{"id":"a5098997-8716-441a-ba6e-e5c40153e562","arxiv_id":"2505.21180","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 3D grid representation built from Gaussian-expanded label differences improves label distribution learning and robustness to noise.","lead":"The paper proposes a new way to represent label uncertainty in machine learning by arranging label differences into a 3D grid where each value is spread out as a Gaussian distribution. This grid is then cleaned up with a low-rank technique and used to predict more accurate label distributions, with competitive results on 15 benchmarks and a medical image dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gaussian prior target is unbounded while B is Tanh-bounded to (-1,1); a large fraction of sampled B_hat entries cannot be represented, so the claimed low-noise representation is not realized as written.","rationale":"The reader's weakest assumption identified the meaningfulness of the hand-crafted Gaussian prior. I sharpen this into a concrete, internally testable issue: even if the prior were meaningful, its support does not match the representable range of B. This is an internal inconsistency rather than a disagreement with external consensus, and it is directly checkable from the paper's own equations. Because the fix is straightforward (clip or truncate the target, or remove the Tanh bound) and the empirical study is otherwise broad, the appropriate disposition remains conditional acceptance rather than rejection. Additional concerns such as the anomalous Table 3 Yeast-cdc Chebyshev entry (0.1587 vs. roughly 0.016 for all baselines) and the unspecified Tucker ranks reinforce the need for revision but are secondary to the range mismatch.","tokens_in":17216,"tokens_out":10124,"duration_ms":101823,"concrete_test":"Reimplement Eq. (2) on the provided label distributions (e.g., Yeast-alpha and Human Gene) with a fixed random seed; compute the fraction of entries of B_hat with absolute value greater than 1 and the minimal achievable Loss_g if B is optimally clipped to [-1,1]. If the fraction is nontrivial (e.g., >5%) or the floor loss is nonzero, clip B_hat to [-1,1] and rerun the Table 3/4 experiments; material changes in results would show that the bounded range, not the Gaussian prior, is responsible for the reported behavior.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction is internally inconsistent as written. Eq. (1) constrains the generated grid to B in (-1,1) via Tanh, explicitly 'to avoid outliers due to stochasticity.' Eq. (2) builds the target grid B_hat by Gaussian sampling with mean a_ij = d_j - d_i and variance 1 - |a_ij|. For diagonal entries a_ii = 0, the variance is 1; for many off-diagonal entries with small |a_ij|, the variance is close to 1. A Gaussian with mean 0 and variance 1 has roughly 32% of its mass outside (-1,1), and for means in 0.5-0.9 the probability is approximately 26-38%. Therefore a substantial fraction of entries of B_hat lie outside the representable range of B. The L2 grid loss in Eq. (8) cannot vanish, and the learned B cannot realize the Gaussian prior; it can at best be the clipped projection of B_hat onto [-1,1]. The paper never specifies clipping or truncation of B_hat, so the claimed 'low-noise representation space' driven by Gaussian uncertainty is not actually representable, and the mechanism behind the reported 17.7% contribution of Loss_g is not what is described.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Latent Label Distribution Grid (LLDG) representation for label distribution learning (LDL). For each sample, a pairwise label difference matrix D is computed from the ground-truth label distribution, and each entry a_ij is expanded into a length-c vector sampled from a Gaussian with mean a_ij and variance 1-|a_ij|, forming a c×c×c target grid B_hat. A network with 1D convolutions and a Transformer produces a grid B bounded by tanh to (-1,1); B is regularized by B_hat through an L2 loss and by a low-rank Tucker reconstruction. The regularized grid is then compressed by an LLDG-Mixer into a predicted label distribution. Experiments cover 15 LDL datasets and a MedMNIST classification extension, with claims of competitive accuracy, noise robustness, and stable representation.","tokens_in":17521,"tokens_out":6354,"duration_ms":81822,"significance":"If the construction were fully validated, the combination of label-relation expansion, Gaussian uncertainty modeling, and Tucker low-rank regularization would be a novel representation-level contribution to LDL. The paper has several concrete strengths: it compares against seven LDL baselines on 15 datasets, includes an ablation study, a noise-disturbance experiment, an extension to MedMNIST, and states that code and datasets are released. The low-rank tensor regularization is a distinctive ingredient. However, the