{"id":"e43e8280-488e-473a-9040-6ff49a1d6e0f","arxiv_id":"2607.25830","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LDAL dynamically reweights classes using online entropy and feature-scale estimates plus an inter-epoch regularizer, and reports top-1 accuracy gains over static reweighting losses on long-tailed image benchmarks.","lead":"A new loss function for long-tailed image classification, LDAL, adjusts how much each class is penalized based on how well the model is actually learning that class rather than on fixed class counts. The authors report consistent top-1 accuracy gains over prior reweighting losses on CIFAR-LT, ImageNet-LT, and iNaturalist-2018 benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 1's stop-gradient denominator is never derived or ablated; it may be the source of reported gains.","rationale":"The reader's weakest assumption is precisely that the stop-gradient denominator in Eq. (1) is unexamined and could distort relative penalties. My analysis confirms this is the most load-bearing concern because the central claim attributes the method's success to dynamic weighting signals, yet the denominator—a non-ablated, poorly understood component—could themselves produce the observed gains. The reader already conditioned acceptance on addressing this and the evaluation issues. Since my concern matches, the verdict remains CONDITIONAL. A concrete ablation of the denominator would provide the decisive evidence needed to either validate its role or reveal it as an artifact.","tokens_in":12205,"tokens_out":10392,"duration_ms":102075,"concrete_test":"Run an ablation that replaces D_i in Eq. (1) with a constant (e.g., set D_i = 1 for all i) and retrain on CIFAR-100-LT (ρ=100) and ImageNet-LT under the same protocol, using at least 5 seeds. Compare the mean and 95% CI of top-1 accuracy to the full LDAL and to AREA. Also log D_i across training epochs and classes to test whether it varies systematically with class frequency or logit magnitude. If removing the denominator changes accuracy by less than the standard error, the concern is resolved; if it drops by >1 point, the denominator is a critical component that requires theoretical justification and independent validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LDAL's dynamic weighting (semantic scale, entropy, inter-epoch regularizer) drives its state-of-the-art results. However, the loss in Eq. (1) includes a per-class denominator D_i = (1/(BC)) Σ_{b,k} (sg(z_k^(b)) - (e_i)_k)^2, introduced in Section II-B with only the assertion that it 'keeps the penalty scale consistent.' This term is never derived, analyzed, or ablated. Because it is a function of the current logits (with stop-gradient), it acts as a dynamic, class-dependent scaling factor on the squared penalty (γ_i \\tilde{N}_i + r_i)^2. The denominator is roughly proportional to the mean squared logit magnitude minus twice the mean logit for class i, so it can vary systematically across classes: classes with higher average logits (likely head classes) get smaller D_i and thus larger penalties, while tail classes with lower logits get larger D_i and smaller penalties. This could either be the intended 'dynamics' or an unintended distortion. If this denominator—not the claimed semantic-scale and entropy signals—is what drives the reported gains, the explanation in the paper is misleading, and the method's generality is unclear. The lack of any ablation or sensitivity analysis for the denominator is a critical gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LDAL, an auxiliary loss for long-tailed classification. In addition to cross-entropy, LDAL adds a per-class penalty (γ_i \\tilde N_i + r_i)^2 divided by a stop-gradient denominator (Eq. 1). γ_i combines a logit-norm 'semantic scale' S_i and prediction entropy H_i (Eqs. 2–4), while r_i is an inter-epoch regularizer based on hard prediction-count changes for minority classes (Eq. 5). Experiments on CIFAR-10/100-LT, ImageNet-LT, and iNaturalist-2018 report accuracy gains over AREA and other CSL baselines, together with ablations and gradient-flow diagnostics.","tokens_in":12539,"tokens_out":5967,"duration_ms":58648,"significance":"If the reported gains hold under a properly controlled evaluation, LDAL would be a valuable plug-and-play objective for long-tailed classification: it operates purely at the loss level, ships with code, and includes component ablations and stability analyses. However, the central claim currently rests on comparisons that are difficult to