{"id":"d66d2c91-1936-44d1-a585-c254a5fd75ac","arxiv_id":"2510.08269","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AdaGC adaptively applies gradient calibration with dual EMA in SPML for RS imagery to recover full labels from single-positive annotations and reports SOTA results on two benchmarks.","lead":"The paper introduces Adaptive Gradient Calibration (AdaGC), a framework for single-positive multi-label learning in remote sensing scene classification that uses dual EMA for pseudo-labels and a training-dynamics indicator to decide when to apply gradient calibration. This targets the practical problem of incomplete labels in expensive RS annotation tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Training-dynamics indicator's reliability in avoiding underfitting/overfitting to label noise remains the least-secured link for the SOTA and robustness claims.","rationale":"The reader's weakest assumption matches the load-bearing component exactly. Because the initial verdict was formed from the abstract and the indicator is presented as the key enabler of adaptivity and robustness, the concern is on target and justifies moving from UNVERDICTED to CONDITIONAL pending the proposed check. No other internal inconsistency is evident from the given material.","tokens_in":1774,"tokens_out":355,"duration_ms":26948,"concrete_test":"On both benchmark datasets, insert controlled phases of deliberate underfitting (train only 10 epochs) and overfitting to noise (train 3x longer than reported schedule while monitoring validation loss); record whether the indicator triggers GC during these phases and measure the resulting mAP drop relative to the published AdaGC numbers. If triggering occurs inappropriately or performance collapses, the indicator does not reliably protect the method.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The headline result (SOTA + robustness on two RS datasets under two noise types) requires that the training-dynamics indicator correctly decides when to apply gradient calibration. The abstract states the indicator is 'theoretically grounded' and 'ensures GC's effectiveness by preventing it from being affected by model underfitting or overfitting to label noise.' If the indicator can be misled by the high class imbalance, multi-scale scenes, or noise patterns typical in RS imagery, then the adaptive mechanism fails and the robustness claim does not follow. No explicit derivation of the indicator or controlled ablation isolating its failure modes is referenced in the abstract, leaving this assumption as the primary point of fragility.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes Adaptive Gradient Calibration (AdaGC) as a framework for single-positive multi-label learning (SPML) tailored to remote sensing (RS) image scene classification. It combines a gradient calibration (GC) mechanism with a dual exponential moving average (EMA) module for pseudo-label generation and introduces a training-dynamics-based indicator that adaptively triggers GC to avoid underfitting or overfitting to label noise. Experiments on two benchmark RS datasets under two label noise types report state-of-the-art performance and robustness across settings.","tokens_in":1915,"tokens_out":523,"duration_ms":20458,"significance":"If the adaptive indicator reliably detects effective calibration points without being misled by RS-specific factors such as class imbalance or multi-scale scenes, the work would advance practical SPML solutions for RS imagery where full annotations are costly. The dual EMA for pseudo-labels and code release are positive elements that could support reproducibility and further adoption in the domain.","major_comments":[{"comment":"§3.2 (Adaptive Trigger): The claim that the training-dynamics indicator is 'theoretically grounded' and prevents GC from being affected by underfitting or overfitting to label noise lacks an explicit derivation or proof sketch; the indicator is defined in terms of optimization trajectory quantities that the model itself produces, raising a circularity risk for the robustness claim.","section":"§3.2"},{"comment":"§4.2 and Table 2: The SOTA and cross-noise-type robustness results rest on the indicator correctly deciding when to apply GC, yet no controlled ablation isolates failure modes under RS-typical conditions (high imbalance, multi-scale scenes); without this, the headline performance does not fully follow from the presented evidence.","section":"§4.2"}],"minor_comments":[{"comment":"Notation for the dual EMA update rules in §3.1 could be clarified with explicit equations for both the label and feature EMAs to avoid ambiguity in implementation.","section":"§3.1"},{"comment":"The abstract mentions 'two distinct label noise types' but the experimental section would benefit from a brief table summarizing the exact noise generation procedures for reproducibility.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's scope aligns well with a CV/RS journal; however, the low soundness rating in the reader's assessment stems primarily from missing details on the indicator's derivation, which should be requested in revision."