{"id":"7c455254-ec52-48bb-8935-bc124587093d","arxiv_id":"2505.10399","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"AXE evaluates local explanations by k-NN predictiveness of top features, requiring neither ground-truth explanations nor model sensitivity.","lead":"This paper introduces AXE, a metric that scores local feature-importance explanations by how accurately a k-nearest-neighbor model can recover the explained model's prediction using only the top-ranked features. It proposes three principles for explanation evaluation and reports that AXE can detect fairwashed explanations without needing ground-truth explanations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 as written trains the k-NN model on the very point it is then asked to predict; without a leave-one-out step, AXE's discriminative scores and fairwashing results are not established.","rationale":"The paper proposes a genuinely useful direction: evaluating local feature-importance explanations by how well the top-n features predict model outputs, without relying on ground-truth explanations or sensitivity-based perturbations. The three principles are reasonable, and the synthetic illustrative example in Section 3.2 is intuitive. However, the single most load-bearing issue is the train/query overlap in Algorithm 1. If the described procedure were executed literally, AXE would score explanations largely by whether the explained point's own label appears in its nearest-neighbour set, not by whether the top-n features generalise across the data manifold. The reported experimental values, especially the near-perfect protection-feature scores in Table 2, imply an unstated leave-one-out step. This is not a disagreement with external consensus; it is an internal specification problem that affects all empirical claims. The fix is small and addressable: document and implement leave-one-out in Algorithm 1, and then re-run the experiments with this clarification. If the code already contains leave-one-out, the manuscript text must be corrected to match. The reader's conditional verdict is therefore appropriate, and I do not see a basis for moving to accept or reject. A secondary concern, not used as the headline, is that k-NN predictiveness is asserted but not validated against human simulatability; that would strengthen the paper but is less immediate than the algorithmic leakage.","tokens_in":19914,"tokens_out":5858,"duration_ms":63334,"concrete_test":"Run the released code on the Section 4.1 fairwashing scenarios twice: (a) exactly as Algorithm 1 is written, and (b) with a leave-one-out step that removes x_i from the training set of M_k^i before querying. Compare AXE(E_rho), AXE(E_phi), AXE(E_psi), and the rankings. If variant (a) yields near-ceiling scores or fails to separate E_rho from E_phi, the published algorithm is not the one that produced the paper's 100% detection claim, and the manuscript must state and implement the leave-one-out variant.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing correctness risk is in Algorithm 1 (Section 3.1). Line 5 constructs X_f from the full dataset X, which includes x_i, and line 6 queries M_k^i on x_i. For k=1, the nearest neighbour of x_i is necessarily x_i itself, so every explanation receives accuracy 1.0; for small k, the self-neighbour systematically inflates scores and compresses differences between explanations. The paper never states that x_i is excluded from the training set, yet the reported results require that exclusion. The fairwashing table (Table 2) shows AXE(E_rho) > AXE(E_phi), which cannot be reproduced by the literal algorithm for small k. Because the central claim is that AXE measures how well an explanation recovers model behaviour from the data, this train/query overlap must be fixed before the empirical claims can be assessed. This concern is independent of whether the k-NN proxy is a valid model of human simulatability; even under the paper's own predictiveness definition, the metric as written does not measure generalisation of the top-n features.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that existing evaluation frameworks for local feature-importance explanations are either ground-truth-dependent or sensitivity-based, and proposes three principles---local contextualization, model relativism, and on-manifold evaluation---for a desirable evaluation framework. It then introduces AXE, a metric that scores an explanation by training a k-NN model on the top-n features identified by that explanation and measuring how accurately the model prediction can be recovered. The paper motivates AXE with a synthetic example, applies it to detect explanation fairwashing, and compares it with prior metrics on OpenXAI datasets. It claims that AXE is the first evaluation metric to detect fairwashing perfectly on the tested suite.","tokens_in":20149,"tokens_out":6134,"duration_ms":63563,"significance":"If the central claim holds, AXE is a genuinely useful contribution: it provides a ground-truth-agnostic, on-manifold way to compare local explanations, with a clear operationalization of the idea that good explanations help users predict model behavior. The paper is clearly written, ships code, and includes both synthetic examples and real-data comparisons. The main caveats are the unstated leave-one-out protocol in Algorithm 1, the lack of statistical reporting in the fairwashing experiment, and the unvalidated