state that the correlation is with un-robustness (or equivalently, that positive correlation means AU is positively associated with the Lipschitz constant). Also clarify in the text why this direction is the one that supports the protocol.","section":"§3.1.2, Table 2"},{"comment":"There is a direct inconsistency in the number of experimental runs. Section 3.1 states: 'Spearman's rank correlation are averaged over 100 runs and concern all decomposition methods.' Appendix D, however, states: 'Each experiment is repeated 5 times (due to SHAP expensive complexity).' The SHAP robustness results in Table 2 and Appendix D are thus based on only a few runs, not 100, and the main text's '100 runs' claim does not apply uniformly. Moreover, the p-values reported in Appendix D (e.g., Tables 9-12) are computed from a single test set and a single run, not from a distribution over runs. The number of runs should be stated consistently for each experiment, and the main text should avoid implying 100-run averages for the SHAP results.","section":"§3.1 and Appendix D"},{"comment":"The paper proposes a decision rule: reject explanations when EU is high, select counterfactuals when AU is high, and select feature-importance when AU is low. Yet the experiments only report per-dataset correlations between AU and each explanation metric; they never evaluate the full rule against a fixed policy such as 'always counterfactual', 'always SHAP', or 'random selection'. A user following the protocol might not experience lower average dissimilarity or lower average Lipschitz constant if the correlations are dominated by extreme points or if the AU thresholds are poorly chosen. To substantiate the protocol's usefulness, please provide a benchmark: for each dataset, compute the average counterfactual dissimilarity (or SHAP un-robustness) under the AU-based selection rule and compare it with the average under a fixed policy and under random selection, with confidence intervals over the repeated runs. Without this comparative evaluation, the paper demonstrates relationships but not the value of the proposed protocol.","section":"§3 overall"},{"comment":"The reject option based on epistemic uncertainty is illustrated with only one extreme example on Glass and one on MNIST, and the rejection curves (Figures 5a, 6a) are not compared against any baseline (e.g., random rejection or rejection by total uncertainty). Additionally, the feature-importance experiment uses an ad-hoc rule—'rejecting 30% of instances with highest epistemic uncertainty'—to 'better capture true aleatoric uncertainty' (§3.1.2). This threshold is not justified or validated, and the appendix reports a significance test only on the remaining 70% without showing that the rejected instances are actually those for which explanations are unreliable. Please quantify the effect of rejection on explanation quality (e.g., error rate, explanation stability) as a function of the rejection threshold, and compare the EU-based rejection with a random rejection baseline.","section":"§3.2 and §3.1.2"}],"minor_comments":[{"comment":"The title in the full text reads 'A P ATH' — presumably a typo for 'A PATH'.","section":"Title"},{"comment":"'many attemps' should be 'many attempts'.","section":"§1"},{"comment":"In Eq. (3), the condition 'h(xt) != yc' is ambiguous because yc is not explicitly defined as a class label different from h(xt). Please define yc clearly (e.g., 'yc is a class in Y such that yc \\neq h(xt)').","section":"Eq. (3)"},{"comment":"The p-values in Appendix C are all near zero even for near-zero correlations (e.g., Ionos. with AUcen: -0.04, p=5.60e-7) because the test sets are large. Report the effect sizes with confidence intervals or standardized effect sizes, and note that statistical significance does not imply practical importance.","section":"Appendix C"},{"comment":"The centroid-based method simplifies the method of Van Amersfoort et al. (2020) by ignoring the learned weight matrix and using Euclidean distance on the raw features. This should be stated as a simplification, and its implications for the interpretation of AUcen should be acknowledged.","section":"Appendix C and D"},{"comment":"The neighborhood N(xt) is defined as the ball of radius ϵ, and the paper says L is estimated with 30 samples drawn from this neighborhood. Please specify how the 30 samples are drawn (e.g., uniform in the ball, or in a shell) and whether the same samples are used for each test point.","section":"§2.2"},{"comment":"The header 'B. C ANCER' contains a spurious space; please fix to 'B. Cancer'.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important and timely topic, but the central empirical support for the counterfactual-selection claim is compromised by the tautological relationship between AUbel and the counterfactual distance. The independent UQ methods give inconsistent evidence, and the paper lacks a protocol-level evaluation against a fixed policy. These issues are load-bearing but potentially fixable with additional analyses; I therefore recommend major revision rather than rejection. I also note that the SHAP experiment's smaller number of runs (5 vs. the stated 100) and the ad-hoc 30% rejection threshold should be addressed explicitly in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper proposes a genuinely new protocol for choosing explanation type from uncertainty decomposition—counterfactuals when aleatoric uncertainty is high, SHAP when low, reject when epistemic is high—and backs it with a wide correlational study. The idea is plausible and useful, and the experimental breadth is real: ten UCI datasets, three UQ families, 100 runs for the counterfactual experiment, and mostly significant Spearman correlations.