{"id":"642e2a49-a820-499f-aeb8-2a29f9b57114","arxiv_id":"2607.23348","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Smoothed logit (natural-parameter) coordinates for categorical variables improve or match one-hot encoding in Flow Matching and diffusion on mixed tabular data, with stability bounds and imbalance-aware nonparametric rates.","lead":"Tabular rows that mix numbers and categories are hard for continuous generative models, which expect Euclidean coordinates. This paper embeds categories as smoothed logit coordinates instead of one-hot flags, and shows that flow matching and diffusion then match or improve mixed table generation, especially for rare category combinations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theory covers only componentwise generator, while headline Logit FM gains come from a flat joint model explicitly excluded by the paper's own §5.4 and §6.","rationale":"The reader's weakest assumption identifies exactly the gap between the componentwise theory and the flat joint model used for the headline Logit FM results. I agree that this is the most load-bearing concern: the paper's central message is a representation-level improvement, and the theory is offered as an explanation (stability + imbalance-aware rates). If that theory does not cover the model that produces the best flat-model results, then the scientific claim is only empirical. The paper explicitly acknowledges the limitation in §5.4 and §6, so it is not a hidden flaw, but the abstract still presents the theory and experiments as a unified contribution. The proposed concrete test—running the componentwise generator that the theory does cover—would either validate the theoretical link or demonstrate that the flat-model gains are not explained by the paper's mechanism. Given the transparency of the limitation and the honesty of the empirical reporting, I do not think the verdict should change from CONDITIONAL: the paper should be accepted after the theory/experiment scope is clarified or the flat model is covered. This matches the reader's recommendation, so no adjustment is needed.","tokens_in":29572,"tokens_out":8973,"duration_ms":96266,"concrete_test":"Re-run the Adult and Buddy benchmarks with the exact componentwise generator of Corollary 5.7 (per-class continuous flows plus a logit marginal flow, common Gaussian base) under both one-hot and scaled-logit coordinates, using the same architectures and ten splits. If the logit advantage over one-hot persists in this theory-covered model, the theoretical mechanism is validated; if it diminishes or reverses, the flat-model gains cannot be attributed to the analyzed framework and the abstract should be relabeled as componentwise-only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim pairs a representation-level theory with empirical improvements: 'scaled-logit coordinates improve or match one-hot coordinates, especially under severe rare-cell imbalance' and 'Logit FM improves the primary distributional metrics on three datasets' (Abstract, §9). The theoretical counterpart, Theorem 6.6, is an imbalance-aware rate for an idealized componentwise generator: a logit flow for categorical mass plus separate per-class conditional continuous flows with common base initialization (Corollary 5.7). The flat joint model that produces the headline Logit FM results is not covered. Section 5.4 states this explicitly: 'For the flat joint model, Theorem 5.2 directly controls Wasserstein error for the full transformed row... but it does not by itself give the class-conditional transport bounds assumed below.' Section 6 repeats the caveat: 'We state one explicit sieve result for the componentwise model used in Corollary 5.7. This avoids treating the calculation as an implementation-level guarantee for the flat joint neural generator.' Consequently, the theory does not substantiate the main empirical comparison; the logit-representation conclusion rests on the experiments alone. This is not an internal inconsistency—the paper is transparent about the gap—but it is load-bearing because the abstract and contributions present the stability/rate analysis and the empirical results as mutually supporting. The rate theorem's effective-sample-size terms also presuppose per-class regression with nπ_k samples, whereas the flat model trains on all rows jointly, so the rare-cell mechanism in Theorem 6.6 is not the mechanism realized in the flat model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a logit-coordinate (additive log-ratio) representation for categorical variables in mixed continuous–categorical tabular data, combined with Flow Matching or Gaussian diffusion. It introduces a mixed-distribution discrepancy Dmix that separates categorical marginal error from class-conditional continuous Wasserstein error, proves softmax Lipschitz and stability bounds, and derives nonparametric convergence rates for an idealized componentwise generator (a logit flow plus separate per-class continuous flows). Simulations compare logit vs one-hot embeddings under rare-cell imbalance, and real-data experiments on Adult, Churn2, Cardio, and Buddy compare flat and block-conditional logit FM, logit/one-hot diffusion, and TabDDPM baselines across ten splits. The paper claims that scaled-logit coordinates improve or match one-hot coordinates, especially under severe imbalance, and that block-conditional structure further improves fidelity.","tokens_in":29928,"tokens_out":6038,"duration_ms":60276,"significance":"If the results