{"id":"3c3432fa-1a20-41c9-96cd-8f4059868be3","arxiv_id":"2509.00876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"TDCE guides a tabular diffusion reverse process with Gumbel-softmax classifier gradients to produce counterfactual explanations, achieving high validity on four benchmarks.","lead":"A new diffusion-based method, TDCE, generates counterfactual explanations for tabular data by relaxing categorical features into a Gumbel-softmax space so classifier gradients can guide the reverse process. On four finance and social-science datasets it reaches high validity and diversity, though its theoretical justification appears flawed and no code is provided.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.1's proof uses a false inequality (Eq. A.17); the KL bound justifying Eq. (16) is unproven, so the categorical guidance update (Eq. 18) lacks a theoretical guarantee.","rationale":"The reader's weakest_assumption identifies exactly the false inequality in Appendix A (Eq. A.17), and my analysis confirms that this inequality is indeed invalid. This is the most load-bearing concern because Eq. (16) is the mathematical foundation for the categorical guidance update in Eq. (18) and Algorithm 1; if the approximation is not close to the Gumbel-softmax posterior, the guided reverse process may generate counterfactuals that do not respect the data manifold, even if the reported metrics look good. The proof error is not a minor typo—it breaks the only theoretical argument for why the approximation should work. A secondary issue is that Algorithm 1's line 7 ('ex_cat ← µ_cat + ||µ_cat|| g_cat') appears to add the gradient to the mean vector on the simplex rather than to the logits as Eq. (18) would suggest; this inconsistency merits clarification, but without code it is hard to determine whether it is a real flaw or a notational shorthand. The paper does provide substantial empirical evaluation across four datasets and multiple baselines, which is independent evidence that the method has some practical value. However, the theoretical guarantee is unproven, and the lack of code and missing diffusion-based baselines further reduce confidence. The appropriate verdict remains CONDITIONAL: the method is worth pursuing, but the manuscript needs a corrected proof or a clear acknowledgment that Eq. (16) is purely heuristic, plus a reproduction package and stronger baselines.","tokens_in":18353,"tokens_out":11084,"duration_ms":125390,"concrete_test":"Numerically compute KL(p_GS(·|π,τ) ∥ p_θ(·|π)) on a grid of (π,τ), using the exact densities in Eqs. (13) and (19), with Z obtained by high-resolution numerical integration over the simplex. Sample π from Dirichlet(1) for K=2 and K=3, and τ ∈ {0.1, 0.3, 1, 5}. Check whether the claimed upper and lower bounds of Theorem 4.1 hold for all tested points. Begin with the counterexample configuration π=(0.95, 0.05), τ=1, x_min=0.2. If any point violates the claimed bounds, Theorem 4.1 is false and Eq. (16)'s theoretical justification is void; if all satisfy the bounds, the theorem may still hold and only the proof step needs repair.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The categorical guidance update (Eq. 18) is derived from the approximation log pθ(x_t|x_{t+1}) ≈ x_t^T log πθ(x_{t+1}) + const (Eq. 16). The paper's only theoretical support for this approximation is Theorem 4.1's KL bound between the Gumbel-softmax density (Eq. 13) and the simplified product density (Eq. 19). The proof of the upper bound in Appendix A relies on the inequality π_i^{1-x_i} ≤ (1-x_min)^{1-x_i} (Eq. A.17). This is false: take π=(0.95, 0.05), x=(0.2, 0.8), x_min=0.2. Then π_1^{1-x_1} = 0.95^0.8 ≈ 0.96, while (1-x_min)^{1-x_1} = 0.8^0.8 ≈ 0.84. The inequality can fail substantially, so the derivation of the upper bound collapses. Without a valid bound there is no demonstrated sense in which Eq. (16) approximates the Gumbel-softmax posterior; the guided process may drift from the true classifier-conditioned reverse process, undermining the claim that generated counterfactuals are realistic. The empirical metrics are promising, but they are not backed by a rigorous approximation guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TDCE, a tabular counterfactual-explanation method that extends classifier guidance to categorical features by relaxing one-hot vectors with a Gumbel-softmax reparameterization. The reverse process is guided by classifier gradients; a product-form approximation to the Gumbel-softmax density is introduced, and Theorem 4.1 claims a KL-divergence bound between the two. Experiments on four tabular datasets compare TDCE with Wachter, CCHVAE, REVISE, CLUE, FACE, and CounterNet on validity, interpretability, diversity, stability, JS divergence, L2 distance, and runtime. The paper reports that TDCE achieves competitive or superior performance on most metrics, especially validity and interpretability, while also being faster than search-based baselines.","tokens_in":18737,"tokens_out":6434,"duration_ms":81249,"significance":"If the method and its supporting analysis are correct, the paper makes a useful contribution: it provides a simple, model-agnostic way to perform classifier guidance on categorical tabular features within a diffusion framework, a setting that has received less attention than image counterfactuals. The empirical tables are presented with standard deviations