{"id":"15f35948-4dc6-4d67-a2c1-a2101e35057f","arxiv_id":"2509.00884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A Gaussian-process autoencoder with a latent-space density estimator generates counterfactual examples for tabular data, with competitive or better scores on several evaluation metrics.","lead":"The paper builds an autoencoder from Gaussian processes with random Fourier features, and uses it to generate counterfactual explanations for tabular classifiers. A learned density estimate in the latent space guides the search so the produced counterfactuals stay close to the data distribution.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validity is not entailed by the boundary constraint: Eq. (18) places counterfactuals at p(target)=0.5, yet the paper reports validity≈0.99 without any crossing mechanism.","rationale":"The reader's weakest assumption is the same as mine: enforcing equality to the decision boundary does not, by itself, flip a sigmoid classifier's label. This is the most load-bearing concern because without a label flip the method does not produce counterfactuals at all, and the reported validity near 0.99 is otherwise unexplained. I considered other candidates: the factor-2 error in Eq. 22 is a concrete bug but does not undermine the conceptual validity claim as directly; the non-class-conditional density estimator weakens the 'in-distribution of the target class' claim but could be interpreted as overall in-distribution. The validity/boundary gap is central. The reader's CONDITIONAL verdict already reflects this concern, so my stress test does not move the verdict; it makes the conditionality concrete and specific. The other issues (factor-2 gradient, missing SAE in masked tables, absent code) are secondary and independently fixable.","tokens_in":25671,"tokens_out":6887,"duration_ms":84499,"concrete_test":"Re-run the GPAE counterfactual search on GMC (or LCD) with the described algorithm; for each query, record the final signed logit ℓ = θ_c^T f_e(x+δ)+θ0 and compute Validity using the same explained classifier with a strict threshold (e.g., target class iff ℓ>0). If the median final logit is near 0 while Validity remains ≈0.99, the reported validity hinges on tie-breaking or an unreported post-processing step and the central claim is unsupported. If the median logit is clearly on the target side, identify which part of the algorithm produces that crossing (e.g., a step in Eq. 25 or an inequality) and document it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central link between the optimization in §4.5 and the main claim of generating valid counterfactuals is broken. The explained classifier is the encoder's linear logit followed by sigmoid (Algorithm 2, step 8: ŷ_i = sigmoid(λ_i^T θ_c)). In §4.5 the search is constrained by θ_c^T f_e(x+δ)=0 (Eq. 18; the Lagrangian in Eq. 19 also includes θ0). This forces the final logit to 0, i.e., p(target)=0.5. For a hard label flip the logit must be strictly on the target side of 0; a point exactly on the boundary is either not label-flipped or is flipped only by an arbitrary tie-break. The objective (17) has no term pushing the logit past the boundary—the density term acts only along/beside the boundary—so the reported Validity≈0.99 (Table 5, e.g., GMC and FICO rows) is not a consequence of the stated optimization. The authors themselves use step=2 for SAE to move beyond the boundary (§4.1), acknowledging that equality is insufficient; GPAE has no analogous step or inequality constraint. Until this is resolved, the empirical validity numbers are unexplained.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GPAE, a Gaussian-process autoencoder with random Fourier feature (RFF) approximations for generating counterfactual explanations on tabular data. The encoder/decoder are linear maps on RFF features; the latent space is supervised by a sigmoid classifier; and a learned density estimator is used in a constrained search over input perturbations to find counterfactuals that are close, diverse, and in-distribution. A separate procedure selects the density regularization strength by minimizing a KL divergence between projected data and the estimated density on the decision boundary. Experiments on five tabular datasets compare GPAE with logistic regression, Wachter, VAE-based methods, CounterNet, SAE, and a GP-inducing-point variant, reporting classification accuracy and counterfactual metrics including L2, diversity, instability, discriminative power, IM1/IM2, and validity. The paper claims that GPAE achieves a balanced performance with fewer learnable parameters and generates diversified, in-distribution counterfactuals.","tokens_in":26030,"tokens_out":6478,"duration_ms":84775,"significance":"If the validity mechanism were fully established, the paper would make a useful contribution: the RFF-based GP autoencoder is a lightweight alternative to neural autoencoders, the density-guided search is a principled