{"id":"0359212f-caad-40bf-8f6e-d5aabeee643a","arxiv_id":"2412.07324","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":10,"one_line_summary":"SNEFY-LDL models the conditional distribution of label distribution vectors on the simplex using the Squared Neural Family, with closed-form mean, variance and covariance.","lead":"The paper builds a probabilistic model that predicts a whole distribution of possible label distributions for each input, instead of just one label distribution. It gives closed-form formulas for the mean and variance of that distribution and uses them in conformal prediction, active learning, and ensemble learning experiments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MLE objective in Eq. (16) is undefined for any observed label distribution containing a zero component, because the model uses t1(ℓ)=log ℓ and the paper states no smoothing or positivity assumption; the experimental claims therefore rest on an unstated boundary condition.","rationale":"The reader identified the same weakest assumption: the sufficient statistic t1(ℓ)=log ℓ and the MLE objective require strictly positive label distribution components, while standard LDL datasets can contain zeros. My stress-test pass found no stronger objection that would change the verdict. The mathematical derivations in Theorems 2 and 3 are internally coherent under the stated W1 > -1/2 condition: the Dirichlet integral and the closed-form moments follow from standard Dirichlet expectations, and the Hadamard-product structure of the moment matrices is consistent with the proof. The conditional distribution is, however, only defined on the open simplex where log ℓ is finite, and the paper does not bridge the gap to observed label vectors on the boundary. The paper's own text contains no explicit positivity assumption or smoothing procedure, and the appendices describe data splits and algorithms without addressing boundary values. This is not a disagreement with an external consensus; it is an internal well-definedness gap in the training objective. Because the gap is potentially fixable by adding a stated positivity assumption, a small smoothing term, or a reformulated sufficient statistic, the appropriate verdict remains CONDITIONAL rather than ACCEPT or REJECT. The availability of source code is helpful and makes the proposed concrete test straightforward: one can check the actual data and code for zero handling. No issue with author conduct or novelty is raised.","tokens_in":22772,"tokens_out":3115,"duration_ms":38901,"concrete_test":"Inspect the released repository's data-loading and training code, and directly count zero entries in the label distribution vectors of the Movie, Natural Scene, SBU_3DFE, and SJAFFE datasets. If any training label has a zero component, check whether the code adds a positive smoothing constant before evaluating log ℓ in Eq. (6). If zeros occur and no smoothing is present, Eq. (16) will produce NaN or infinite losses; retraining with a small smoothing δ would be needed, and the sensitivity of Tables 2–7 to δ should be reported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that SNEFY-LDL can be trained by maximum likelihood on arbitrary LDL datasets. In §5, Theorem 2 sets t1(ℓ)=(log ℓ_y1,...,log ℓ_yL), and Algorithm 1 computes batched likelihoods with Eq. (6), which contains t1(ℓ). The MLE objective in Eq. (16) evaluates the density at the observed label distribution ℓ_x. If any component ℓ_xl = 0, then log(0) is undefined, and the exponential integrand in Eq. (9) is either zero, infinite, or undefined at that point depending on the sign of w1_il+w1_jl. The paper does not state that all observed label distributions lie in the open simplex, nor does it describe an epsilon-smoothing or boundary-avoiding preprocessing step. The benchmark datasets described in Table 1, especially facial-expression datasets such as SJAFFE and SBU_3DFE, are standard LDL datasets in which exact zero label proportions are common; the released code may include undocumented smoothing, but the manuscript does not. If the training data contain zeros and no smoothing is applied, the objective in Eq. (16) is not well-defined, so the fitted model and all downstream uncertainty-quantification experiments lack a valid training procedure. This is the load-bearing assumption: the closed-form moments in Theorem 3 and the empirical comparisons all presuppose a well-defined MLE, which in turn presupposes strictly positive observed label proportions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SNEFY-LDL, a conditional distribution model over label distribution vectors on the probability simplex, built on the squared neural family (SNEFY). With sufficient statistic t1(ℓ)=log ℓ, exponential activation, and Lebesgue base measure, the model obtains a closed-form normalizing constant (Theorem 2) and closed-form conditional mean, variance, and covariance (Theorem 3). Training is done by maximum likelihood (Eq. 16) with weight clipping to enforce W1 > -1/2. The paper reports experiments on conformal prediction, active learning, and ensemble learning across four LDL datasets, claiming that the closed-form moments enable instance-wise uncertainty quantification without sampling.","tokens_in":23111,"tokens_out":7628,"duration_ms":82087,"significance":"If the training procedure is well-defined, the contribution is useful: SNEFY-LDL moves beyond point estimation of label distributions