current empirical support is weakened by an apparent data-entry error, statistically insignificant ablation differences, and an underexplained classification extension; the central representational mechanism also has a mismatch between the unbounded Gaussian target and the bounded generated grid. These issues are fixable but are load-bearing for the paper's claims.","major_comments":[{"comment":"The generated grid B is constrained by tanh to lie in (-1,1), while the target grid B_hat is sampled from a Gaussian with mean a_ij and variance 1-|a_ij|. For diagonal entries a_ij=0, the variance is 1, so about 32% of sampled values lie outside (-1,1); for off-diagonal entries with |a_ij| between roughly 0.5 and 0.9, the fraction outside (-1,1) ranges from about 26% to 38%. No clipping or truncation of B_hat is specified. Consequently, a substantial part of the target cannot be represented by B, the L2 grid loss in Eq. (8) cannot approach zero by construction, and the learned B is at best a projection of B_hat onto [-1,1]. This contradicts the paper's description of B being 'bounded ... to avoid outliers' while being driven by the Gaussian prior. Please specify the target construction precisely (e.g., truncation, a different activation, or a principled reason for allowing unrepresentable targets), or the claimed 'low-noise representation space' is not realized as written.","section":"§3, Eq. (1), Eq. (2), Eq. (8)"},{"comment":"The reported Chebyshev distance for Ours on Yeast-cdc is 0.1587±0.0004, which is an order of magnitude worse than all baselines (which are near 0.016) and inconsistent with the paper's claim that LLDG outperforms the neural-network baselines on most datasets. The tight standard deviation and the healthy values on the other five metrics strongly suggest a typo, most likely a missing leading zero (0.01587). Please correct the entry and re-verify the corresponding comparison; as printed, this single row undermines the cross-dataset claim.","section":"Table 3, Yeast-cdc row"},{"comment":"The ablation results do not support the stated conclusion that 'the Tucker reconstruction algorithm has significant benefits.' For example, Chebyshev is 0.0522±0.0011 with LLDG versus 0.0524±0.0009 without LLDG and 0.0529±0.0011 without Tucker; Cosine is 0.8368±0.0027 versus 0.8346±0.0019 versus 0.8349±0.0008. These differences are all within one standard deviation. Please report significance tests, confidence intervals, or repeated-run distributions, and temper the claim accordingly. In addition, the 'w/o LLDG' condition is described only as 'modify the output dimension of LLDG-Mixer and remove the Lossg'; this does not specify what remains of the grid architecture, making the ablation difficult to interpret.","section":"Table 5 and §4, Ablation study"},{"comment":"The MedMNIST extension is a central part of the claimed generality, but the experimental setup is not reproducible from the text. The paper says 'for every single label, by modeling the Gaussian prior is expanded into a vector shape as a learning target for the LLDG,' yet it does not specify how integer class labels are converted into label distributions, what the target grid is for a classification task, what classifier head is used after the grid, or the network input representation for medical images. Without these details, Table 7 cannot be checked or reproduced. Please provide a complete description, ideally in the main text or appendix, including any label-distribution construction and training hyperparameters.","section":"§4, Potential of LLDG and Table 7"},{"comment":"The noise-robustness claim is not supported by the reported experiment. Table 6 shows only LLDG's performance under Gaussian noise of increasing standard deviation; no baseline method is evaluated under the same noise conditions. The statement that 'it is still competitive compared to other methods' therefore has no empirical basis in the table. Please add comparable baselines under the same noise protocol, or explicitly restrict the claim to the stability of LLDG itself.","section":"§4, Noise disturbance and Table 6"}],"minor_comments":[{"comment":"Equation (8) contains a stray closing parenthesis: '||Bijk − ˆBijk )||2' should be '||Bijk − ˆBijk||2'.","section":"§3, Eq. (8)"},{"comment":"In the text describing Tucker decomposition, the second factor matrix is called 'U' twice, and then 'V' is introduced; the factor matrices in Eq. (3) are U, V, and X, so the text should say V for the second factor.","section":"§3, Eq. (3)"},{"comment":"In Table 4, the SBU-3DFE rows for IIS-LLD and Duo-LDL are identical to four decimal places across all metrics, which appears to be a transcription error; please verify these entries.","section":"§4, Visualization and Table 4"},{"comment":"The text says 'a Gaussian noise with a variance of 1 is added,' but Table 6 