hold, the paper offers a useful representation-level insight: the coordinate system used to embed categorical variables materially affects the fidelity of continuous generative models on mixed tabular data. The Dmix criterion is a sensible decomposition, and the stability and rate arguments are standard and largely correct. The paper is transparent about several limitations—most importantly, that the main theory covers only a componentwise generator and not the flat joint Logit FM used for several headline experiments, and that the experiments use maximum-probability decoding while the theory controls randomized softmax decoding. The ten-split real-data benchmark is a practical strength, and the within-family comparisons (logit vs one-hot under fixed architectures) are carefully framed. However, the gap between theory and the headline experiments, together with the absence of a comparative theoretical guarantee, means the representation-level claim rests primarily on the empirical results.","major_comments":[{"comment":"The rate analysis and mixed-fidelity bounds (Corollary 5.7, Theorem 6.6) are proved for a componentwise generator—a separate logit flow plus per-class conditional continuous flows—whereas the headline Logit FM results come from flat joint models. This is stated explicitly in §5.4 ('For the flat joint model, Theorem 5.2 ... does not by itself give the class-conditional transport bounds assumed below') and in §6 ('This avoids treating the calculation as an implementation-level guarantee for the flat joint neural generator'). Since the real-data simulations and benchmarks use flat joint generators, the theory does not substantiate the main empirical comparisons; the logit-representation conclusion rests on the experiments alone. The paper should either reframe the theory as an idealized analysis that motivates, but does not explain, the flat-model results, or add experiments with the compon","section":"§5.4, §6, §7.2"},{"comment":"The theoretical guarantees apply to randomized softmax decoding: Corollary 5.6 and Theorem 6.6 control Dmix for bX(d) ~ Categorical(p̂), while the experiments use maximum-probability decoding for both one-hot and scaled-logit variants (§4.6, §8.1). Remark 6.7 admits that maximum-probability decoding 'requires margin control' and is 'outside the scope of the theorem.' Thus the primary theoretical quantity Dmix is not the quantity evaluated in the experiments. The authors should either implement randomized softmax decoding in the experiments, or provide margin diagnostics showing that max-probability decoding is consistent with the theoretical decoder, or explicitly separate the theory from the experimental protocol.","section":"§4.6, Theorem 6.6, Remark 6.7"},{"comment":"The paper does not provide a formal argument that logit coordinates improve over one-hot coordinates. Proposition 3.1 and Proposition 3.4 compare KL control in natural-parameter space vs probability space, but the Flow Matching training objective is Euclidean in the embedding space and the decoding is argmax; no analogue of Theorem 6.6 is derived for one-hot embeddings, so no comparative rate is available. The abstract's claim 'scaled-logit coordinates improve or match one-hot coordinates, especially under severe rare-cell imbalance' is therefore an empirical assertion rather than a consequence of the theory. The authors should either derive a comparative bound (e.g., using margin or Lipschitz properties of one-hot decoding) or temper the theoretical framing so it does not imply a proof of the representation-level advantage.","section":"§3.4, Theorem 6.6"},{"comment":"In the controlled simulations, the reported logit-vs-one-hot differences at severe imbalance are comparable to the standard deviations across the ten seeds. For example, at ρ=0.01, Logit-FM vs OneHot-FM gives Dmix 0.157±0.007 vs 0.165±0.008 and Dbal 0.195±0.007 vs 0.203±0.008; the diffusion comparisons show similar effect sizes. No significance tests or paired comparisons are reported. Given the abstract's emphasis on 'especially under severe rare-cell imbalance,' the authors should quantify the uncertainty (e.g., paired tests or confidence intervals) or restate the simulation claim as a directional improvement, rather than a definitive demonstration.","section":"Table 1, Table 3, §7.4"}],"minor_comments":[{"comment":"The smoothing parameter ε and the logit scaling aK are central to the method, but the values used in the simulations and real-data experiments are not reported. This is needed for reproducibility and for interpreting the stability constants.","section":"§4.2, §7, §8"},{"comment":"The definition of Dmix assumes population distributions. For empirical evaluations, generated samples can have empty cells; the paper mentions a support convention but does not state which convention is used in the reported tables. Please specify.","section":"§5.3, Eq. (1)"},{"comment":"The ALR reference category is chosen as a high-frequency category in the training data, and sensitivity to this choice is deferred. This should at least be documented per dataset, since the reference category can affect finite-sample training.","section":"§3.3, §7.2"},{"comment":"Minor typographical issues: 'H¨ older' should be 'Hölder', 'ETrain' appears in §A.5, and some inline math is inconsistently formatted. These do not affect the substance.