{"id":"1a981733-20f8-4e1e-8428-b03866af96d3","arxiv_id":"2501.13756","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A weighted linear combination of SCL, RSG, and LDAM losses is shown to improve tail-class accuracy while largely preserving head-class accuracy on several long-tailed benchmarks.","lead":"The paper combines three existing long-tail recognition techniques, supervised contrastive learning, a rare-class sample generator, and a label-distribution-aware margin loss, by summing their loss functions with tuned weights. It reports balanced accuracy gains across head and tail classes on four long-tailed image datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RSG loss weights 1e-5/1e-6 are reported insensitive yet Table 6 attributes gains to RSG; the causal role of RSG is unsupported unless the module's forward path is shown to affect training.","rationale":"The reader's weakest assumption is exactly the point I would stress. The paper's central claim is a three-way synergy, so each component must have an identifiable causal contribution. For RSG, the only reported mechanism terms are LCESC and LMV (Eq. 6), but the paper says the model is insensitive to their weights and sets them to 1e-5/1e-6. If those losses contribute negligibly, then the observed accuracy and ICD differences between SCL-LDAM and RSG-SCL-LDAM must be explained by something else. The most plausible alternative is that the RSG module changes the forward computation (e.g., by generating synthetic tail features that are then used by SCL or the classifier), but this channel is never stated or isolated in the paper. Without that isolation, the ablation results in Table 6 and the ICD evidence in Table 2 cannot be attributed to the mechanism claimed in Section 3.2. This is a correctness risk in the causal narrative, not merely a missing baseline. It is also empirically checkable: run the Table 6 ablation with η=μ exactly zero, and with the generation module disabled but losses present. The paper's own reporting makes this check necessary before the synergy claim can be accepted. Other issues (no error bars, validation-set-tuned weights, trailing BCL) are secondary and already reflected in the CONDITIONAL verdict; they do not change my assessment. I do not see a reason to move to reject or accept: the empirical numbers, if reproducible, still show a useful combination, but the RSG attribution must be pinned down.","tokens_in":15492,"tokens_out":3827,"duration_ms":34210,"concrete_test":"Run the mini-ImageNet-LT ablation from Table 6 with three configurations: (i) SCL-LDAM without the RSG module; (ii) RSG-SCL-LDAM with η=μ=0.0 exactly, not 1e-5/1e-6; (iii) RSG-SCL-LDAM with the reported η=1e-5, μ=1e-6. Use at least 3 seeds and report Many/Medium/Few/All accuracy. If (ii)≈(iii) and both differ from (i), the gains come from the module's forward pass/architecture, not from the RSG loss terms, and the paper's stated insensitivity requires an explanation of that pathway. If (ii)≈(i), the RSG effect vanishes when its losses are exactly zero, contradicting the reported ablation and the synergy claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 states that the model is insensitive to the weights η for LCESC and μ for LMV, and therefore sets them to 1e-5 and 1e-6 in Eq. 6. Section 4.3/Table 6 attributes the accuracy gap between SCL-LDAM (95.22) and RSG-SCL-LDAM (95.61) to RSG, and Table 2 attributes ICD changes to RSG. These two statements are in tension. If the RSG loss terms are truly negligible at these weights, the ablation gap cannot be caused by LCESC or LMV; it would have to come from the RSG module's forward pass (e.g., generated tail samples entering the batch) or from uncontrolled differences. The paper does not describe such a pathway. This matters because the central claim is a synergistic and compensatory relationship among SCL, RSG, and LDAM; if RSG's only measurable effect is through an unstated forward-path channel, the mechanism story (RSG expands tail feature space, SCL clusters, LDAM shifts boundary) is not supported. The same issue weakens Table 2's ICD evidence: the comparison 'without RSG' vs 'with RSG' changes both the architecture and the loss, so the ICD decrease cannot be unambiguously assigned to the RSG losses. A minimal fix is to test η=μ=0 exactly, and to ablate the RSG module's sample generation independently of its loss weights.