reports standard deviations of 0.1, 0.2, 0.5, and 1.0. Please clarify whether the parameter is the standard deviation or the variance.","section":"§4, Noise disturbance"},{"comment":"The dataset name 'Yest-dtt' in the visualization section should be 'Yeast-dtt'.","section":"§4, Visualization"},{"comment":"The sentence 'The experimental results are summarized in Table 4' should refer to Table 5, which contains the ablation results.","section":"§4, Ablation study"},{"comment":"The paper states 'we do not make any prior assumptions about label correlation' but then constructs a label correlation matrix based on differences between label values; this is itself a modeling assumption and should be acknowledged.","section":"§3, Latent label distribution grid"},{"comment":"The variance choice 1-|a_ij| is justified by saying 'the clearer the relationship between label values, the smaller the uncertainty,' but under this rule equal label values (a_ij=0) receive the largest variance, which seems to say that identical labels have the most uncertain relationship. Please clarify the intended interpretation.","section":"§3, Latent label distribution grid"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to build closely on the authors' own prior works, references [6] and [25], and the novelty relative to those works should be checked by the editor. The comparison protocol also relies heavily on numbers extracted from other papers; it would be worth asking the authors to confirm that the train/test splits and evaluation code are identical for all methods. The unbounded-target/bounded-grid mismatch in Section 3 is the most substantive technical concern and should be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper has a plausible idea and a lot of experiments, but the core construction has an internal inconsistency that the authors have not addressed. Worth a referee, but it needs a real revision.\n\nWhat's actually new: LLDG extends the authors' earlier 2D label distribution matrix to a 3D grid: build a label difference matrix, expand each entry into a Gaussian-distributed vector with variance 1-|a_ij|, and regularize the resulting grid with Tucker low-rank reconstruction. That is a natural continuation of [6] and [25], and the paper is honest that it builds on those. The experimental sweep across 15 LDL datasets plus the MedMNIST classification results is genuinely broad.\n\nWhat's good: the method is mostly clearly described, the ablation on Human Gene shows the grid and Tucker contribute something, and the noise-disturbance table suggests robustness. The limitations section is refreshingly direct about compute cost.\n\nThe soft spots, in order of severity:\n\n1. The central construction doesn't work as written. Eq. (1) bounds B to (-1,1) with Tanh. Eq. (2) builds the target grid B_hat by Gaussian sampling with variance 1-|a_ij|, which for diagonal entries (variance 1, mean 0) puts about 32% of the mass outside (-1,1), and similar for small off-diagonal entries. So a substantial fraction of B_hat entries cannot be represented in B. The L2 grid loss cannot vanish; the learned B can at best be a clipped projection. The paper never specifies clipping or truncation, so the mechanism behind the 'low-noise representation space' is not what is described. This is a load-bearing issue, not a nitpick.\n\n2. The reported Chebyshev distance of 0.1587 for Yeast-cdc in Table 3 is an order of magnitude worse than every baseline and is almost certainly a typo. That sort of error makes the tables hard to trust.\n\n3. The abstract promises 'executable code and datasets are released in the Appendix,' but there is no appendix in the arXiv version. That should be fixed.\n\n4. Tucker ranks R1,R2,R3 are never given. Lambda is stated as 0.5, but no sensitivity analysis.\n\n5. The novelty over [6] and [25] is not clearly delimited. The paper says it integrates label correlation and extended representation, but the reader has to reverse-engineer what exactly is new.\n\n6. The LDL baselines are mostly classical; there are no modern deep LDL competitors, so the comparison is weaker than it looks.\n\nVerdict: this deserves a serious referee because the idea is plausible and the experiments are broad, but it needs major revision. If the Gaussian-prior inconsistency is fixed and the missing details supplied, it could be a modest, citable contribution. As is, I'd send it back for revision rather than reject outright.","headline":"A broadly evaluated but internally inconsistent grid representation for label uncertainty; the mechanism fails as written, though the empirical work is substantial.","tokens_in":18020,"tokens_out":4375,"would_cite":false,"duration_ms":45967,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A 3-D label grid turns noisy labels into a low-noise target, the paper argues.","keywords":["label