way to encourage in-distribution counterfactuals, and the mask mechanism for immutable features is practically relevant. The empirical study is broad, uses standard benchmarks, and evaluates against several external baselines; IM1/IM2 are computed with independently trained autoencoders, which partially mitigates self-referentiality of the in-distribution claim. However, the central derivation connecting the optimization in §4.5 to the main claim of generating valid counterfactuals is incomplete: the stated equality constraint places counterfactuals at the classifier's decision boundary, which does not, in general, flip the predicted label. The reported validity values near 0.99 are therefore unexplained by the stated optimization. This gap, together with a factor-2 error in the reported gradient and a class-conditional gap in the density estimator, prevents the paper from being acceptable in its current form. The contributions are potentially salvageable with a corre","major_comments":[{"comment":"The claim that the boundary constraint 'naturally makes the counterfactual sample valid' is not justified. With the sigmoid classifier in Algorithm 2 (ŷ_i = sigmoid(λ_i^T θ_c)), the constraint θ_c^T f_e(x+δ)=0 gives p(target)=0.5. Under the standard threshold rule this is a tie, not a label flip from the original side; for queries whose original class is the positive class, the boundary point is still classified as the original class. Unlike the SAE baseline in §4.1, which explicitly uses step=2 to move beyond the boundary, Eq. (17) contains no term that pushes the logit past zero, and Eq. (18) is an equality, not an inequality. The reported Validity≈0.99 in Table 5 (e.g., GMC and FICO rows) is therefore not a consequence of the stated optimization. The authors should replace Eq. (18) with a strict inequality or margin constraint (e.g., θ_c^T f_e(x+δ) ≥ ε or ≤ −ε for the target side), ad","section":"§4.5, Eqs. (18)-(19) and Table 5"},{"comment":"The gradient in Eq. (22) has a factor-2 error in the Gaussian term of the density estimator. Differentiating the log-density term in Eq. (20), −β[w^T φ(f_e) − ½ f_e^T Σ^{-1} f_e + μ^T Σ^{-1} f_e], with respect to f_e gives −β[(∂φ)^T w − Σ^{-1} f_e + Σ^{-1} μ]. The coefficient of the μ^T Σ^{-1} term in Eq. (22) should therefore be 1, not 2. This changes the balance between the distance and density forces in the counterfactual search. The authors should correct Eq. (22) and confirm which gradient was actually implemented in the experiments.","section":"§4.5, Eq. (22) vs. Eq. (20)"},{"comment":"The density estimator in Eqs. (12)-(14) is learned from all N observed latent codes without conditioning on class, and the search objective (17) uses this global density. The third Highlight states that the density estimator 'direct[s] the generated samples to be in-distribution of the target class,' but the construction does not target the target class specifically; it can equally favor the original class density near the boundary. The IM1/IM2 results suggest empirically that the generated counterfactuals lie near the target manifold, but this is not entailed by the stated density objective. The authors should either train a class-conditional density estimator, restrict the density model to target-class samples, or provide a clear argument for why the global density suffices.","section":"§4.4-§4.5 and Highlights"},{"comment":"The β-selection criterion minimizes KL(q(γ) || p(γ)), where p(γ) is the same parametric density estimator used in the search objective. This makes the 'optimal β' claim partly self-referential: the criterion can only measure how well the counterfactual distribution matches the model's own density estimate, not how well that estimate matches the true data density. The external IM1/IM2 metrics mitigate the impact, but the text should be revised to state that β is selected by internal density matching, and ideally the sensitivity of the final counterfactual metrics to β should be reported.","section":"§4.6, Eqs. (28)-(31)"},{"comment":"The optimization algorithm is under-specified. The paper states that δ and η are optimized 'iteratively and interchangeably,' but does not give the update rule for η, the stopping criterion, or whether the equality constraint is enforced exactly at convergence. Given that the validity claim depends entirely on the final point satisfying the boundary condition, the manuscript should specify the primal-dual updates, the convergence tolerance, and how the final counterfactual is projected/rounded. This is needed to reproduce the reported validity numbers.","section":"§4.5, Eq. (22)-(25)"}],"minor_comments":[{"comment":"There are numerous typos and formatting errors: 'Fgiure 4', 'W ac.', 'GP AE', 'Disucssion', 'Futhermore', 'high-stack fields', 'SA W', 'We All experiments', and 'the IM2 and IM2' in the caption of Figure 6. A careful proofread is needed.","section":"Throughout"},{"comment":"The density estimator