and the runtime comparison is informative. However, the theoretical bound is not currently reliable, the algorithm as written does not transparently implement the derived update, and the empirical comparison lacks diffusion-based baselines and uses per-dataset tuning of the temperature. These issues affect the strength of the main claims, but the underlying idea is plausible and the experimental results are promising. The paper does not provide code or a reproducibility artifact, so independent verification is not yet possible.","major_comments":[{"comment":"The proof of Theorem 4.1 relies on the inequality π_i^{1-x_i} ≤ (1-x_min)^{1-x_i}, but this is false in general. For example, with π=(0.95,0.05), x=(0.2,0.8), x_min=0.2, the left-hand side is 0.95^0.8 ≈ 0.96 while the right-hand side is 0.8^0.8 ≈ 0.84. Consequently, the upper-bound derivation leading to Eq. (A.22) collapses. Since Theorem 4.1 is the only theoretical justification for the approximation in Eq. (16), and Eq. (16) is used in deriving the guided update in Eq. (18), the theoretical guarantee for the categorical reverse process is not established. The authors should either repair the proof, provide a valid bound, or explicitly reframe Eq. (18) as a heuristic approximation supported only by experiments.","section":"Appendix A, Eq. (A.17)"},{"comment":"The update `ex_cat ← µ_cat + ||µ_cat|| g_cat` does not match the derived guided density in Eq. (18), which suggests updating the logits as `log π_θ + λ g_cat`. The algorithm also omits λ. It is unclear what µ_cat represents (predicted x_0? predicted logits? the mean of a Gumbel-softmax transition?) and how it is computed from the denoiser output. If the implemented update differs from Eq. (18), the derivation and the experiments may not correspond to the same method. Please specify the exact relationship and provide the sampling step or the closed-form update that follows from Eq. (18).","section":"Algorithm 1, line 7"},{"comment":"The empirical comparison is weakened by two issues. First, the temperature τ is searched per dataset over [0.1, 5] using the benchmark metrics themselves, while the baselines are reported with CARLA defaults. This can inflate the relative performance of TDCE. Second, there are no diffusion-based counterfactual baselines, despite the existence of recent discrete-diffusion guidance methods that the paper cites (e.g., [32,33]). A fairer comparison with baselines tuned on validation splits, plus at least one diffusion-based baseline, is needed to support the claim that TDCE 'outperforms popular baseline methods.'","section":"Section 5.3 and Tables 2–3"}],"minor_comments":[{"comment":"The upper bound in the main text uses `log[1 - x_min]`, while Appendix A, Theorem A.2 states `log x_min` and the derivation in A.22 uses `log(1 - x_min)`. These versions should be reconciled.","section":"Theorem 4.1 vs. Appendix A.2"},{"comment":"The figure lacks axis labels and numerical values. Since the paper claims a 'tight bound,' it would help to plot the actual KL and both bounds on the same axes, or at least report the numerical values used.","section":"Figure 3"},{"comment":"The temperature schedule (warm start, then decrease) is described in the text but not reflected in Algorithm 1. Please state the schedule explicitly or add it to the pseudocode.","section":"Section 4.4 and Algorithm 1"},{"comment":"Typo: 'low IM2/IM2' should read 'low IM1/IM2.' Also, 'Efficency' in the subsection heading should be 'Efficiency.'","section":"Section 5.3"},{"comment":"The function `f_dn` is used in Eq. (8) before being defined. Define it at first use.","section":"Section 3.2"},{"comment":"Caption says 'on the same computer setting'; should be 'same hardware and software settings' or similar.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The central idea is of interest and the empirical results are encouraging, but the theoretical bound is not proven and the algorithm as written is ambiguous. I recommend major revision rather than rejection because the empirical contribution may stand even if the bound is removed or replaced by a heuristic justification, provided the experiments are made more rigorous and the algorithm is clarified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mike,\n\nQuick take on arXiv:2509.00876. The paper proposes TDCE, a diffusion-based counterfactual generator for tabular data that handles continuous and categorical features together. The key trick is to relax categorical one-hot vectors with Gumbel-softmax during the reverse diffusion, so you can backprop classifier gradients through them and do guidance similar to the Gaussian case. That is a sensible extension of Schiff et al. and Gruver et al. to the tabular mixed-type setting, and the empirical results on four datasets look decent: validity near 1, low JS divergence, lower instability than most baselines, and competitive diversity/interpretability. The authors also study temperature effects and handle immutable features with a masking trick. That is genuinely useful for applied explainability in credit and similar domains.