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is largely sound but overclaims theoretical support for its headline flat-joint Logit FM experiments. The central theory is for a componentwise generator and a randomized decoder, while the experiments use flat joint models and maximum-probability decoding. This is not an internal inconsistency—the authors are transparent about it—but it is load-bearing for the abstract's mutually-supporting framing. The authors could resolve it by repositioning the theory as an idealized analysis, adding componentwise/block-conditional experiments, or deriving conditional-stability bounds for the flat joint model. The simulation significance issue should also be addressed before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the paper is more honest than the abstract suggests, and the stress-test is right about the disconnect. The rates are explicitly for a componentwise generator, while the flat joint Logit FM that produces many of the real-data tables is excluded by the paper's own §5.4 and §6. To the authors' credit, the limitation is stated plainly. But it is load-bearing: the theory does not explain the flat-model gains, and the representation-level claim rests on the experiments alone. That is a gap, not a concealed flaw.\n\nWhat is genuinely new is the package: a mixed-discrepancy Dmix separating categorical TV from conditional Wasserstein error, a straightforward stability bound (Gronwall plus softmax Lipschitzness), and a nonparametric rate that makes the nπ_k effective-sample-size loss explicit. Individually these are standard tools, but the combination is useful and clearly presented. The simulations and the ten-split real-data comparisons are a serious empirical effort, and the within-family design is the right way to isolate the representation effect. The block-conditional extension adds a meaningful structural axis. The geometric propositions are simple and correct, and the paper does not overclaim about Fisher-Rao geometry.\n\nSoft spots, in proportion: first, the real-data Dmix conditions on the label only, not the joint categorical-cell structure that the theory and simulations use. So the headline empirical metric is not literally the same object as the theoretical Dmix. Second, no code, data, or values for the smoothing parameter ε, the ALR reference category, or the scaling aK are reported, which hurts reproducibility and prevents sensitivity checks. Third, several reported improvements are modest—a few hundredths in Dmix with overlapping standard deviations—and no significance tests are given. The consistency across seeds and datasets is suggestive, but not overwhelming. Fourth, the claimed advantage over one-hot is not derived from theory; there is no analogous rate for the one-hot representation, so the comparison is empirical. That is acceptable, but the abstract should say it more loudly.\n\nThe citation pattern is sound: the paper positions itself against the existing logit/simplex flow work without claiming priority for the coordinate transform.\n\nWho should read this: anyone building mixed tabular generative models and anyone who wants a clear statement of why rare cells are hard. It deserves a serious referee. In revision I would ask for code and data, the scope clarification in the abstract, a definition that matches the empirical Dmix to the theoretical one, and a comparison with the closest prior logit-based flow methods. With those changes I would accept it.\n\nSend it to referees; do not desk reject.","headline":"Honest and solid within-family empirical study of logit coordinates for mixed tabular generation, but the rate theory covers only the componentwise generator while the headline gains come from the flat joint model, so the experiments carry the main claim alone.","tokens_in":30398,"tokens_out":3720,"would_cite":false,"duration_ms":38251,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62B10","62G05","62G20","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Encoding categorical variables as smoothed logit coordinates, rather than one-hot vectors, improves fidelity of continuous generative models on mixed tabular data, especially under rare-category imbalance.","keywords":["mixed tabular data","logit coordinates","flow matching","diffusion models","rare-category imbalance","information geometry","nonparametric rates","Wasserstein discrepancy"],"falsifier":"Run the severe-imbalance simulation (rho = 0.01) with the ALR reference category cycled through all 24 joint cells and with the smoothing parameter eps varied over a grid; if there exists a reference choice or a eps value where One-Hot FM matches or beats Logit-FM on Dmix, the claimed representation-level advantage is not invariant to coordinate conventions.","tokens_in":29471,"feed_emoji":"📊","tokens_out":4632,"duration_ms":46032,"temperature":0.7,"pith_summary":"The paper argues that how categorical variables are embedded into Euclidean space is a consequential modeling choice for continuous generative models trained on mixed tabular data. It proposes embedding categories as smoothed natural parameters (logit coordinates) instead of one-hot vectors, then running standard Flow Matching or Gaussian diffusion in the resulting space. If the paper is right, a simple coordinate swap—no change to architecture or training objective—improves fidelity on imbalanced mixed data, and the theoretical analysis explains why: logit coordinates give local quadratic control of KL divergence, while one-hot Euclidean error does not. The paper also introduces a mixed-distribution discrepancy that separates categorical mass error from conditional continuous transport