notation is inconsistent: Eq. (12) uses a scalar σ, while Eq. (20) uses a covariance matrix Σ. Please unify the notation and define the Gaussian envelope precisely.","section":"Eq. (12)"},{"comment":"The validity metric in Eq. (35) uses the explained classifier's predicted label. Since the paper's own classifier is a sigmoid, the tie-breaking convention at p=0.5 should be stated; otherwise the validity numbers are ambiguous.","section":"§5.4, Table 5"},{"comment":"Step 8 writes ŷ_i = sigmoid(λ_i^T θ_c), while Eq. (10) defines the logit as λ_i^T θ_c and Eqs. (19)-(24) also include a bias θ0. Please clarify whether the classifier has a bias term and, if so, include it consistently in the constraint and updates.","section":"Algorithm 2"},{"comment":"The description of baselines is brief. In particular, the text says CARLA default parameters are used for all methods, but it is not stated whether the explained classifier is the same for all baselines and whether the same immutable-feature masking protocol is applied to methods that do not natively support it. This information is important for the fairness of the comparison.","section":"§5.2"},{"comment":"The ablation text says 'all the other metrics improve significantly' while Figure 6 shows Diversity decreasing; the wording should be adjusted to acknowledge the trade-off.","section":"§5.6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the journal's scope and the empirical study is reasonably thorough. However, the validity derivation in §4.5 is a load-bearing gap: the stated equality constraint does not explain the near-perfect validity scores. The factor-2 gradient error and the class-conditional density issue are additional correctness concerns. I believe the central idea is salvageable, but the authors need to fix the validity constraint, correct the gradient, and clarify the density objective. I would be willing to review a revised version. No concerns about novelty or citation behavior beyond the need to better position the contribution relative to the authors' prior SAE work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is worth a look: an RFF-based GP autoencoder with a supervised latent space, a learned latent density estimator, and a KL-guided beta search for counterfactual generation. The components are mostly known, but the combination is new, and the empirical results are competitive across five tabular datasets. The paper is also honest about its own limitations in the discussion, and the external IM1/IM2 evaluation with independently trained autoencoders is a good check against pure self-referential density. Credit where due: the ablation on latent dimension and RFF size is useful, and the comparison set is broad.\n\nThat said, the central validity claim has a real hole. The optimization in Section 4.5 enforces equality on the decision boundary, but the classifier is sigmoid-headed, so a logit of zero gives p=0.5, not a label flip. The paper even uses step=2 for the SAE baseline to pass beyond the boundary, which is precisely the mechanism GPAE lacks. The reported Validity near 0.99 is therefore unexplained by the stated optimization. That is not a minor typo; it is the load-bearing link between the objective and the main claim. The factor-2 error in Eq. (22) versus Eq. (20) is real but less damaging; it changes the gradient magnitude, not the conceptual argument. The density estimator also appears to be trained on all classes while the abstract claims target-class in-distribution generation; the mismatch should be cleaned up.\n\nThe beta-selection algorithm minimizing KL against the model's own density estimate is somewhat self-referential, but it is a heuristic for picking a regularization parameter, and the paper already tempers this by evaluating with external metrics. That is acceptable if the rest holds. The omission of SAE from masked evaluations is annoying but not fatal; the unmasked table includes it. No code or full hyperparameters means I cannot reproduce, which matters more here because the validity mechanism is already unclear.\n\nThis paper is for readers working on counterfactual generation for tabular data, especially in recourse and fairness settings. It deserves a serious referee, but it needs substantive revision: justify the boundary-validity mechanism or change the constraint to an inequality, fix the gradient, clarify the density estimator's class conditioning, and release code. I would not cite it in its current form, but I would put it on the reading list for a lab meeting on counterfactual autoencoders.","headline":"A promising but not-yet-trustworthy counterfactual method: the architecture and experiments are real, but the validity claim rests on a broken constraint, and the missing code/baselines make it hard to verify.","tokens_in":26477,"tokens_out":886,"would_cite":false,"duration_ms":13375,"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":"GPAE builds an autoencoder from Gaussian