\n\nThe problem is Theorem 4.1. The bound between the Gumbel-softmax density and their simplified product form is load-bearing for Equation 16, which justifies the categorical guidance update. The proof in Appendix A uses the inequality pi_i^{1-x_i} ≤ (1-x_min)^{1-x_i} (Eq. A.17). That inequality is false; a concrete counterexample is pi=(0.95,0.05), x=(0.2,0.8), x_min=0.2, where 0.95^0.8≈0.96 > 0.8^0.8≈0.84. So the upper bound derivation collapses, and with it the claim that Equation 16 is a controlled approximation. The lower bound also has some hand-wavy steps involving the Lebesgue measure of the simplex. In essence, the theoretical contribution as stated is not proven.\n\nThat doesn't sink the method entirely. The empirical story is plausibly independent of the bound—the approximation could still be good in practice, and the results suggest it is. But the paper currently overclaims. I'd recommend requesting a correction: either fix the proof, replace it with a valid bound, or drop the theorem and present Equation 16 as a heuristic with a small empirical validation of the approximation quality. Also missing are the most relevant baselines—diffusion-based counterfactual generators like TabDDPM or DiDDM-style methods—and no code is released, which makes reproducibility harder.\n\nWho is this for? Practitioners building recourse systems on tabular data and researchers working on guided discrete diffusion. The idea has legs, but the current manuscript needs revision. A serious referee should be engaged; a desk reject would be a mistake. I'd send it to review and ask the authors to fix the proof or reframe the contribution.\n\nWorth a read, but with a red flag on the theorem.","headline":"Useful empirical recipe for tabular counterfactuals via Gumbel-softmax guidance, but the theoretical bound in Theorem 4.1 is unproven—the proof rests on a false inequality.","tokens_in":19168,"tokens_out":2604,"would_cite":false,"duration_ms":28340,"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":"Tabular classifiers can be explained by guiding a diffusion reverse process, with categorical features relaxed through Gumbel-softmax so classifier gradients can flow.","keywords":["counterfactual explanations","diffusion models","tabular data","Gumbel-softmax","classifier guidance","categorical features","explainable machine learning","recourse"],"falsifier":"Compute the guided and unguided reverse trajectories on the same datasets and compare validity and JS divergence; if they match, the categorical guidance is not the active mechanism. Also, the proof's key inequality (Appendix A, A.17) fails for simple draws like π=(0.95,0.05), x=(0.2,0.8), so the bound cannot be relied on as a guarantee.","tokens_in":18266,"feed_emoji":"📊","tokens_out":8397,"duration_ms":87440,"temperature":0.7,"pith_summary":"Counterfactual explanations for tabular data have mostly been handled by latent-space searches or per-sample optimization; this paper proposes a diffusion-based alternative that works directly in the feature space. By relaxing categorical one-hot vectors through Gumbel-softmax, the reverse process can accept classifier gradients and guide the sample toward the target class without enumerating category combinations. The authors test on four credit and law datasets and report that their TDCE procedure achieves the highest validity, low JS divergence between counterfactual and target categorical distributions, and competitive interpretability, in a fixed number of reverse steps. A working version would give finance and social-science practitioners a model-agnostic explainer that does not require per-sample search.","feed_headline":"Diffusion guidance now flips tabular predictions with categorical edits","feed_subtitle":"Categorical features stay realistic while the reverse process moves the classifier to the target class.","key_machinery":"The load-bearing object is the Gumbel-softmax reparameterization of each categorical feature combined with a linearized guided reverse update. The relaxed vector replaces the one-hot value; the reverse log-density is approximated by the inner product of the relaxed vector with log probabilities from the denoiser, and the classifier's gradient is added to those logits. This makes the categorical reverse step resemble the Gaussian classifier-guidance update, allowing backpropagation through discrete features without enumerating all category combinations. The temperature controls how close the relaxation is to a one-hot vector and how much gradient variance appears.","core_discovery":"Central claim: a tabular diffusion model becomes a counterfactual explainer when classifier guidance is added to the reverse process and categorical features are relaxed with Gumbel-softmax vectors. Continuous features use the usual mean-plus-covariance gradient shift; categorical features approximate the Gumbel-softmax log-density linearly and add the classifier gradient to the logits (Eq. 18). The final category is an argmax. On four credit and law datasets, the paper reports the best validity, lower JS divergence for categorical distributions, and better interpretability than baselines at a fixed number of reverse steps, plus a temperature-dependent bound linking low temperature to a bett","pith_inferences":["Because Eq. (16) linearizes the Gumbel-softmax log-density, the guidance update effectively shifts logits by a constant gradient; a natural extension is to learn a per-step logit correction that keeps the reverse trajectory on the data manifold even at high temperature.","The approximation error is controlled by the minimum relaxed value, so categorical variables with skewed category distributions are the likely