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end long-tailed recognition model that minimizes a weighted linear combination of four losses, L_total = αL_SCL + λL_LDAM + ηL_CESC + μL_MV (Eq. 6), where L_SCL is supervised contrastive loss, L_LDAM is label-distribution-aware margin loss, and the latter two terms come from the Rare-Class Sample Generator (RSG). The authors report class-grouped accuracy on CIFAR-10-LT, CIFAR-100-LT, mini-ImageNet-LT, and ImageNet-LT, and claim that the three techniques act synergistically and compensatorily: RSG expands the tail-class feature space, SCL clusters same-class features and separates classes, and LDAM shifts the decision boundary toward tail classes, thereby improving tail accuracy without sacrificing head accuracy. The evidence consists mainly of ablation tables, an intra-class distance (ICD) measurement, and qualitative reasoning in Section 3.2.","tokens_in":15764,"tokens_out":5908,"duration_ms":54514,"significance":"If the claimed synergy is real, the paper offers a simple, compute-efficient recipe (Table 4) for improving tail-class accuracy while preserving head-class accuracy, which would be practically useful. The paper is commendably transparent about the loss formulas, training configurations, and hyperparameter-search procedure, and it reports class-grouped results that go beyond single-number accuracy. However, the current empirical support is not yet convincing: the RSG loss weights are reported to be insensitive and set to extremely small values while the ablation attributes measurable gains to RSG; the mini-ImageNet experiments use a model pretrained on a superset that likely contains the mini-ImageNet classes; and the per-dataset tuning of α and λ without error bars leaves the main comparisons vulnerable to selection effects. The paper's contribution is better described as an empirical study of a loss combination than as a new method with a demonstrated mechanism.","major_comments":[{"comment":"The statement that \"the model exhibits insensitivity to the weights η for LCESC and μ for LMV\" and the decision to set η=1e-5 and μ=1e-6 (Section 4, first paragraph) are in tension with the ablation in Table 6, where removing RSG changes the overall accuracy from 95.61 to 95.22 and the Few-shot accuracy from 95.42 to 93.17, and with Table 2, where RSG is credited with reducing the average ICD from 1.55 to 1.39. If the two RSG losses are truly negligible at these weights, the observed effect must come from the RSG module's forward path (e.g., generated tail samples entering the training batch or feature-space changes), but the paper does not describe or measure such a pathway. Please run the control η=μ=0 with the RSG module still active, and separately ablate the module's sample generation while keeping the losses, so that the observed gains can be assigned to a specific mechanism. Without this, the claimed synergistic role of RSG is not supported by the reported experiments.","section":"Section 4, first paragraph; Tables 2 and 6"},{"comment":"For mini-ImageNet-LT, the authors use a ResNeXt-50-32x4d backbone pretrained on ImageNet-1K. Mini-ImageNet is commonly constructed as a subset of ImageNet-1K, and the paper does not state whether the 100 mini-ImageNet classes are contained in the pretraining class set. If they are, the reported Few-shot accuracy of 95.42% in Table 6 and the RSG ablation results are not a valid test of generalization to unseen classes and are likely inflated by test-class leakage. Please disclose the exact mini-ImageNet class split and its overlap with the pretraining classes; if overlap exists, rerun the mini-ImageNet experiments with a pretraining set that excludes those classes or with a backbone not pretrained on the same classes. This point is load-bearing because Table 6 is the primary ablation supporting the synergy claim.","section":"Section 4.2, mini-ImageNet-LT"},{"comment":"The weights α and λ for L_SCL and L_LDAM are optimized separately for each dataset and each imbalance factor β using a genetic algorithm, with candidate combinations evaluated on validation accuracy for only five epochs before the final runs. The reported test numbers therefore reflect post-hoc selection, as the Limitation in Section 5 acknowledges. In addition, the tables contain no error bars or multiple-seed results, and some differences are very small (e.g., Table 7: Many-shot 66.8 vs. 66.6, Few-shot 21.9 vs. 23.2). Please report the variance across at least three runs for the main configurations, and either fix the weighting rule a priori or treat the genetic-algorithm search as part of the method with nested validation. Without this, the claimed \"balanced improvement across all classes\" is not robustly established.","section":"Section 4.2, Table 3, and Supplementary Material"}],"minor_comments":[{"comment":"Equation (2) uses the spelling \"LCECS\" while the text uses \"LCESC\", and the angle-bracket notation for averaging in Equations (2) and (3) is not formally defined; please clarify these notations for reproducibility.","section":"Equations (2) and (3)"},{"comment":"The column header \"A VG.\" should read \"Avg.