k-NN proxy for human simulatability. These issues are fixable, and the core idea is worth publishing after the experimental protocol is corrected.","major_comments":[{"comment":"As written, Algorithm 1 trains the k-NN model M_k^i on X_f constructed from the full dataset X, which includes the explained point x_i, and then queries M_k^i at x_i. For k=1 the nearest neighbor is x_i itself, so every explanation receives an AXE score of exactly 1; for small k, the self-neighbor systematically inflates all scores and compresses differences between explanations. The illustrative example in Section 3.2 (q(e_b) ~ 0.5 for a wide range of k) and the fairwashing results in Table 2 cannot be reproduced by the literal algorithm: with self-inclusion, the E_phi and E_psi explanations would also receive near-perfect scores from their own points. The paper never states a leave-one-out step. Because the central claim is that AXE measures how well the top-n features generalize in recovering model behavior, the algorithm must exclude x_i from the training set for its own prediction (or equivalently exclude the zero-distance self-match at query time), and the reported experiments must be rerun under that protocol.","section":"Section 4.1, Table 2"},{"comment":"The claim that AXE is 'the first evaluation metric to be able to do this perfectly' is not supported by the experimental reporting. Table 2 contains a single run per dataset/model/foil configuration, with no error bars, no number of repetitions, and no statement of the AXE hyperparameter k used (only n=1 is fixed). The PGI/PGU baseline numbers are also reported without the perturbation-width hyperparameter, which Section 3.2 itself shows to be decisive for these metrics. I therefore cannot assess whether the 100% success rate is robust or an artifact of a favorable hyperparameter choice or of the train/query overlap in Algorithm 1. The authors should report standard errors or confidence intervals, the k value(s) used, and the PGI/PGU width(s), and should recompute the table after fixing Algorithm 1.","section":"Section 4.1, Table 2"},{"comment":"The load-bearing premise of AXE is that a good local explanation is one whose top-n features allow a k-NN model to recover the model's prediction. The cited user studies support the broader idea that useful explanations help people predict model behavior, but they do not establish that k-NN accuracy is a faithful proxy for human simulatability; in particular, k-NN ignores feature magnitude, sign, and human cognitive factors. If AXE is proposed as a formal operationalization, the paper should state this more carefully and provide at least a small human-subject validation or a clear argument for why the proxy is sufficient. Without this, the empirical comparisons only show that AXE agrees with the authors' chosen synthetic and fairwashing constructions, not that it tracks explanation quality in the human-centered sense invoked in the introduction.","section":"Section 3.1"}],"minor_comments":[{"comment":"The entry '0-0.524' appears to be a typo for '-0.524'; please fix.","section":"Table 2, German Credit m_L (2 foils) PGU row"},{"comment":"The sentence 'For a dataset with ν features' should read 'For a dataset with N features and ν datapoints'; the notation is inconsistent and confusing.","section":"Section 3.1, cache-size discussion"},{"comment":"The notation AXE1, AXE3, AXE5, AXE9 is ambiguous: the paper uses AXE_k^n elsewhere, and the reader cannot tell whether the subscript refers to k or n. Please make the notation consistent.","section":"Section 4.2, Figures 6 and 7"},{"comment":"The caption misspells 'reliably' as 'relibably'.","section":"Section 3.2, Figure 5"},{"comment":"The illustrative example treats 'average near 0.5' as poor accuracy, but Algorithm 1 uses hard accuracy on predicted labels; please clarify the k-NN decision rule (majority vote versus averaged score) and the tie-breaking procedure.","section":"Section 3.2, k-NN decision rule"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a reasonable fit for the venue and the core idea is publishable after the train/query overlap is fixed and the experiments are rerun with proper statistical reporting. The central flaw in Algorithm 1 is local and fixable, so I recommend major revision rather than rejection. The authors should also temper the 'first metric to do this perfectly' claim unless error bars support it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: this is a real contribution to the XAI evaluation literature. The core idea—score a local explanation by how well the top-n features recover the model's prediction using a k-NN model trained on the data—is new relative to prior predictiveness work (Colin et al., Hooker et al.) and it cleanly sidesteps the two standard crutches of ground-truth and sensitivity. The three principles in Section 2 are a useful synthesis, and the simple 2D illustration in Section 3.2 does genuine work: it shows why PGI needs perturbation-width luck while AXE gives a stable ranking. The fairwashing experiment, though small, is a smart stress test, and AXE's perfect separation there is notable.