interpret, and the most novel term in the loss—the stop-gradient denominator—is neither derived nor ablated. The paper's explanatory narrative could therefore be misleading if that denominator, rather than the semantic-scale/entropy/regularizer signals, drives the improvements.","major_comments":[{"comment":"The denominator D_i = (1/(BC)) Σ_{b,k}(sg(z_k^b)-(e_i)_k)^2 is introduced only with the assertion that it 'keeps the penalty scale consistent.' It is not a constant normalizer: D_i is a per-class, batch-dependent function of all current logits. Expanding for class i gives D_i = (1/B) Σ_b (Σ_{k≠i} z_{k,b}^2 + (z_{i,b}-1)^2), so classes with large logit magnitudes receive smaller penalties, and the effect changes over training. This term is never derived, ablated, or given a sensitivity analysis. If the reported gains are driven by this denominator rather than by S_i, H_i, or r_i, the interpretation in the paper is misleading. Please analyze the denominator's behavior, ablate it by replacing D_i with a constant or a per-class logit-norm normalizer, and report the resulting accuracy.","section":"Section II-B, Eq. (1)"},{"comment":"The hyperparameter α is selected by sweeping on the same CIFAR benchmarks used for the main results: α=1.0 is chosen from Table VIII and then used in Table I. No held-out validation split or cross-validation procedure is described. This amounts to test-set selection for a free parameter and can inflate the reported gains. Please either tune α on a separate validation split and report those numbers, or show that the Table I conclusions are unchanged when α is fixed to a value chosen without access to the test set.","section":"Section V-C / Table VIII"},{"comment":"LDAL results are reported as the 'best top-1 accuracy across all runs,' whereas the baselines are point estimates taken from prior papers. Table V shows that mean LDAL accuracy is lower than the headline numbers: CIFAR-100 ρ=100 mean 48.88 vs. best 49.79; ImageNet-LT mean 49.67 vs. best 50.10; iNaturalist mean 66.53 vs. best 67.10. Comparing best-of-runs against literature point estimates biases the comparison in LDAL's favor. Please report mean±std for LDAL and, where possible, for baselines, and state whether the differences are statistically significant.","section":"Tables I and III (footnotes)"},{"comment":"The training protocol for baselines is not fully specified: the paper states that implementation details were 'adopted from AREA [4],' but does not list epochs, learning-rate schedules, augmentations, batch sizes, or the exact values of τ, α, and the target minority set U used per dataset. This makes it hard to verify that the comparison is fair and to reproduce the results. Please provide a protocol table and release configuration files for all methods.","section":"Section IV-A / IV-B"}],"minor_comments":[{"comment":"Typos: 'eliminting' (Section IV-B), 'throughtout' (Section V-A), 'Remooving' (Section V-B). Section IV-B refers to 'Section 3.3' instead of the numbered section for γ_i.","section":"Various"},{"comment":"The upper bound τ=5.0 and the stability constant ε=1.0 are fixed without sensitivity analysis or justification. The target minority class set U is also not defined in a reproducible way.","section":"Eq. (4) / Section II-D"},{"comment":"The 95% confidence intervals appear to use z=1.96 rather than a t-distribution for n=5. The intervals would be wider with the correct t-quantile (2.776).","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially useful contribution and provides code and ablations, but the evaluation protocol and the unexplained denominator are significant enough that I cannot recommend acceptance without revision. If the authors address the major comments, the paper may become publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper with the reader's concerns in mind. The core idea — replace static class-frequency weights with a loss penalty driven by feature-strength (logit norm), prediction entropy, and an inter-epoch prediction-count regularizer — is a legitimate new combination. The ablations in Table VII show each component matters, and the gradient-flow analysis (Table VI) plus the repeated-run statistics (Table V) are honest attempts to back the method's stability. The entropy–accuracy correlation (Spearman -0.94) is a nice sanity check that supports the difficulty measure.