\", and the class labels 0-9 should be explicitly stated to be ordered from the most frequent to the least frequent class, since this ordering is necessary to interpret the ICD pattern.","section":"Table 2"},{"comment":"The text says CIFAR-10 and CIFAR-100 contain 10,000 images for validation, but these datasets are normally partitioned into training and test sets; please clarify whether the reported \"validation\" numbers actually refer to the standard test split or to an internal validation split.","section":"Section 4.1"},{"comment":"References [15] and [16] appear to be the same work (Kang et al., \"Decoupling representation and classifier for long-tailed recognition\") listed twice with different reference numbers; please merge them or clarify why they are distinct.","section":"References"},{"comment":"The Limitation paragraph should be integrated into the main experimental section as well, since the substantial cost of hyperparameter search directly conditions the interpretability of the reported results and should be disclosed wherever the tuned weights are presented.","section":"Section 5, Limitation"}],"recommendation":"major_revision","confidential_remarks":"The most serious issue is the likely test-class leakage in the mini-ImageNet experiment; the authors should be required to report the class overlap with ImageNet-1K pretraining. The RSG loss-weight inconsistency is also central and should be resolved with the proposed control experiments. The novelty is modest relative to Hybrid-SC and BCL, but the paper could still be a useful empirical contribution if the evidence is made robust."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible but modest engineering paper. The genuinely new thing is the exact combination of SCL, RSG and LDAM as a weighted loss sum, plus a few ablation measurements that were missing from the literature. It does not reorganize the field, but it is a useful baseline for anyone doing long-tail recognition with balanced accuracy in mind.\n\nThe paper does some things well. The research gap it identifies is real and simply stated: Hybrid-SC is SCL+CE, BCL is a modified SCL plus logit adjustment, but nobody had reported the plain SCL+LDAM combination or the addition of RSG. The mini-ImageNet ablation is clean: adding SCL to RSG-LDAM improves every group, replacing CE with LDAM helps the few-shot classes, and the full model improves many/medium/few simultaneously. The FLOPs/parameter comparison with BCL is fair. And the supplementary material is unusually honest about the genetic-algorithm weight search and its limitations; the limitation section explicitly admits suboptimal weights hurt.\n\nThe soft spots are real but mostly fixable. The stress-test concern lands: Section 4 says the model is insensitive to eta and mu, then sets them to 1e-5 and 1e-6, but Table 6 attributes a 2.2-point few-shot gain to adding RSG. If the RSG losses are negligible at those weights, the gain has to come from the RSG forward path — probably generated tail features entering the batch for LDAM and SCL. The paper never says that. This is not fatal; it is likely the actual mechanism from the original RSG paper. But the authors should state it directly and, ideally, ablate eta=mu=0 exactly or ablate sample generation independently of loss weights. Second, the synergy narrative is assembled post hoc from the same experiments, and the sole mechanism evidence is one ICD table that confounds architecture with loss changes. No error bars, and alpha and lambda are tuned per dataset on the validation set, so the test numbers are partly selected. Those caveats lower confidence in the precise numbers, not in the overall direction.\n\nWho is this for? People working on long-tail recognition in applications where head-class accuracy matters. It is a reasonable empirical contribution and deserves a serious referee, but it needs a revision: explain the RSG forward path, report multiple seeds, add the missing SCL-family baselines. I would not cite it in my own work in the next year, but I would not desk-reject it.","headline":"A credible but modest engineering paper: the SCL+RSG+LDAM combination is new and the ablations are honest, but the synergy claim is post hoc and the RSG ablation needs a clarified mechanism.","tokens_in":16318,"tokens_out":4188,"would_cite":false,"duration_ms":36656,"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":"This paper claims that a weighted sum of SCL, RSG, and LDAM losses yields balanced gains across all classes in long-tailed image classification, improving tail accuracy without sacrificing head accuracy.","keywords":["long-tailed recognition","supervised contrastive learning","rare-class sample generator","label-distribution-aware margin loss","loss combination","feature