\n\nThe soft spot is real and load-bearing. Algorithm 1 as written trains the k-NN on the full dataset X and then queries it on the same x_i. With k=1 the nearest neighbor of any point is itself, so every explanation gets accuracy 1.0; with small k the self-example inflates all scores and compresses differences. The illustrative example and Table 2 are only reproducible if the explained point is excluded from the training set, but the paper never says that. This is not a philosophical objection—it's an algorithmic bug in the description, and it has to be fixed before the numbers mean anything. The fix is easy: build X_f from X \\ {x_i}, or use leave-one-out cross-validation. But until it is stated, the empirical claims are unsubstantiated.\n\nThe other concerns are milder. The k-NN proxy is motivated by user studies but never validated against human simulatability; I'd want at least a small human or behavioral check before treating AXE as a measure of 'usefulness.' The fairwashing table has no error bars or hyperparameter reporting, and the 100% claim is over the specific attack suite, not a general guarantee. Those are minor relative to the leakage issue.\n\nBottom line: the idea deserves a serious referee. With the leave-one-out step spelled out and some error bars, this could be a standard evaluation tool for tabular XAI. I'd send it to review with a request for revision. I'd cite it once the algorithm is corrected, not before. For a reading group, yes—it generates good discussion about what 'predictiveness' should mean.","headline":"AXE is a genuinely new and promising ground-truth-free evaluation idea, but Algorithm 1 leaks the explained point into the k-NN training set; the empirical claims hold only after a leave-one-out fix is stated and implemented.","tokens_in":20656,"tokens_out":2364,"would_cite":false,"duration_ms":21311,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"AXE scores explanations by whether their top features recover the model's prediction, with no ground truth needed.","keywords":["explainable AI","feature importance explanations","explanation evaluation","ground-truth agnostic","k-nearest neighbors","fairwashing detection","local explanations","model faithfulness"],"falsifier":"A direct test is a user study: give humans the same top-n features and ask them to predict the model's output, then compare their success rate with AXE scores; if the two rankings diverge on real datasets, the predictiveness proxy is not the useful notion. A cheaper computational check is to rerun Algorithm 1 exactly as written, without excluding the explained point from the k-NN training set, since with that point present distance-based recovery becomes trivially easy and would inflate every explanation's score.","tokens_in":19726,"feed_emoji":"🔍","tokens_out":5432,"duration_ms":54204,"temperature":0.7,"pith_summary":"Competing local explanations of the same prediction often contradict one another, and existing evaluation metrics either need a \"ground-truth\" explanation nobody can supply or reward sensitivity to off-manifold perturbations. This paper argues that neither approach is adequate and proposes three principles an evaluation should satisfy: explanations should depend on the datapoint and the model, and should not depend on off-manifold model behavior. To meet them, it introduces AXE, a score that measures whether the top-n features named by an explanation can be used by a k-nearest-neighbor model to recover the model's prediction. The paper reports that AXE ranks a set of real explanations coherently, agrees across its own hyperparameters, and detects adversarially \"fairwashed\" explanations—explanations that hide a protected attribute's role—in every one of ten tested settings, a result the authors attribute to AXE's avoidance of sensitivity-based assumptions.","feed_headline":"K-NN check scores explanations without a ground truth","feed_subtitle":"A ground-truth-free score that catches fairwashed explanations in all ten tested cases.","key_machinery":"The carrying object is a per-datapoint k-nearest-neighbor accuracy check on the explanation's top-n features. For each explained point, AXE extracts the n features with largest absolute importance, fits a separate k-NN model to the model's predictions over the dataset using only those features, and records whether that model reproduces the explained point's prediction; the dataset-level quality score is the average accuracy, and an area-under-curve over n turns the top-n scores into a single scalar. The deliberate separation of k-NN models per datapoint is what makes the score local, and training on the dataset's own points is what keeps it on-manifold.","core_discovery":"The paper's central claim is that explanation quality for local feature-importance methods can be measured without any ground-truth explanation and without sensitivity analysis. The proposed measure, AXE, takes a datapoint, its model prediction, and a candidate explanation, keeps only the features the explanation marks most important, and asks whether a k-NN model trained on those features over the dataset recovers the model's prediction. A good explanation is one whose top features separate model predictions in feature space. The paper shows that this operationalizes the user-study finding that useful explanations help users predict model behavior, and it argues AXE satisfies local contextualization, model relativism, and on-manifold evaluation. In fairwashing experiments built on an adversarial attack, AXE scored truthful explanations above manipulated ones in all ten settings, which the paper states is the first evaluation metric able to do so.","pith_inferences":["If the