},"author_rebuttal":{"model":"grok-4.3","summary":"We sincerely thank the referee for the constructive and detailed feedback on our manuscript. The comments highlight important aspects of the theoretical motivation and experimental validation that we will address in the revision. Below we respond point by point to the major comments.","responses":[{"response":"We appreciate the referee's observation on the presentation of the adaptive trigger. The indicator is motivated by monitoring the divergence between the model's evolving predictions on the fixed single-positive labels and the dual-EMA pseudo-labels, which empirically signals the transition out of underfitting before noise overfitting dominates. We acknowledge that the original submission did not include an explicit derivation or proof sketch supporting this choice. In the revised manuscript we will add a dedicated paragraph in §3.2 that provides a step-by-step motivation derived from the expected behavior of gradient descent under partial label noise, together with a short proof sketch showing that the chosen threshold corresponds to a point where the expected gradient bias begins to increase. Regarding potential circularity, the trigger quantities are computed solely from the observed single-positive supervision and the EMA estimates; the gradient-calibration step is applied only after the trigger decision and does not feed back into the indicator. We believe these additions will remove any ambiguity while preserving the original design.","revision_made":"yes","referee_comment":"[§3.2] §3.2 (Adaptive Trigger): The claim that the training-dynamics indicator is 'theoretically grounded' and prevents GC from being affected by underfitting or overfitting to label noise lacks an explicit derivation or proof sketch; the indicator is defined in terms of optimization trajectory quantities that the model itself produces, raising a circularity risk for the robustness claim."},{"response":"We thank the referee for underscoring the need for more targeted validation of the indicator under remote-sensing-specific conditions. The reported experiments already cover two standard RS benchmarks that exhibit natural class imbalance and multi-scale scene content, and AdaGC maintains SOTA performance under both symmetric and asymmetric noise. Nevertheless, we agree that controlled ablations that explicitly vary imbalance ratios and scene-scale complexity would strengthen the robustness claim. In the revised version we will insert a new subsection in §4.2 containing two additional ablation tables: one that sweeps class-imbalance ratios while keeping other factors fixed, and another that partitions the test sets according to a multi-scale complexity metric. These will report both the trigger decision accuracy and the final mAP to demonstrate that the indicator remains reliable under the conditions highlighted by the referee.","revision_made":"yes","referee_comment":"[§4.2] §4.2 and Table 2: The SOTA and cross-noise-type robustness results rest on the indicator correctly deciding when to apply GC, yet no controlled ablation isolates failure modes under RS-typical conditions (high imbalance, multi-scale scenes); without this, the headline performance does not fully follow from the presented evidence."}],"tokens_in":1416,"tokens_out":608,"duration_ms":34304,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper takes gradient calibration from prior CV work on single-positive multi-label learning and adds a dual EMA module plus an adaptive trigger based on training dynamics, then applies the whole thing to remote sensing scene classification. They test it on two benchmark RS datasets under two label noise types and report state-of-the-art numbers plus robustness across settings. Code release is planned, which is helpful.","headline":"AdaGC adds a training-dynamics trigger to gradient calibration plus dual EMA for single-positive multi-label learning on remote sensing images, with reported SOTA results on two datasets, but the indicator's reliability under RS-specific imbalance and noise is the least secured part of the robustness claim.","tokens_in":2428,"tokens_out":183,"would_cite":false,"duration_ms":24004,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Cost/FunctionalEquation.lean","rs_theorem":"washburn_uniqueness_aczel","paper_passage":"We introduce a theoretically grounded, training-dynamics-based indicator to adaptively trigger GC, which ensures GC's effectiveness by preventing it from being affected by model underfitting or overfitting to label noise."},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean","rs_theorem":"J_uniquely_calibrated_via_higher_derivative","paper_passage":"The final objective is a combination of the binary cross-entropy loss and the GC regularization: L(θ) = L_AN(θ) + λ · R_GC_MLC(θ)"}],"headline":"AdaGC training-dynamics indicator and dual-EMA pseudo-labeling for SPML noise in RS imagery share no structural overlap with RS cost or forcing machinery","alignment":"orthogonal","rationale":"The paper's core contributions (gradient calibration via R_GC_MLC, stationarity theorem on noisy mAP, dual-EMA combination t_i = γ p_T + (1-γ) p̃_S, early-learning trigger on teacher mAP plateau) are standard empirical ML techniques for label noise. No reference to J-cost, reciprocal symmetry, φ-ladder, 8-tick periodicity, or any RS theorem appears; the work operates entirely within conventional BCE + EMA + Mixup optimization on remote-sensing datasets.","tokens_in":62784,"confidence":"high","tokens_out":350,"duration_ms":11300,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Adaptive gradient calibration with dual