failure mode; testing on datasets with many rare categories would show where the method's validity advantage erodes.","If the reported validity and interpretability gains hold, they imply that feature-space diffusion explainers stay closer to the data manifold than latent-space VAE explainers; this is directly testable by measuring the distance from generated counterfactuals to a held-out training set."],"forward_implications":["Counterfactual generation for categorical features becomes a fixed-cost reverse pass rather than a per-sample search, so explanation time scales with the diffusion depth, not with the data size.","Continuous and categorical features are handled jointly in the same reverse process, so explanations can respect numeric and discrete structure at the same time.","Immutable features can be enforced with a mask that blends the noisy input into the guided sample, which keeps protected or unchangeable attributes fixed.","Temperature becomes a practical dial: lower temperatures make the relaxed features closer to one-hot but increase gradient variance, so each dataset needs its own temperature setting."],"supporting_citations":[{"why":"Supplies the Gumbel-softmax reparameterization that turns one-hot categorical vectors into differentiable continuous vectors.","marker":"[34]"},{"why":"Provides the denoising diffusion probabilistic model framework and the noise-prediction objective the tabular model builds on.","marker":"[23]"},{"why":"Introduces classifier guidance for continuous features, which the paper adapts to the categorical case.","marker":"[25]"},{"why":"Supplies the adaptive parameterization with distance constraint used for the continuous-feature guidance.","marker":"[26]"},{"why":"Defines multinomial diffusion for discrete data, the base categorical diffusion model being relaxed.","marker":"[27]"},{"why":"Establishes the tabular diffusion modeling setting the paper extends to counterfactual generation.","marker":"[30]"},{"why":"Formulates the counterfactual explanation objective and serves as the direct feature-space baseline.","marker":"[12]"},{"why":"Provides a conditional-VAE tabular counterfactual baseline that the experiments compare against.","marker":"[22]"}],"fun_headline_variants":["Diffusion counterfactuals for tables get realistic categorical edits","Gumbel-softmax guide makes tabular counterfactuals robust","Tabular diffusion flips predictions with honest categorical changes","Guided reverse process beats baselines on credit data"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The categorical guidance pipeline assumes the approximation log pθ(ext|ext+1) ≈ ex⊤ log πθ(ext+1) + const is accurate enough that the guided reverse process still produces realistic categorical distributions; the paper's stated bound for this closeness relies on an inequality that does not hold for all Gumbel-softmax draws, so the theoretical guarantee is not actually established.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion counterfactuals for tables get realistic categorical edits","Gumbel-softmax guide makes tabular counterfactuals robust","Tabular diffusion flips predictions with honest categorical changes","Guided reverse process beats baselines on credit data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000146,"raw_usage":{"total_tokens":976,"prompt_tokens":660,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":404,"completion_tokens_details":{"reasoning_tokens":257}},"tokens_in":404,"tokens_out":316,"duration_ms":3944,"temperature":1.0,"reasoning_tokens":257,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:07:29.378778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the guided and unguided reverse trajectories on the same datasets and compare validity and JS divergence; if they match, the categorical guidance is not the active mechanism. Also, the proof's key inequality (Appendix A, A.17) fails for simple draws like π=(0.95,0.05), x=(0.2,0.8), so the bound cannot be relied on as a guarantee.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gumbel-softmax reparameterization that turns one-hot categorical vectors into differentiable continuous vectors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the denoising diffusion probabilistic model framework and the noise-prediction objective the tabular model builds on."},{"cited_title":"Dhariwal, A","cited_arxiv_id":null,"evidence_quote":"Introduces classifier guidance for continuous features, which the paper adapts to the categorical case."},{"cited_title":"Augustin, V","cited_arxiv_id":null,"evidence_quote":"Supplies the adaptive parameterization with distance constraint used for the continuous-feature guidance."},{"cited_title":"Hoogeboom, D","cited_arxiv_id":null,"evidence_quote":"Defines multinomial diffusion for discrete data, the base categorical diffusion model being relaxed."},{"cited_title":"Kotelnikov, D","cited_arxiv_id":null,"evidence_quote":"Establishes the tabular diffusion modeling setting the paper extends to counterfactual generation."},{"cited_title":"Wachter, B","cited_arxiv_id":null,"evidence_quote":"Formulates the counterfactual explanation objective and serves as the direct feature-space baseline."},{"cited_title":"Pawelczyk, K","cited_arxiv_id":null,"evidence_quote":"Provides a conditional-VAE tabular counterfactual baseline that the experiments compare against."}],"review_version":1}