processes with random Fourier features and uses a latent density estimator during counterfactual search, reporting balanced, in-distribution counterfactuals on five tabular datasets.","keywords":["Gaussian process","random Fourier features","autoencoder","counterfactual explanations","algorithmic recourse","density estimation","tabular data","interpretable machine learning"],"falsifier":"On a tabular dataset, train GPAE and, for a held-out query, record counterfactuals at the converged solution of the constrained search, then evaluate the explained classifier's argmax label at those counterfactuals. If a substantial fraction are not assigned the target class, then the boundary constraint theta_c^T f_e(x+delta)=0 does not by itself make the sample valid, and the reported ~0.99 validity must be attributed to the density-regularized optimization rather than the equality constraint.","tokens_in":25583,"feed_emoji":"🤖","tokens_out":8501,"duration_ms":98411,"temperature":0.7,"pith_summary":"The paper proposes an autoencoder built from Gaussian processes, approximated with random Fourier features, and a latent linear classifier that supervises the encoder. On top of this, it adds a learned density estimator in the latent space and uses it during counterfactual search: the search minimizes the change to the input while staying on the classifier's decision boundary and inside high-density regions of the target class. The paper reports that this setup, called GPAE, produces counterfactual samples that are simultaneously close to the query, diverse, stable, and in-distribution across five tabular datasets, while needing far fewer trainable parameters than neural-autoencoder baselines. A beta-selection algorithm chooses the strength of the density term by matching the counterfactual distribution to the data projected on the boundary. The practical payoff is an explainable-AI recipe for recourse: what to change in an application to flip an outcome, with immutable features fixed by a mask.","feed_headline":"GP autoencoder yields in-distribution counterfactuals","feed_subtitle":"Latent density guides the search for small, plausible input changes; five tabular datasets show balanced results.","key_machinery":"The carrier of the argument is the random-Fourier-feature Gaussian process autoencoder coupled to a latent density estimator. The RFF map phi(x)=sqrt(2/S)[cos(z_s^T x + c_s)] is a fixed, untrained cosine layer; learning happens only in the linear projections W_e and W_d and the latent classifier weight theta_c. The density estimator p(lambda) proportional to exp(w^T phi(lambda)) N(lambda|mu,sigma) approximates the latent data density with an exponentiated RFF function under a Gaussian envelope, and its gradient steers the counterfactual search toward high-density target regions. The decision-boundary constraint theta_c^T f_e(x+delta)=0, plus the projection methodology for choosing beta, is w","core_discovery":"GPAE replaces the usual neural encoder and decoder with Gaussian processes whose RBF kernel is approximated by random Fourier features. The encoder becomes a fixed random cosine layer followed by one trainable linear projection W_e, and the decoder the same with W_d; only the two projections and the latent classifier's boundary theta_c are learned. The paper's claim is that this parameter-light supervised autoencoder, searched with a density regularizer, generates counterfactuals that are valid, diverse, and in-distribution. The counterfactual for a query x is x+delta, where delta is found by minimizing ||delta||^2/2 - beta log p(f_e(x+delta)) subject to theta_c^T f_e(x+delta)=0; a mask mult","pith_inferences":["The beta-selection scheme, which projects latent codes onto the decision boundary and minimizes KL divergence between the projected data and the density estimate, could be reused by any latent-space counterfactual method that has a linear classifier, not just GPAE.","Because the density estimator is learned once in latent space, it could also serve as an out-of-distribution detector, flagging query points whose latent codes fall in low-density regions before an explanation is attempted.","The same architecture could incorporate non-RBF kernels through alternative random feature maps, making the density estimator and counterfactual search applicable to heterogeneous feature scales without architectural changes."],"forward_implications":["Counterfactuals are generated directly in feature space, so immutable or non-actionable features can be kept fixed by masking gradients, without post-hoc replacement that creates out-of-distribution samples.","Because only two linear projections and a boundary vector are trained, the explainer is much lighter than deep autoencoder baselines and less prone to overfitting on small tabular sets.","The density regularizer and beta-selection procedure make the counterfactual distribution track the target class's latent distribution, which is what yields