to a full conditional distribution over the simplex, and the closed-form moments make uncertainty quantification cheap and instance-specific. The derivations in Theorems 2 and 3 are transparent and internally consistent, and the paper includes a code release and a parameter sensitivity study, which strengthen reproducibility. The main risk is that the maximum-likelihood objective, as stated, is undefined on boundary label vectors, which is a load-bearing gap for the experimental claims.","major_comments":[{"comment":"The MLE objective in Eq. (16) evaluates the conditional density at each observed label distribution ℓ_x′ through t1(ℓ)=log ℓ. If any component of an observed ℓ_x′ is zero, log is undefined; moreover, for ℓ_l→0 the unnormalized kernel in Eq. (6) behaves as ∏_l ℓ_l^{w1_il+w1_jl}, which can be zero, infinite, or undefined at the boundary. The paper states no assumption that all observed label distributions lie in the open simplex, and it describes no ε-smoothing or boundary-avoiding preprocessing. The benchmark datasets in Table 1, especially SJAFFE and SBU_3DFE, are facial-expression LDL datasets where zero label proportions are plausible, and the paper gives no reason to exclude them. The same boundary problem affects the ensemble weighting in Algorithm A4, which evaluates the density at base-learner predictions. This makes the training objective not well-defined on standard LDL data as presented. Please state and justify an explicit positivity assumption or a preprocessing step, or modify the model and objective to handle zero components; if the released code already implements smoothing, that must be described in the manuscript.","section":"§5, Eq. (16) and Algorithm 1"},{"comment":"The Dirichlet baseline is specified only as 'modeling the distribution ... using a Dirichlet distribution centered at the point prediction.' A Dirichlet distribution is determined by L concentration parameters; centering fixes their ratios but not their scale, and the scale controls the variance and hence the conformal interval width, coverage, and the resulting FSC scores. The manuscript does not state how the concentration parameters are estimated or set. Without this information the comparison is not reproducible, and the claimed greater adaptivity of SNEFY-LDL over the Dirichlet baseline is not established. The same underspecification applies to the Dirichlet active-learning baseline in §6.2.","section":"§6.1, Tables 2–3"}],"minor_comments":[{"comment":"The text says 'SEU_3DFE' but the dataset is called 'SBU_3DFE' in Table 6 and elsewhere; please unify the spelling.","section":"§6.3"},{"comment":"The dataset is called 'Motive' in Table 4 and Appendix B.2 but 'Movie' in Section 6.2, Table 1, and Table 5; please use a consistent name.","section":"Table 4 and Appendix B.2"},{"comment":"The header contains 'Clarky↓' where it should read 'Clark↓'.","section":"Table 7"},{"comment":"The horizontal axis is labeled 'bit_size' instead of 'bin_size'.","section":"Figure A1(d)"},{"comment":"The equality ∫_{Δ^{L-1}} ∏_l ℓ_l^{w1_il+w1_jl} dℓ = B(α) holds over the open simplex; please use consistent notation distinguishing the open simplex from the closed simplex Δ^{L-1} throughout the paper.","section":"Appendix A, proof of Theorem 2"},{"comment":"Differential entropy is estimated by importance sampling with a uniform proposal; no effective sample size or variance diagnostics are reported, so the reliability of the entropy estimates used for active learning is unclear.","section":"Algorithm A2"}],"recommendation":"major_revision","confidential_remarks":"The paper builds directly on SNEFY by two of the authors; the dependency is transparent and SNEFY is published, so I do not see a novelty-disclosure problem. The main technical obstacle is the boundary issue in the MLE objective, which must be resolved before the experimental claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: SNEFY-LDL is a genuinely new and mathematically clean way to put a tractable distribution on the probability simplex conditional on features, with closed-form means, variances and covariances. The theory holds up under the stated W1 > -1/2 condition. The empirical story is real, but the training objective has an unstated boundary assumption that needs fixing before the experiments can be fully trusted.\n\nWhat's new: the construction with t1(ℓ) = log ℓ and exponential activation turns the normalizing constant into a Dirichlet integral, and the moment formulas in Theorems 2 and 3 are correct and not in the previous LDL literature. That is a useful tool for uncertainty quantification in LDL—conformal intervals, active learning, and ensemble weighting all become possible without sampling. The paper also ships code and runs several datasets, which is more than most ML submissions do.\n\nSoft spots, in order of severity. First, the boundary: Eq. (16) evaluates the density at observed ℓ, and with t1(ℓ) = log ℓ, any zero component makes log undefined. The paper does not state that all training labels lie in the open simplex, nor does it give a smoothing procedure. The facial-expression datasets (SJAFFE, SBU_3DFE) commonly have exact zeros in label distributions. As written, the objective is not well-defined on those datasets. This is a load-bearing problem for the experimental claims, but it is fixable: an epsilon-smoothing preprocess, or a modified statistic that handles zeros, would likely resolve it. The authors should also state whether their released code does such smoothing.