space","class imbalance","decision boundary"],"falsifier":"Measure the gradient norm contributed by the two RSG losses at the chosen weights at several training checkpoints on mini-ImageNet-LT; if it is orders of magnitude below the SCL and LDAM gradient norms, the ablation gains attributed to RSG cannot be explained by RSG's loss signal.","tokens_in":15231,"feed_emoji":"⚖️","tokens_out":10321,"duration_ms":78836,"temperature":0.7,"pith_summary":"This paper tries to show that three established remedies for long-tailed image classification—supervised contrastive learning, a rare-class sample generator, and a label-distribution-aware margin loss—work better together than any of them alone. The claimed reason is a mutually compensatory geometry: SCL pulls each class into a tight cluster, RSG manufactures new tail features to refill the tail region that SCL compresses, and LDAM then tilts the decision boundary toward tail classes without misclassifying head samples because the clusters are already well separated. On CIFAR-10-LT, CIFAR-100-LT, mini-ImageNet-LT, and ImageNet-LT, the combined loss raises few-shot accuracy while keeping many-shot accuracy roughly flat, which the paper presents as a balanced improvement across all classes. The practical stake is that a simple weighted sum of existing losses could replace more complex two-stage or ensemble long-tail pipelines.","feed_headline":"Three losses together lift tail accuracy without head-class loss","feed_subtitle":"SCL clusters classes, RSG expands tail space, LDAM adjusts boundaries for balanced gains across long-tail benchmarks.","key_machinery":"The load-bearing object is the four-term loss in Eq. 6, where the SCL loss (supervised contrastive loss, which clusters same-class features and separates different classes) is combined with the LDAM loss (a label-distribution-aware margin loss that gives classes with fewer samples a larger decision margin) and the two RSG losses: CESC (center estimation with sample contrastive, pulling features toward class centers) and MV (maximized vector loss, generating new tail samples from head-class feature displacements). The geometric picture that carries the argument is that RSG expands the tail feature space, SCL keeps clusters tight and separated, and LDAM tilts the boundary; each component is meant to fix the distortion the others introduce.","core_discovery":"The central claim is that minimizing the weighted linear combination $\\mathcal{L}_{\\text{total}} = \\alpha \\mathcal{L}_{\\text{SCL}} + \\lambda \\mathcal{L}_{\\text{LDAM}} + \\eta \\mathcal{L}_{\\text{CESC}} + \\mu \\mathcal{L}_{\\text{MV}}$ produces a synergistic and compensatory relationship among the three techniques. SCL encourages intra-class clustering and inter-class separation but favors head classes; RSG generates new tail features and pulls features toward class centers, expanding the tail feature space that SCL squeezes; LDAM assigns larger margins to classes with fewer samples, and those margins become effective precisely because the SCL/RSG clusters leave clear voids between classes. The paper argues that SCL compensates for the head-class accuracy that RSG and LDAM would otherwise sacrifice, and ablation results on mini-ImageNet-LT show that removing any component lowers accuracy, with the full model reaching 95.61 overall and 95.42 on few-shot classes. The paper also introduces intra-class distance as a diagnostic, reporting that RSG tightens the clustering of tail classes without appreciably loosening head classes.","pith_inferences":["A testable consequence the paper leaves implicit: if the synergy is geometric, the gain from adding RSG should scale with how well SCL already separates classes; measuring per-class intra-class distance and correlating it with RSG's per-class accuracy change would check this.","The reported insensitivity to the RSG weights suggests the RSG module's main contribution may come from its center-clustering term rather than from generating new samples; an ablation that removes only the vector-transformation module would separate the two mechanisms.","The 'balanced improvement' claim could be formalized as a Pareto criterion (no tail gain without head loss), and future weighting schemes could optimize that objective directly instead of relying on genetic-algorithm searches.","Because the paper used a fixed set of datasets, an extension to more extreme imbalance factors or to class-imbalanced non-image data would test whether the compensation story holds outside the current regime."],"forward_implications":["If the synergy claim holds, long-tail recognition can be improved by a simple weighted sum of off-the-shelf losses, without two-stage decoupled training or ensembles.","Tail-class accuracy can rise without