EMA and training-dynamics triggers recovers full labels from single-positive annotations in remote sensing scenes.","keywords":["single-positive multi-label learning","remote sensing scene classification","gradient calibration","pseudo-label generation","label noise robustness","exponential moving average","adaptive training"],"falsifier":"On the same two remote sensing benchmarks and the same two label-noise protocols, a re-implementation of AdaGC that removes the training-dynamics trigger (or replaces it with a fixed schedule) yields lower mean average precision than the full method or than prior SPML baselines.","tokens_in":2661,"feed_emoji":"🛰️","tokens_out":708,"duration_ms":22033,"temperature":0.7,"pith_summary":"The paper introduces AdaGC as a framework for single-positive multi-label learning in remote sensing imagery, where each training image carries only one positive label yet the model must infer the complete set. It combines a dual exponential moving average module to generate stable pseudo-labels with a training-dynamics indicator that decides when gradient calibration should be applied. This adaptive trigger prevents the calibration step from being applied too early, when the model is underfit, or too late, when it has begun to overfit the noisy single-positive supervision. Experiments on two standard remote sensing benchmarks under two different label-noise regimes show that the resulting method reaches state-of-the-art accuracy while remaining stable across varied settings.","feed_headline":"Adaptive calibration recovers full labels from single-positive remote sensing annotations","feed_subtitle":"Dual EMA pseudo-labels and a training-dynamics trigger let the model decide when gradient correction helps, yielding top accuracy on two RSI","key_machinery":"Adaptive Gradient Calibration (AdaGC) driven by a training-dynamics indicator that decides when to apply gradient updates based on pseudo-labels produced by a dual exponential moving average module.","core_discovery":"AdaGC adopts a gradient calibration mechanism together with a dual EMA module for robust pseudo-label generation and introduces a theoretically grounded, training-dynamics-based indicator that adaptively triggers calibration only when it is likely to be effective, thereby avoiding degradation from underfitting or overfitting to label noise; extensive experiments on two benchmark remote sensing datasets under two distinct label noise types establish that this approach attains state-of-the-art performance while preserving strong robustness.","pith_inferences":["Similar adaptive triggering could reduce the annotation burden in other image domains that rely on multi-label ground truth.","The training-dynamics signal might be combined with other semi-supervised regularizers to further stabilize learning from partial labels.","If the indicator proves reliable, it could be used to schedule other forms of label correction beyond gradient calibration."],"forward_implications":["Full multi-label recovery becomes feasible from far cheaper single-positive annotations in remote sensing scene classification.","Gradient calibration steps remain beneficial across both uniform and instance-dependent label noise without manual retuning.","The dual EMA pseudo-label generator supplies sufficiently stable targets for the calibration step on typical remote sensing imagery.","The overall pipeline generalizes across the two standard benchmark datasets without dataset-specific hyper-parameter changes."],"fun_headline_variants":["AdaGC uses dual EMA for robust pseudo-labels in single-positive RS multi-label learning","Adaptive trigger calibrates gradients for single-positive RS multi-label learning","Dynamics indicator adaptively triggers calibration to avoid label noise in RS SPML","Training dynamics indicator selects when to apply gradient calibration in RS imagery"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The training-dynamics indicator can correctly identify moments when gradient calibration will help rather than harm, without being misled by the model's early underfitting or later overfitting to the single-positive noise.","fun_headline_variants_meta":{"raw":{"variants":["AdaGC uses dual EMA for robust pseudo-labels in single-positive RS multi-label learning","Adaptive trigger calibrates gradients for single-positive RS multi-label learning","Dynamics indicator adaptively triggers calibration to avoid label noise in RS SPML","Training dynamics indicator selects when to apply gradient calibration in RS imagery"]},"model":"grok-4.3","cost_usd":0.015465,"raw_usage":{"total_tokens":6574,"prompt_tokens":722,"num_sources_used":0,"completion_tokens":76,"cost_in_usd_ticks":154653000,"prompt_tokens_details":{"text_tokens":722,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":5776,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":722,"tokens_out":76,"duration_ms":42348,"temperature":1.0,"reasoning_tokens":5776,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-18T08:51:21.442400+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On the same two remote sensing benchmarks and the same two label-noise protocols, a re-implementation of AdaGC that removes the training-dynamics trigger (or replaces it with a fixed schedule) yields lower mean average precision than the full method or than prior SPML baselines.","supporting_citations":[],"review_version":1}