\n\nSecond, the Dirichlet conformal baseline is underspecified: \"centered at the point prediction\" does not say how the concentration parameter is chosen, and the comparison could be sensitive to that choice. The FSC tables also show SNEFY-LDL losing on several class/bin combinations, so the \"great effectiveness\" claim is a bit stronger than the tables support.\n\nThird, the differential entropy estimator in Algorithm A2 uses importance sampling with a uniform proposal on the simplex; the density p can be heavy-tailed, and no diagnostic is given for the estimator's variance. This is minor and standard practice, but worth a sentence.\n\nThe math itself is sound; I have no concern about Theorem 2 or 3. The self-reference to SNEFY is fine—it's a published, independently derived framework, and the new part is the simplex-specific integral.\n\nWho should read this: people working on LDL, compositional data, or tractable distributions on the simplex. It deserves a serious referee; the boundary issue is fixable and the core idea is worth engaging.\n\nRecommendation: send to peer review, but ask the authors to address the boundary assumption and the Dirichlet baseline specification before acceptance.","headline":"New tractable family on the simplex with closed-form moments; the theory is clean, but the training objective silently assumes positive label components and needs a boundary fix before the empirical claims can be trusted.","tokens_in":23636,"tokens_out":2038,"would_cite":true,"duration_ms":18416,"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":"A tractable distribution over label distributions, with closed-form moments, brings uncertainty quantification to label distribution learning.","keywords":["label distribution learning","squared neural family","probability simplex","uncertainty quantification","closed-form moments","conformal prediction","active learning","ensemble learning"],"falsifier":"Train SNEFY-LDL exactly as described on a standard LDL dataset that contains a label distribution with a zero component (for instance, an emotion absent from a facial-expression label vector). If the training objective in Eq. (16) is evaluated at that sample, $\\log 0$ is undefined and the density in Eq. (5) is not defined, so the published results can only be reproduced with an unstated preprocessing step; checking the released code for such a step would settle whether the modeling claim extends to the closed simplex or only to its interior.","tokens_in":22535,"feed_emoji":"📊","tokens_out":7995,"duration_ms":73563,"temperature":0.7,"pith_summary":"Label distribution learning (LDL) predicts a distribution over classes for each sample, but existing methods return only a single point estimate on the simplex. This paper argues that the right object is a distribution over all possible label distributions, and that the squared neural family (SNEFY) can provide it tractably when restricted to the simplex with an exponential activation and a logarithmic sufficient statistic. The central payoff is Theorem 3: closed-form conditional mean, variance, and covariance for each label's composition ratio, so the fitted model can be summarized and used for uncertainty quantification without sampling. The paper demonstrates the utility on conformal prediction, active learning, and ensemble learning, where SNEFY-LDL outperforms point-estimation baselines and a unimodal Dirichlet model.","feed_headline":"Closed-form moments yield label-distribution uncertainty","feed_subtitle":"Mean, variance, and covariance are computed in closed form, enabling confidence intervals and smarter active learning.","key_machinery":"The central object is the squared neural family (SNEFY) restricted to the probability simplex: $P(d\\ell|x) \\propto \\|V\\sigma(W_1 \\log \\ell + W_2 t_2(x) + b)\\|^2 d\\ell$. The argument is carried by the closed-form kernel integral in Eq. (9), which evaluates $\\int_{\\Delta^{L-1}} \\prod_l (\\ell_l)^{w_{1il}+w_{1jl}} d\\ell$ as the Dirichlet-type ratio $\\prod_l \\Gamma(1+w_{1il}+w_{1jl}) / \\Gamma(L+\\sum_l (w_{1il}+w_{1jl}))$. This identity, combined with the corresponding Dirichlet moments, yields the mean, variance, and covariance formulas of Theorem 3; the constraint $W_1 > -1/2$ keeps all Dirichlet parameters positive.","core_discovery":"SNEFY-LDL models the conditional law $P(d\\ell|x)$ of the label distribution vector $\\ell \\in \\Delta^{L-1}$ as a squared neural family distribution on the simplex, with exponential activation, sufficient statistic $t_1(\\ell)=(\\log \\ell_{y_1},\\dots,\\log \\ell_{y_L})$, and Lebesgue base measure. Under the elementwise constraint $W_1 > -1/2$, the kernel function integrates in closed form to a Dirichlet-type gamma ratio (Theorem 2), giving a tractable normalizing constant. Theorem 3 then provides closed-form expressions for $\\mathbb{E}[\\ell_{y_r}|x]$, $\\mathrm{Var}[\\ell_{y_r}|x]$, and $\\mathrm{Cov}[\\ell_{y_r},\\ell_{y_s}|x]$ as ratios of Hadamard products of the kernel matrix with moment matrices $F$, $G$, and $H$. This turns the fitted probabilistic model into an immediate source of prediction, confidence intervals, and label-correlation estimates, without Monte Carlo.","pith_inferences":["The Dirichlet-integral trick suggests SNEFY-LDL is essentially a mixture of Dirichlet-type components indexed by pairs of hidden