\n\nThe soft spots are real, but I'd call them evaluation rigor rather than fatal flaws. First, α is swept on the same test sets (Table VIII) and then fixed at 1.0; that is test-set tuning for the headline numbers. The paper even admits in Section V-C that the optimal α likely depends on C, the number of classes, which undercuts the 'transfers without dataset-specific tuning' claim. Second, LDAL is reported as best-of-runs, while baseline numbers are single point estimates from prior papers; Table V shows the mean is 0.3–1.0 lower than the best, which could change comparisons on ImageNet-LT (50.10 best vs 49.53 AREA). Third, the denominator D_i in Eq. 1 is never derived or ablated. The stress-test note is on target: D_i scales each class penalty by a function of current logits, and because head classes tend to have larger logit magnitudes, this could systematically shift relative penalties. It might be a useful part of the 'dynamics' or it might be the main driver; the paper doesn't tell us. A proper ablation that replaces D_i with a constant or with a per-class norm would settle it.\n\nThe method itself is not obviously circular; the stop-gradient ops prevent direct manipulation of the statistics. The citation pattern looks fair, and the code is promised. This is a serious paper for the long-tailed reweighting community. It deserves a careful referee, not a desk reject. I'd recommend conditional acceptance — but the authors should be required to add a held-out α selection or justify why test-set tuning is acceptable, report mean±std for all methods or at least match baselines' protocols, and analyze/ablate D_i. If the gains survive that, the method is a useful plug-and-play loss.","headline":"A plausible plug-and-play dynamic loss for long-tailed classification, with honest ablations and stability checks, but the empirical claims are weakened by test-set-tuned hyperparameters, best-of-runs reporting, and an unexamined normalizer in the loss.","tokens_in":13021,"tokens_out":3412,"would_cite":true,"duration_ms":33150,"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":"Reweighting classes by real-time learning progress, not static counts, boosts long-tailed accuracy.","keywords":["long-tailed classification","reweighting loss","learning dynamics","class imbalance","Shannon entropy","semantic scale","inter-epoch regularization","soft prediction counts"],"falsifier":"Train on CIFAR-100-LT with the LDAL denominator replaced by (i) a constant 1.0, (ii) a per-class stop-gradient distance to a fixed centroid, and (iii) a detached version of the original denominator. If top-1 accuracy stays within run-to-run noise across all three, the denominator is inert; if accuracy drops sharply when the denominator is constant or when the stop-gradient is removed, the denominator is the carrier of the effect.","tokens_in":12122,"feed_emoji":"⚖️","tokens_out":8099,"duration_ms":67605,"temperature":0.7,"pith_summary":"The paper is trying to establish that loss reweighting for long-tailed classification should respond to how well each class is being learned during training, not to fixed sample frequencies. It proposes LDAL, a drop-in auxiliary loss that up-weights classes whose soft prediction counts are high relative to their current difficulty, where difficulty is measured by prediction entropy and feature-strength (semantic scale). On CIFAR-10-LT, CIFAR-100-LT, ImageNet-LT, and iNaturalist-2018, LDAL reports top-1 accuracy above the prior best pure reweighting loss, e.g., 45.04 vs. 43.85 on CIFAR-100-LT (ρ=200) and 50.10 vs. 49.53 on ImageNet-LT. If correct, this shows that objective-level dynamic weighting can capture much of the benefit of module-heavy long-tailed methods while remaining a simple plug-in.","feed_headline":"Loss that reweights by learning speed beats static baselines","feed_subtitle":"Reweights by prediction entropy and feature strength, beating static reweighting losses on four long-tailed benchmarks.","key_machinery":"The load-bearing object is the LDAL auxiliary loss term in Eq. (1): for each class i, L_i = (γ_i * \\tilde{N}_i + r_i)^2 / ( (1/(BC)) Σ_{b,k}(sg(z_k^{(b)}) - (e_i)_k)^2 + ε ), where \\tilde{N}_i is a differentiable soft prediction count, γ_i = min( S_i / ( (1 + max_j S_j) * H_i ), τ ) combines semantic scale S_i (squared mean feature logit norm) with mean prediction entropy H_i, r_i is an inter-epoch regularizer based on hard argmax count changes for target minority classes, and the stop-gradient denominator is asserted to keep penalty scale consistent. This mechanism carries the argument by converting static class-frequency priors into epoch-level learning-progress signals while retaining