the reported low IM1/IM2 and high validity.","The same encoder doubles as a competitive classifier, so the model being explained is not an arbitrary black box; the latent space it supervises is the one being searched.","The masking mechanism extends the same search to fair-recourse settings where protected attributes must not be altered."],"supporting_citations":[{"why":"Supplies the random Fourier feature approximation used to make the GP encoder, decoder, and density estimator scalable.","marker":"[43]"},{"why":"Establishes the link between infinitely wide neural networks and Gaussian processes that motivates building the autoencoder from GPs.","marker":"[30]"},{"why":"The supervised-autoencoder architecture this paper extends; contributes the probit-supervised latent space and boundary-projection counterfactual idea.","marker":"[59]"},{"why":"Defines the counterfactual optimization objective and provides the gradient-search baseline.","marker":"[56]"},{"why":"A VAE-based counterfactual baseline evaluated in the experiments, supplying the comparison for latent nearest-neighbor counterfactual search.","marker":"[36]"},{"why":"A VAE-based counterfactual baseline evaluated in the experiments, contributing the comparison on decoder-generated counterfactuals.","marker":"[19]"},{"why":"The end-to-end prediction-aware counterfactual baseline, including the immutable-feature replacement behavior the paper contrasts with its mask.","marker":"[17]"},{"why":"Defines the IM1/IM2 interpretability metrics used to score plausibility of the generated counterfactuals.","marker":"[55]"}],"fun_headline_variants":["GP autoencoder: fewer parameters, in-distribution counterfactuals","Lightweight GP autoencoder for realistic counterfactuals","Density-guided GP autoencoder diversifies counterfactuals","Lean GP autoencoder yields in-distribution counterfactuals"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The search assumes that a latent code exactly on the decision boundary flips the classifier's predicted label to the target class; with a sigmoid output, that boundary is only the 50% probability contour, so label flipping is not logically entailed by the constraint.","fun_headline_variants_meta":{"raw":{"variants":["GP autoencoder: fewer parameters, in-distribution counterfactuals","Lightweight GP autoencoder for realistic counterfactuals","Density-guided GP autoencoder diversifies counterfactuals","Lean GP autoencoder yields in-distribution counterfactuals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001788,"raw_usage":{"total_tokens":6854,"prompt_tokens":683,"completion_tokens":6171,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":6103}},"tokens_in":427,"tokens_out":6171,"duration_ms":42896,"temperature":1.0,"reasoning_tokens":6103,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:06:36.976462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a tabular dataset, train GPAE and, for a held-out query, record counterfactuals at the converged solution of the constrained search, then evaluate the explained classifier's argmax label at those counterfactuals. If a substantial fraction are not assigned the target class, then the boundary constraint theta_c^T f_e(x+delta)=0 does not by itself make the sample valid, and the reported ~0.99 validity must be attributed to the density-regularized optimization rather than the equality constraint.","supporting_citations":[{"cited_title":"Random features for large-scale kernel machines","cited_arxiv_id":null,"evidence_quote":"Supplies the random Fourier feature approximation used to make the GP encoder, decoder, and density estimator scalable."},{"cited_title":"An interpretable deep clas- sifier for counterfactual generation","cited_arxiv_id":null,"evidence_quote":"The supervised-autoencoder architecture this paper extends; contributes the probit-supervised latent space and boundary-projection counterfactual idea."},{"cited_title":"Counterfactual explanations without opening the black box: Automated decisions and the gdpr","cited_arxiv_id":null,"evidence_quote":"Defines the counterfactual optimization objective and provides the gradient-search baseline."},{"cited_title":"Learning model-agnostic counterfactual explanations for tabular data","cited_arxiv_id":null,"evidence_quote":"A VAE-based counterfactual baseline evaluated in the experiments, supplying the comparison for latent nearest-neighbor counterfactual search."},{"cited_title":"Counternet: End- to-end training of prediction aware counterfactual explanations","cited_arxiv_id":null,"evidence_quote":"The end-to-end prediction-aware counterfactual baseline, including the immutable-feature replacement behavior the paper contrasts with its mask."},{"cited_title":"Interpretable counterfactual explanations guided by prototypes","cited_arxiv_id":null,"evidence_quote":"Defines the IM1/IM2 interpretability metrics used to score plausibility of the generated counterfactuals."}],"review_version":1}