error, and derives nonparametric rates that make rare-cell sample-size loss explicit.","feed_headline":"Logit coordinates beat one-hot for mixed tabular data","feed_subtitle":"Rare categories get explicit contrast in logit space; block conditioning adds fidelity gains across four datasets.","key_machinery":"The additive log-ratio (ALR) coordinate map eta_k = log(q_k / q_r) with softmax decoding, combined with the mixed discrepancy Dmix = TV(pi*, pi_hat) + sum_k pi*_k W2(P*_k, P_hat_k). The stability argument uses the Wasserstein transport bound W2(P*_1, P_hat_1) <= C * integral sqrt(E||v_hat - v*||^2) dt, plus the softmax Lipschitz bound ||S(eta) - S(eta')||_1 <= (sqrt(K)/2) ||eta - eta'||_2, which converts logit-space error into categorical probability error.","core_discovery":"Embedding categorical variables via smoothed additive log-ratio coordinates eta_k = log(q_k / q_r) turns the categorical part of mixed data into an unconstrained Euclidean vector while preserving exponential-family structure. For this representation, the paper proves a stability bound of the form Dmix <= eps + C1 n^{-alpha_Z/(2alpha_Z+d_Z+1)} + sum_k pi*_k C2,k (n pi*_k)^{-alpha_c/(2alpha_c+p+1)}, making explicit that rare cells learn from effective sample size n pi*_k. Empirically, within-family comparisons across ten splits show Logit Flow Matching improves or matches One-Hot FM on the primary distributional metrics on three of four datasets; Block-Conditional Logit FM consistently improve","pith_inferences":["If the representation effect is general, the same coordinate swap should help other Euclidean-space generative families on mixed tabular data—for example, score-based models with different noise schedules or variational autoencoders—though the paper does not test this.","Because ALR coordinates depend on a reference category, the reported logit gains could be partly reference-dependent; a permutation-invariant log-ratio formulation would test whether the improvement is intrinsic to natural-parameter structure or to the particular reference choice.","The rare-cell rate suggests a practical resource bound: with total sample n, cells with mass below roughly 1/n cannot support consistent conditional continuous estimation on their own, so generators should pool across rare cells or share structure—an extension the paper's block-conditional results hint at but do not develop.","A direct testable extension is to vary the smoothing parameter eps and the scaling constant a_K; if the logit advantage over one-hot disappears for large eps, the benefit is tied to boundary behavior rather than to natural-parameter geometry per se."],"forward_implications":["Logit Flow Matching improves the primary distributional metrics over One-Hot Flow Matching on Adult, Cardio, and Buddy, and is comparable on Churn2.","Block-Conditional Logit Flow Matching consistently improves over flat Logit FM across all four datasets, showing that categorical representation and conditional factorization address complementary sources of error.","Logit Diffusion generally improves over or matches One-Hot Diffusion, with near ties on Buddy, suggesting the coordinate benefit transfers across generative dynamics.","Under severe rare-cell imbalance (rho = 0.01), logit coordinates reduce categorical total variation, Dmix, Dbal, conditional label error, and two-sample AUC relative to one-hot in controlled simulations.","The nonparametric rate shows rare-cell conditional components converge at rate (n pi*_k)^{-alpha_c/(2alpha_c+p+1)}, so preserving rare categorical mass is statistically harder than preserving common cells."],"fun_headline_variants":["Logit coordinates tame rare categories in tabular data","Logit Flow Matching wins on 3 of 4 mixed-data benchmarks","Smoothed logits make generative models respect categorical laws","Rare-cell bias tackled by logit encoding in generative models","Block-conditional logit flow boosts fidelity on tabular data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The theoretical guarantees are proved for an idealized componentwise generator—a separate logit flow plus per-categorical-class continuous flows—not for the flat joint Logit FM that produces several headline empirical results; if flat joint training cannot satisfy the needed conditional-stability assumptions, the reported gains rest on the experiments alone.","fun_headline_variants_meta":{"raw":{"variants":["Logit coordinates tame rare categories in tabular data","Logit Flow Matching wins on 3 of 4 mixed-data benchmarks","Smoothed logits make generative models respect categorical laws","Rare-cell bias tackled by logit encoding in generative models","Block-conditional logit flow boosts fidelity on tabular data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000277,"raw_usage":{"total_tokens":1474,"prompt_tokens":716,"completion_tokens":758,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":673}},"tokens_in":460,"tokens_out":758,"duration_ms":7411,"temperature":1.0,"reasoning_tokens":673,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T23:40:27.437947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the severe-imbalance simulation (rho = 0.01) with the ALR reference category cycled through all 24 joint cells and with the smoothing parameter eps varied over a grid; if there exists a reference choice or a eps value where One-Hot FM matches or beats Logit-FM on Dmix, the claimed representation-level advantage is not invariant to coordinate conventions.","supporting_citations":[],"review_version":1}