\n\nThe problem is that the paper's strongest pillar may be partly circular. The AUbel measure in Eq. (13) is the Discord over the K=7 neighborhood of a test point, while the counterfactual distance in Eq. (7) is the distance to the nearest different-class neighbor in the same training set, found with the same KNN classifier at K=7 (Appendix C). A test point that sits close to a different-class neighbor will mechanically tend to have high class conflict in that neighborhood, so the −0.83 correlations in Table 5 are at least partly a restatement of shared neighborhood structure. The independent estimators do not rescue it: AUens goes positive on Glass and Ionosphere, and AUcen is near zero on Ionosphere. The paper even concludes that \"more local uncertainty quantification should be privileged,\" which effectively leans on the entangled estimator.\n\nOther soft spots: the protocol is never benchmarked against a fixed explanation policy, so the correlations don't demonstrate that the routing improves user-facing outcomes; the SHAP experiment uses only 5 runs; the rejection evidence is anecdotal (one instance per setting); and code is promised only on acceptance. These are fixable.\n\nThat said, the paper is honest about some limits, the related work is well covered, and the central idea—using uncertainty to route explanations—is worth taking seriously. I'd send it to review, but flag the circularity question to the referees. If the authors re-run the counterfactual correlation with a distance computed outside the KNN structure (or with a separate held-out neighborhood), the claim would stand on firmer ground.","headline":"A plausible routing protocol for explanation choice, but the headline counterfactual correlation may be baked into the shared KNN machinery; worth refereeing with revision.","tokens_in":18400,"tokens_out":2016,"would_cite":false,"duration_ms":22390,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the type of prediction uncertainty tells you which explanation to offer: counterfactuals for high aleatoric uncertainty, feature importance for low, and an 'insufficient training' rejection for high epistemic…","keywords":["uncertainty decomposition","aleatoric uncertainty","epistemic uncertainty","counterfactual explanations","feature importance","SHAP","explanation robustness","reject option"],"falsifier":"Compute AUbel and counterfactual distance on the same datasets but with different K values for the two computations (e.g., K=3 for the belief function and K=15 for the counterfactual search); if the negative Spearman correlation is robust, it should persist beyond a shared K=7 neighborhood. A second check is to construct a dataset with high class overlap but distant different-class neighbors, where the predicted negative correlation should vanish.","tokens_in":17302,"feed_emoji":"🧭","tokens_out":6206,"duration_ms":60591,"temperature":0.7,"pith_summary":"The paper tries to establish that uncertainty decomposition can be used as a practical selector among explanation methods. The central proposal is to split prediction uncertainty into aleatoric (data noise/class overlap) and epistemic (lack of training data) parts; high epistemic uncertainty means no post-hoc explanation is trustworthy, and the uncertainty itself should be shown as an explanation: the model lacks training. For the remaining points, high aleatoric uncertainty points to counterfactual explanations, because class-confused regions tend to yield close, attainable counterfactuals, while low aleatoric uncertainty points to feature-importance explanations such as SHAP, which are more locally stable where classes are separated. If the correlations reported hold, explanation systems become context-aware rather than one-size-fits-all.","feed_headline":"Uncertainty type tells you which AI explanation to trust","feed_subtitle":"Epistemic uncertainty is itself an explanation: 'the model lacks training'; aleatoric uncertainty picks the method.","key_machinery":"The load-bearing objects are two uncertainty estimates derived from a Dempster-Shafer belief function on the K-nearest neighbors: epistemic uncertainty EUbel (non-specificity, Equation 12) and aleatoric uncertainty AUbel (discord, Equation 13). Around these sit the counterfactual defined as the nearest training point with a different predicted label (Equation 3), SHAP feature attributions (Equation 2), and two robustness metrics: discrete local Lipschitz continuity for feature-importance explanations (Equation 4) and Euclidean distance to the counterfactual as a measure of attainability (Equation 7). The argument runs by Spearman rank correlations between AUbel and these robustness scores, with the rejection rule driven by EUbel.","core_discovery":"On its own terms, the paper's discovery is a pair of empirical regularities plus a rejection rule. Across ten tabular datasets, the belief-based aleatoric uncertainty AUbel is significantly negatively correlated with counterfactual dissimilarity (Spearman's r roughly -0.45 to -0.89), meaning that in high-aleatoric regions the nearest different-class neighbor is closer and more attainable; and AUbel is positively correlated with SHAP un-robustness, meaning low-aleatoric regions give more locally stable feature-importance explanations. The paper further claims that epistemic uncertainty is itself a legitimate explanation: when it is high, the instance is out-of-distribution or under-represented, so the correct statement is that the model lacks training, and the usual explanation should be withheld. These findings hold across three uncertainty decomposition families and carry over to a deep learning MNIST example, where the most epistemically uncertain test image yields an unintelligible counterfactual that should be rejected.","pith_inferences":["The paper leaves implicit that the same selector could be applied adaptively per instance at deployment time, for example a dashboard that switches explanation type based on AUbel.","Because AUbel and the counterfactual both come from the same K-nearest-neighbor structure, some of the observed correlation may be mechanical; varying K between the two computations would separate genuine data effects from shared-machinery artifacts.","A testable extension is to use the epistemic reject option as an acquisition criterion: instead of explaining uncertain points, query them for labels, since high EUbel marks regions where additional training data would most improve explanations.","A stricter test of the paper's claim would compare the recommended explanation type on the same low-AU and high-AU points against user-satisfaction or decision-quality measures, not just robustness metrics."],"forward_implications":["A deployed explanation system can use AUbel directly: when it is high, show the counterfactual; when it is low, show feature attributions.","High EUbel becomes a signal to withhold explanations and instead state that the model has insufficient evidence, turning explanation failure into a transparent outcome.","Local, density-based uncertainty estimates (belief-based and centroid-based) outperform global ensemble-entropy estimates as selectors, so more local uncertainty quantification is worth investing in.","The reject option gives a principled way to clean test sets before reporting SHAP values, since filtering the 30% most epistemically uncertain points often strengthens the AU-robustness correlation.","The protocol transfers from tabular classifiers to deep image classifiers, so the same uncertainty-decomposition guidance can be applied in deep learning settings."],"supporting_citations":[{"why":"Supplies the ensemble entropy decomposition used to compute AUens and EUens in Equations (8)-(9).","marker":"Shaker & Hüllermeier, 2020"},{"why":"Provides the centroid-based certainty measure whose inverse gives EUcen.","marker":"Van Amersfoort et al., 2020"},{"why":"Provides the evidential K-NN setting under which AUbel and EUbel are computed locally.","marker":"Hoarau et al., 2024"},{"why":"Defines the belief-function K-NN rule that produces the mass function mt used in Equations (12)-(13).","marker":"Denœux, 1995"},{"why":"Defines the discord measure used as aleatoric uncertainty AUbel in Equation (13).","marker":"Klir & Ramer, 1990"},{"why":"Defines non-specificity used as epistemic uncertainty EUbel in Equation (12).","marker":"Dubois & Prade, 1987"},{"why":"Defines SHAP feature attributions used for the feature-importance robustness experiment.","marker":"Lundberg & Lee, 2017"},{"why":"Introduces the discrete local Lipschitz continuity metric used to measure explanation robustness in Equation (4).","marker":"Alvarez-Melis & Jaakkola, 2018"}],"fun_headline_variants":["Uncertainty type determines which AI explanation to trust","Epistemic uncertainty is itself an AI explanation","Aleatoric uncertainty chooses explanation method","Uncertainty-aware selection improves explanation robustness","How uncertainty breakdown picks and rejects AI explanations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim collapses if the negative AU-vs-counterfactual-attainability and positive AU-vs-SHAP-unrobustness correlations are largely artifacts of both quantities being computed from the same K-nearest-neighbor structure, rather than reflections of genuine data geometry.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty type determines which AI explanation to trust","Epistemic uncertainty is itself an AI explanation","Aleatoric uncertainty chooses explanation method","Uncertainty-aware selection improves explanation robustness","How uncertainty breakdown picks and rejects AI explanations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000219,"raw_usage":{"total_tokens":1409,"prompt_tokens":879,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":462}},"tokens_in":495,"tokens_out":530,"duration_ms":5885,"temperature":1.0,"reasoning_tokens":462,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:35:45.162152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute AUbel and counterfactual distance on the same datasets but with different K values for the two computations (e.g., K=3 for the belief function and K=15 for the counterfactual search); if the negative Spearman correlation is robust, it should persist beyond a shared K=7 neighborhood. A second check is to construct a dataset with high class overlap but distant different-class neighbors, where the predicted negative correlation should vanish.","supporting_citations":[{"cited_title":"A random forest with 100 estimators and a maximum depth of 4 is used","cited_arxiv_id":null,"evidence_quote":"Supplies the ensemble entropy decomposition used to compute AUens and EUens in Equations (8)-(9)."},{"cited_title":"This result is highlighted in Table","cited_arxiv_id":null,"evidence_quote":"Provides the evidential K-NN setting under which AUbel and EUbel are computed locally."}],"review_version":1}