units, so the closed-form moment formulas extend to any function whose Dirichlet expectation is known, such as entropy or cross-label moment bounds.","The strict positivity assumption could be relaxed by replacing $\\log \\ell$ with $\\log(\\ell+\\varepsilon)$ or by choosing a Dirichlet base measure, a modification that would make the model applicable to compositional data with zeros; the paper does not explore this.","The instance-wise density weighting for ensembles points toward a broader principle: the same model could score any candidate label distribution, not just base-learner outputs, opening a route to rejection or calibration of arbitrary LDL predictions.","If the closed-form moments are as accurate as claimed, they also provide a cheap initialization or regularizer for other simplex-valued models, since they give an analytic objective on the simplex without sampling."],"forward_implications":["Given a fitted model, per-label means, variances, and covariances for a new input $x$ are available in closed form in $O(mn^2 + Ln^2 + dD_2 + D_2^2)$ time, enabling real-time uncertainty quantification without sampling.","The variance can be combined with Chebyshev's inequality and conformal calibration to give instance-adaptive confidence intervals for each label's composition ratio.","The model's differential entropy, estimated by importance sampling, can drive max-entropy active learning, selecting the most informative unlabeled samples.","Base learner predictions in an ensemble can be weighted by their SNEFY-LDL density, yielding instance-wise ensemble weighting that outperforms uniform averaging in the paper's experiments.","Because the conditional distribution is a full density, the same fitted model supports reliability evaluation of a prediction via direct density values."],"supporting_citations":[{"why":"Supplies the SNEFY definition, the closed-form kernel table, and the conditional distribution theorem that SNEFY-LDL builds on.","marker":"[Tsuchida et al., 2023]"},{"why":"Defines the LDL problem, provides the benchmark datasets, the evaluation metrics, and the SA-BFGS baseline.","marker":"[Geng, 2016]"},{"why":"Provides the conformal prediction calibration procedure and the Feature-Stratified Coverage metric used in the experiments.","marker":"[Angelopoulos and Bates, 2021]"},{"why":"Supplies Chebyshev's inequality used to construct per-label confidence intervals from mean and variance.","marker":"[Grimmett and Stirzaker, 2020]"},{"why":"Supplies the Dirichlet distribution background used for the baseline and the Dirichlet moment identities underlying Theorem 3.","marker":"[Ng et al., 2011]"},{"why":"Supplies the weight clipping technique used to enforce the $W_1 > -1/2$ constraint during training.","marker":"[Arjovsky et al., 2017]"}],"fun_headline_variants":["SNEFY-LDL: closed-form moments for label distribution uncertainty","Probabilistic LDL with closed-form uncertainty moments","Squared neural family yields label distribution moments","Closed-form mean and variance for label distribution prediction","Tractable uncertainty for label distributions via SNEFY"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes every training label distribution has strictly positive components, because the sufficient statistic is the logarithm of each label proportion and the density is evaluated at the observed $\\ell$; the paper gives no smoothing or boundary handling for zero proportions, yet real LDL datasets often contain them.","fun_headline_variants_meta":{"raw":{"variants":["SNEFY-LDL: closed-form moments for label distribution uncertainty","Probabilistic LDL with closed-form uncertainty moments","Squared neural family yields label distribution moments","Closed-form mean and variance for label distribution prediction","Tractable uncertainty for label distributions via SNEFY"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1854,"prompt_tokens":986,"completion_tokens":868,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":793}},"tokens_in":602,"tokens_out":868,"duration_ms":62878,"temperature":1.0,"reasoning_tokens":793,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:56:10.280894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SNEFY-LDL exactly as described on a standard LDL dataset that contains a label distribution with a zero component (for instance, an emotion absent from a facial-expression label vector). If the training objective in Eq. (16) is evaluated at that sample, $\\log 0$ is undefined and the density in Eq. (5) is not defined, so the published results can only be reproduced with an unstated preprocessing step; checking the released code for such a step would settle whether the modeling claim extends to the closed simplex or only to its interior.","supporting_citations":[{"cited_title":"Squared neural families: a new class of tractable density models","cited_arxiv_id":null,"evidence_quote":"Supplies the SNEFY definition, the closed-form kernel table, and the conditional distribution theorem that SNEFY-LDL builds on."},{"cited_title":"Label distribution learning","cited_arxiv_id":null,"evidence_quote":"Defines the LDL problem, provides the benchmark datasets, the evaluation metrics, and the SA-BFGS baseline."},{"cited_title":"Probability and Random Processes","cited_arxiv_id":null,"evidence_quote":"Supplies Chebyshev's inequality used to construct per-label confidence intervals from mean and variance."}],"review_version":1}