proxy holds up in human tests, AXE-style scores could become a cheap selection criterion in model documentation and explanation dashboards, something the paper motivates but does not itself demonstrate.","The paper's pseudocode omits a leave-one-out step; unless the released implementation excludes the explained point when training the k-NN model, the reported near-perfect scores could be an artifact of self-matching, so the code should be checked before relying on the numbers.","The same k-NN recovery idea extends naturally to regression targets by replacing classification accuracy with a continuous agreement measure, and to explanation types beyond top-n feature sets, such as rules or counterfactuals.","A testable consequence of the paper's account is that high-AXE explanations should be the ones human users find most useful in prediction tasks; if user studies show no such link, the metric would need to be recalibrated rather than abandoned."],"forward_implications":["Practitioners can rank competing local explanations for any tabular model using only the dataset and model outputs, with no oracle explanation.","AXE can serve as an audit signal: explanations that hide a protected attribute's contribution receive lower scores than explanations that surface it, so a regulator could flag fairwashed explanations.","For neural networks, where coefficient-based ground truths do not exist, AXE provides a comparison metric that the paper shows is stable across its own hyperparameters.","The three principles give later evaluation methods a checklist: local contextualization, model relativism, and on-manifold evaluation.","Because AXE measures on-manifold predictiveness rather than sensitivity, it avoids rewarding explainers that merely match a particular perturbation scheme."],"supporting_citations":[{"why":"Supplies the user-study finding that useful explanations help users predict model behavior, which AXE operationalizes.","marker":"[12]"},{"why":"Provides the benchmark, datasets, and baseline metrics that the paper compares AXE against.","marker":"[2]"},{"why":"Defines a perturbation-based explainer whose sensitivity assumptions AXE is designed to avoid.","marker":"[48]"},{"why":"Defines the other main perturbation-based explainer used in the comparison and fairwashing experiments.","marker":"[36]"},{"why":"Supplies the adversarial fairwashing attack and the datasets and models used to test whether a metric can detect manipulated explanations.","marker":"[54]"},{"why":"Documents disagreement among ground-truth-based metrics and is the source of those metric definitions.","marker":"[31]"},{"why":"Provides the k-NN classification method that AXE uses as its predictive-recovery mechanism.","marker":"[13]"}],"fun_headline_variants":["AXE scores explanations without any ground truth","k-NN based AXE catches fairwashing in every test","Explanation quality measured sans ground truth or sensitivity","AXE: the ground-truth-free score that exposes fairwashing","No gold standard needed: AXE judges explanation quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"AXE's entire validity rests on the premise that a local explanation is good exactly when the model's prediction at that point can be recovered by a k-NN model built on the dataset using the explanation's top-n features, a proxy for human usefulness that the paper does not test against human behavior.","fun_headline_variants_meta":{"raw":{"variants":["AXE scores explanations without any ground truth","k-NN based AXE catches fairwashing in every test","Explanation quality measured sans ground truth or sensitivity","AXE: the ground-truth-free score that exposes fairwashing","No gold standard needed: AXE judges explanation quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1510,"prompt_tokens":868,"completion_tokens":642,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":565}},"tokens_in":484,"tokens_out":642,"duration_ms":6496,"temperature":1.0,"reasoning_tokens":565,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:09:31.239056+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is a user study: give humans the same top-n features and ask them to predict the model's output, then compare their success rate with AXE scores; if the two rankings diverge on real datasets, the predictiveness proxy is not the useful notion. A cheaper computational check is to rerun Algorithm 1 exactly as written, without excluding the explained point from the k-NN training set, since with that point present distance-based recovery becomes trivially easy and would inflate every explanation's score.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the user-study finding that useful explanations help users predict model behavior, which AXE operationalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the benchmark, datasets, and baseline metrics that the paper compares AXE against."},{"cited_title":"Lundberg and Su-In Lee","cited_arxiv_id":null,"evidence_quote":"Defines the other main perturbation-based explainer used in the comparison and fairwashing experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the adversarial fairwashing attack and the datasets and models used to test whether a metric can detect manipulated explanations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents disagreement among ground-truth-based metrics and is the source of those metric definitions."}],"review_version":1}