{"id":"eec8431f-0a76-4026-95c5-c9ce1a3b8bd8","arxiv_id":"2508.01594","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"CLIMD combines curriculum learning with confidence and complementarity measures to improve multimodal diagnosis under severe class imbalance.","lead":"CLIMD is a new training strategy that teaches multimodal medical AI models to diagnose rare disease classes by ordering samples from easy to difficult and balancing signals across imaging and clinical data. The authors claim it beats existing methods on imbalanced medical datasets and that it can be added to other models without retraining them from scratch.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The curriculum measurer's confidence/complementarity ordering is untested for minority classes; if miscalibrated early, CLIMD's curriculum could suppress minority samples and reverse its claimed gains.","rationale":"The reader and I identify the same load-bearing uncertainty: the confidence/complementarity-based difficulty measurer is assumed to provide a correct easy-to-hard curriculum for imbalanced data. I would make that worry a required empirical check rather than a fatal objection. The method is plausible; the difficulty is that the central claim of SOTA gains depends on this ordering being beneficial, and the available abstract and corrupted full text do not demonstrate it. Because the full text supplied here is unreadable mojibake, I also cannot independently verify the reported tables; the random-ordering ablation would settle the mechanistic concern, while recovering the original PDF would settle the quantitative claim. My recommendation is not to change the reader's UNVERDICTED verdict: the claim remains unverified, and the key hypothesis needs a direct test before acceptance.","tokens_in":10031,"tokens_out":7166,"duration_ms":86560,"concrete_test":"Run the released code on one multimodal medical dataset with three training conditions: (a) CLIMD as proposed, (b) same framework but with per-class difficulty scores replaced by random permutations, and (c) reversed difficulty ordering. If (a) does not significantly exceed (b) on minority-class F1/AUC (e.g., more than 3 points and consistent across seeds), the curriculum measurer is not the source of the claimed benefit; additionally report per-class expected calibration error (ECE) of confidence scores at the end of each curriculum stage to confirm early miscalibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central effectiveness claim requires that the curriculum measurer's difficulty scores—intra-modal confidence and inter-modal complementarity—order samples from easy to hard in a way that helps minority classes, and that the class-distribution scheduler then introduces minority samples at the right time. The weakest point is the first half of that chain. Early in training, a model trained on imbalanced data has poorly calibrated confidence: minority-class samples tend to receive low confidence regardless of their true difficulty, and majority easy samples receive high confidence. If CLIMD therefore selects high-confidence samples first and defers low-confidence minority samples until late, it can reduce minority-class exposure during feature learning, worsening the imbalance it is designed to fix. The abstract asserts extensive SOTA gains but provides no calibration analysis, no oracle comparison for the difficulty measure, and no minority-stratified ablation of the curriculum in the supplied text (which is also a corrupted rendering, so no quantitative result is inspectable). This is a missing validation of a load-bearing mechanism, not an internal contradiction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLIMD, a curriculum learning framework for imbalanced multimodal diagnosis. It combines two indicators in a curriculum measurer (intra-modal confidence and inter-modal complementarity) with a class-distribution-guided training scheduler, aiming to progressively adapt the model to imbalanced class distributions. The abstract claims state-of-the-art performance on multiple multimodal medical datasets, as well as plug-and-play integration into other models. However, the supplied full text is heavily corrupted mojibake, so the technical definitions, equations, experimental tables, and ablations could not be inspected; only the abstract and a few fragmentary English tokens are readable.","tokens_in":10176,"tokens_out":3894,"duration_ms":45613,"significance":"If substantiated, the paper addresses an important practical problem: class imbalance in multimodal medical diagnosis. The proposed mechanism of ordering samples by multimodal difficulty, rather than relying on resampling or loss reweighting, is a plausible and interesting direction, and the plug-and-play claim would make the framework broadly useful. The manuscript also provides a public code URL, which is a strength. However, the submitted version contains no readable experimental evidence: the abstract gives no quantitative results, and the full text is corrupted. The core mechanism of confidence/complementarity-based difficulty ordering is also unvalidated for minority classes, which is the population the method is designed to help. The significance is therefore conditional on a clean version providing the missing empirical support.","major_comments":[{"comment":"The central empirical claim that CLIMD 'outperforms state-of-the-art approaches across various metrics' is asserted without any quantitative support: no dataset names, metrics, baseline identifiers, effect sizes, or variance estimates appear in the abstract. Because this is an empirical paper, the experimental evidence is load-bearing, and as submitted there is nothing to verify.","section":"Abstract"},{"comment":"The full text is rendered as mojibake, so no equation, algorithm, table, or ablation can be examined. I could not locate readable definitions