gra","core_discovery":"On the paper's own terms, the central discovery is that a loss whose per-class penalty is a function of dynamic learning state—soft prediction counts, prediction entropy, semantic scale, and an inter-epoch prediction-shift regularizer—consistently outperforms static inverse-frequency and margin-based reweighting losses across four long-tailed image benchmarks. The penalty for class i is ((γ_i * soft-count_i + r_i)^2) divided by a stop-gradient normalizer that measures mean squared deviation between the batch logits and the one-hot vector of class i. The dynamic weight γ_i is built so that classes that are rare but already well learned (low entropy, high semantic scale) are not heavily penali","pith_inferences":["A cheap falsification test would replace the stop-gradient denominator with a per-class constant and re-run the benchmarks; if accuracy is unchanged, the denominator is inert and the gains come from the numerator terms alone.","The reported Spearman ρ=-0.94 between training-time entropy and final test accuracy suggests entropy could be repurposed as an early-stopping or sample-scheduling signal beyond its role in the loss.","The α-sweep shows instability on CIFAR-10 at α>5, hinting that the regularizer's safe range may shrink as the number of classes grows; a class-count-aware α schedule would be a natural extension the paper does not explore.","Since LDAL is purely objective-level, a head-to-head against sampling-based hybrids under identical backbones would clarify whether dynamic weighting subsumes sampling or they are complementary."],"forward_implications":["LDAL can be appended to any classification objective with negligible overhead, so dynamic reweighting becomes a drop-in change rather than a new architecture or training schedule.","If the reported accuracy holds across five seeds, LDAL sets a new reference point among pure class-sensitive losses, closing part of the gap to decoupled and ensemble methods without multi-stage training.","The inter-epoch regularizer's benefit appears concentrated after learning-rate decay, suggesting dynamic reweighting methods should pay attention to late-stage convergence, not just early rebalancing.","The gradient-flow analysis shows the auxiliary term's relative contribution grows as training matures, implying the method's mechanism is strongest during fine-grained feature learning.","The paper's proposed extension to object detection and transformer backbones would test whether the same learning-dynamics signals transfer outside the evaluated CNN classification setting."],"fun_headline_variants":["Loss that reweights by learning speed beats static baselines","Adaptive loss weights based on learning progress outperform static reweighting","Use prediction entropy and feature strength to set class weights dynamically","Dynamic penalty per class based on learning state beats fixed reweighting"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The stop-gradient denominator in Eq. (1) is assumed to be a harmless scale-consistent normalizer, but the paper never derives or ablates it; if it couples class i's penalty to the logits of all samples in ways that skew relative weights, the reported gains could be an artifact of the normalizer rather than of dynamic weighting.","fun_headline_variants_meta":{"raw":{"variants":["Loss that reweights by learning speed beats static baselines","Adaptive loss weights based on learning progress outperform static reweighting","Use prediction entropy and feature strength to set class weights dynamically","Dynamic penalty per class based on learning state beats fixed reweighting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000542,"raw_usage":{"total_tokens":2449,"prompt_tokens":775,"completion_tokens":1674,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":1603}},"tokens_in":519,"tokens_out":1674,"duration_ms":13990,"temperature":1.0,"reasoning_tokens":1603,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T01:20:55.674532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train on CIFAR-100-LT with the LDAL denominator replaced by (i) a constant 1.0, (ii) a per-class stop-gradient distance to a fixed centroid, and (iii) a detached version of the original denominator. If top-1 accuracy stays within run-to-run noise across all three, the denominator is inert; if accuracy drops sharply when the denominator is constant or when the stop-gradient is removed, the denominator is the carrier of the effect.","supporting_citations":[],"review_version":1}