distribution learning","latent representation","label uncertainty","Gaussian expansion","Tucker reconstruction","low-rank regularization","noise robustness","MedMNIST"],"falsifier":"Train the same pipeline with the grid prior computed from randomly shuffled label values, so the difference matrix no longer reflects true label relations. If accuracy and noise robustness stay essentially unchanged, the specific label-difference prior is not what carries the result; if they degrade, the prior is doing real work. A second check is to sweep the Tucker ranks: if performance is flat across very different ranks, the low-rank constraint is not the active noise-reduction mechanism.","tokens_in":17005,"feed_emoji":"🧊","tokens_out":5238,"duration_ms":51543,"temperature":0.7,"pith_summary":"Label distribution learning assumes annotations are clean, but real label distributions are expensive to collect and often inexact. This paper proposes a latent label distribution grid (LLDG): a three-dimensional tensor in which every pairwise label difference is expanded into a Gaussian vector, so the grid encodes both label relations and their uncertainty. The learned grid is pulled toward this Gaussian prior and simultaneously passed through a Tucker decomposition–reconstruction step to enforce low rank and strip noise. The central claim is that this grid representation forms a low-noise learning space, yielding more accurate and noise-robust label distributions than existing LDL methods on a series of benchmarks.","feed_headline":"A 3-D label grid turns noisy labels into a low-noise target","feed_subtitle":"The grid expands every label difference into a Gaussian and uses Tucker reconstruction to strip noise, beating baselines on 15 datasets.","key_machinery":"The central object is the latent label distribution grid (LLDG): a $c\\times c\\times c$ tensor $B$. Its construction combines (1) a label difference matrix $D$ with entries $a_{ij}=d^j-d^i$ from the true label distribution; (2) a Gaussian expansion that turns each entry into a length-$c$ vector with mean $a_{ij}$ and variance $1-|a_{ij}|$, producing the prior grid $\\hat{B}$; and (3) a Tucker decomposition–reconstruction operator $B^*=\\operatorname{Tucker}^{[R]}(\\operatorname{Tucker}^{[D]}(B))$ that enforces low rank on the learned grid. A local-global feature extractor (1D convolutions plus a Transformer) generates $B$ from the input features, and an LLDG-Mixer with three linear layers squeezes the grid along its axes to output the label distribution. The loss combines an L2 label-distribution error with an L1 grid-prior error, weighted by 0.5.","core_discovery":"The paper's central claim is that modeling uncertainty directly in the label representation space—rather than only in the loss function—improves LDL accuracy and robustness. To do this, it builds a $c\\times c\\times c$ grid $B$ from the sample's features, then regularizes $B$ toward a prior grid $\\hat{B}$ obtained by taking the pairwise difference matrix $D$ of the ground-truth label distribution and replacing each entry $a_{ij}$ by a length-$c$ Gaussian sample with mean $a_{ij}$ and variance $1-|a_{ij}|$. Because the grid is a tensor, a Tucker decomposition–reconstruction scheme enforces low-rank structure, which the authors describe as noise reduction. A small mixer network then collapses the regularized grid into a predicted label distribution. The authors report the method outperforms or matches seven LDL baselines on 15 datasets, degrades gracefully under added Gaussian noise, and transfers competitively to MedMNIST classification.","pith_inferences":["Inference: the Gaussian prior is anchored to ground-truth label differences, which makes the training signal depend on label quality; a natural extension is to bootstrap the prior from predicted distributions for unlabeled or partially labeled samples.","Inference: the variance rule $1-|a_{ij}|$ asserts that similar labels are confidently related and dissimilar labels are uncertain; testing alternative variance schedules (for instance, variance proportional to observed disagreement) could show whether this specific schedule matters.","Inference: the reported gains are small in absolute terms on several Yeast datasets, so a practical reading is that LLDG buys robustness more than raw accuracy; comparing area under the curve over noisy regimes rather than point metrics would clarify the trade-off.","Inference: since the Tucker ranks are never specified, the method's reproducibility depends on them; an explicit rank-selection rule would turn the method into a drop-in module for other LDL pipelines."],"forward_implications":["On the 15 LDL datasets, the method posts the best or tied-best score on the majority of the six evaluation metrics, with the largest margins on wc-LDL and SBU-3DFE.","The grid representation keeps the model stable