a Many-shot penalty: on ImageNet-LT the combined model improves Few-shot accuracy to 23.2 percent while Many-shot accuracy stays at 66.6 percent, near the Softmax baseline.","The intra-class distance result suggests that feature-space tightness is a usable diagnostic: models that cluster tail classes well are candidates for adding margin-based losses.","The comparison with BCL indicates that similar balanced accuracy can be approached with roughly one-third the FLOPs, since BCL's data-augmentation strategy triples its computational cost.","The paper's observation that excessive data augmentation hurts accuracy identifies a practical constraint for SCL-based long-tail models: augmentation that disperses within-class features is counterproductive."],"supporting_citations":[{"why":"Supplies the LDAM loss with per-class margins $\\Delta_j = C/n_j^{1/4}$, the classifier-level component of the combined loss.","marker":"[3]"},{"why":"Supplies the supervised contrastive loss that clusters same-class features and drives the feature-learning branch.","marker":"[17]"},{"why":"Provides the logit-adjustment view of decision-boundary rebalancing that the paper uses to frame LDAM's role.","marker":"[28]"},{"why":"Supplies the Rare-Class Sample Generator module, including the CESC and MV losses that generate tail features and pull features toward class centers.","marker":"[34]"},{"why":"Provides the end-to-end Hybrid-SC baseline combining SCL with cross-entropy, which the paper modifies by replacing CE with LDAM.","marker":"[35]"},{"why":"Documents the known trade-off that LDAM and RSG improve tail accuracy at the cost of head accuracy, the problem the combination is meant to solve.","marker":"[43]"},{"why":"Provides the BCL baseline whose balanced accuracy and computational cost frame the paper's comparison and motivation.","marker":"[45]"}],"fun_headline_variants":["Synergy of three losses balances long-tail accuracy","Three techniques combine for balanced long-tail gains","Balanced tail accuracy via synergistic loss trio","Three losses synergize to lift tail without harming head"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole synergy account depends on the RSG losses actually exerting a meaningful influence at their tiny weights ($\\eta=10^{-5}$, $\\mu=10^{-6}$); the paper reports the model is insensitive to those weights, so the ablation gains credited to RSG are not demonstrably caused by the RSG mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Synergy of three losses balances long-tail accuracy","Three techniques combine for balanced long-tail gains","Balanced tail accuracy via synergistic loss trio","Three losses synergize to lift tail without harming head"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000701,"raw_usage":{"total_tokens":3225,"prompt_tokens":1070,"completion_tokens":2155,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":686,"completion_tokens_details":{"reasoning_tokens":2097}},"tokens_in":686,"tokens_out":2155,"duration_ms":13295,"temperature":1.0,"reasoning_tokens":2097,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:37:33.604910+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the gradient norm contributed by the two RSG losses at the chosen weights at several training checkpoints on mini-ImageNet-LT; if it is orders of magnitude below the SCL and LDAM gradient norms, the ablation gains attributed to RSG cannot be explained by RSG's loss signal.","supporting_citations":[{"cited_title":"Learning imbalanced datasets with label- distribution-aware margin loss","cited_arxiv_id":null,"evidence_quote":"Supplies the LDAM loss with per-class margins $\\Delta_j = C/n_j^{1/4}$, the classifier-level component of the combined loss."},{"cited_title":"Rsg: A simple but effective mod- ule for learning imbalanced datasets","cited_arxiv_id":null,"evidence_quote":"Supplies the Rare-Class Sample Generator module, including the CESC and MV losses that generate tail features and pull features toward class centers."},{"cited_title":"Contrastive learning based hybrid networks for long- tailed image classification","cited_arxiv_id":null,"evidence_quote":"Provides the end-to-end Hybrid-SC baseline combining SCL with cross-entropy, which the paper modifies by replacing CE with LDAM."},{"cited_title":"Deep long-tailed learning: A survey","cited_arxiv_id":null,"evidence_quote":"Documents the known trade-off that LDAM and RSG improve tail accuracy at the cost of head accuracy, the problem the combination is meant to solve."},{"cited_title":"Balanced contrastive learn- ing for long-tailed visual recognition","cited_arxiv_id":null,"evidence_quote":"Provides the BCL baseline whose balanced accuracy and computational cost frame the paper's comparison and motivation."}],"review_version":1}