under Gaussian label noise with standard deviation up to 1.0, with only small metric degradation.","LLDG can serve as an upstream representation for classification: on MedMNIST it achieves the best or near-best AUC and accuracy on most of the ten tasks among the reported methods.","Ablation on the Human Gene dataset shows that removing the grid or the Tucker reconstruction worsens all metrics, supporting that both components contribute.","Because the grid is built from labels alone, the representation can be attached to any feature extractor, not just the proposed CFormer."],"supporting_citations":[{"why":"Supplies the prior idea of extending a label distribution vector into a matrix via Gaussian sampling, which LLDG generalizes to a grid.","marker":"[25]"},{"why":"Introduces the label correlation grid concept that the LLDG construction directly builds upon.","marker":"[6]"},{"why":"Provides the Transformer architecture used as the global feature extractor in the CFormer pipeline.","marker":"[3]"},{"why":"Defines the label distribution learning paradigm and the six evaluation metrics used throughout the experiments.","marker":"[5]"},{"why":"Demonstrates how label correlation can boost LDL performance, motivating the grid's relation-based regularization.","marker":"[4]"},{"why":"Motivates the noise-robustness goal by showing that label noise is a known problem in LDL.","marker":"[12]"}],"fun_headline_variants":["Grid of Gaussian differences tames noisy label distributions","Low-rank label grid cuts noise in label distribution learning","Label grid with Gaussian spread models uncertainty, beats baselines","From noisy labels to clean targets via a 3D grid","Tucker-smoothed label grid ups accuracy on 15 datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's success rests on the assumption that a Gaussian prior built from ground-truth label differences—with variance set to one minus the absolute difference—captures the true uncertainty structure of the label space, so that pulling the learned grid toward it removes noise rather than information.","fun_headline_variants_meta":{"raw":{"variants":["Grid of Gaussian differences tames noisy label distributions","Low-rank label grid cuts noise in label distribution learning","Label grid with Gaussian spread models uncertainty, beats baselines","From noisy labels to clean targets via a 3D grid","Tucker-smoothed label grid ups accuracy on 15 datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000688,"raw_usage":{"total_tokens":3146,"prompt_tokens":998,"completion_tokens":2148,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":614,"completion_tokens_details":{"reasoning_tokens":2067}},"tokens_in":614,"tokens_out":2148,"duration_ms":14892,"temperature":1.0,"reasoning_tokens":2067,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:33:23.497399+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same pipeline with the grid prior computed from randomly shuffled label values, so the difference matrix no longer reflects true label relations. If accuracy and noise robustness stay essentially unchanged, the specific label-difference prior is not what carries the result; if they degrade, the prior is doing real work. A second check is to sweep the Tucker ranks: if performance is flat across very different ranks, the low-rank constraint is not the active noise-reduction mechanism.","supporting_citations":[{"cited_title":"Label Distribution Learning via Implicit Distribution Representation","cited_arxiv_id":"2209.13824","evidence_quote":"Supplies the prior idea of extending a label distribution vector into a matrix via Gaussian sampling, which LLDG generalizes to a grid."},{"cited_title":"Label distribution learning via label correlation grid","cited_arxiv_id":"2210.08184","evidence_quote":"Introduces the label correlation grid concept that the LLDG construction directly builds upon."},{"cited_title":"An image is worth 16x16 words: Transformers for image recognition at scale","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer architecture used as the global feature extractor in the CFormer pipeline."},{"cited_title":"Label distribution learning.IEEE TKDE, 2016","cited_arxiv_id":null,"evidence_quote":"Defines the label distribution learning paradigm and the six evaluation metrics used throughout the experiments."},{"cited_title":"Age estimation using expectation of label distribution learning","cited_arxiv_id":null,"evidence_quote":"Demonstrates how label correlation can boost LDL performance, motivating the grid's relation-based regularization."},{"cited_title":"Label distribution learning with noisy labels via three-way decisions.International Journal of Approximate Reasoning, 2022","cited_arxiv_id":null,"evidence_quote":"Motivates the noise-robustness goal by showing that label noise is a known problem in LDL."}],"review_version":1}