of the curriculum measurer, the inter-modal complementarity score, the class-distribution scheduler, or the training objective. This prevents verification of every central technical step and makes the manuscript unreviewable in its current form.","section":"Full text (corrupted rendering)"},{"comment":"The difficulty-ordering mechanism is a correctness risk for the very minority classes the method targets. Early in training, model confidence on imbalanced data is likely miscalibrated: minority-class samples may receive low confidence regardless of true difficulty, while majority easy samples receive high confidence. If CLIMD defers low-confidence minority samples until late, it could suppress minority-class feature learning. The manuscript provides no calibration analysis, no oracle comparison for the difficulty ordering, and no minority-stratified ablation. Since the curriculum measurer is the central novelty, this missing validation is load-bearing for the claimed imbalance improvements.","section":"Curriculum measurer (minority-class calibration)"},{"comment":"No readable evidence of experimental rigor is present: there are no visible error bars, number of independent runs, statistical tests, or per-class performance metrics. Even if the original tables contain such information, the submitted version does not allow me to determine whether the claimed state-of-the-art gains are statistically significant or consistent across datasets.","section":"Experimental evaluation"}],"minor_comments":[{"comment":"The phrase 'excels in handling imbalanced multimodal medical data' is vague; please specify the imbalance ratios used and the minority-class metrics reported (e.g., F1, recall, or AUC).","section":"Abstract"},{"comment":"The rendering of the manuscript is corrupted at the level of character encoding; please ensure the arXiv submission source is correctly encoded before resubmission.","section":"Full text"},{"comment":"The abbreviation 'CL' appears in the abstract without being expanded in the abstract body; consider writing 'curriculum learning' at first use.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":"The recommendation reflects the unreadable full text rather than a judgment on the scientific content. If a clean, correctly encoded version is available, I would be willing to review it; with only the abstract and corrupted text, neither the technical soundness nor the empirical claims can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a plausible, well-scoped curriculum-learning framework for imbalanced multimodal diagnosis, combining intra-modal confidence, inter-modal complementarity, and a class-distribution-guided scheduler. The abstract is coherent and the public code link is a plus. But our copy of the full text is corrupted mojibake, so every empirical claim is unverifiable from what I have. The paper could be completely sound; I just can't check.\n\nWhat is actually new: the specific integration of confidence-based and complementarity-based difficulty scoring into a curriculum for imbalanced multimodal medical data, presented as a plug-and-play framework. The individual ingredients are established, but the combination for this problem is not something I can point to as prior work. The abstract does not overreach in tone—it claims SOTA gains without numbers, which is normal for an abstract.\n\nSoft spots, in proportion: the evidence problem dominates. No tables, ablations, baselines, or dataset names are inspectable in the supplied text. The stress-test concern about early-training confidence miscalibration is plausible but speculative: if the curriculum relies on the model's own confidence, minority samples with low confidence could be deferred, and the abstract mentions no calibration analysis or minority-stratified ablation to rule that out. That is a missing validation, not a demonstrated flaw. If the full paper includes those ablations, the concern disappears. I would also want the datasets named in the abstract or introduction, since 'multiple multimodal medical datasets' is too vague to assess scope.\n\nWho this is for: researchers working on medical imaging AI with imbalanced multimodal data, and anyone interested in curriculum learning as a training strategy. A serious referee should get this paper because the problem is real, the method is plausible, and the code is promised—reviewers can actually run it.\n\nMy recommendation: send it to peer review. Put the confidence-calibration interaction on the reviewer checklist, and ask for a minority-stratified ablation of the curriculum. I won't cite it until I see the actual numbers, but it deserves a fair read.","headline":"Plausible curriculum-learning framework for imbalanced multimodal diagnosis, but the only inspectable evidence is the abstract; the supplied full text is corrupted, so the empirical claims cannot be checked.","tokens_in":10701,"tokens_out":1663,"would_cite":false,"duration_ms":19438,"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":"The paper claims that CLIMD, a curriculum-learning framework that orders multimodal samples by intra-modal confidence and inter-modal complementarity under a class-distribution-guided scheduler, outperforms state-of-the-art methods on…","keywords":["curriculum learning","multimodal fusion","class imbalance","medical diagnosis","intra-modal confidence","inter-modal complementarity","plug-and-play training","minority-class learning"],"falsifier":"Train CLIMD and an otherwise identical baseline with fixed random sample order on the same imbalanced multimodal dataset, using the same epochs and seeds; if the curriculum does not improve minority-class recall over that baseline, the claimed benefit of difficulty ordering is not present.","tokens_in":9832,"feed_emoji":"🩺","tokens_out":3249,"duration_ms":41827,"temperature":0.7,"pith_summary":"The paper is trying to establish that class imbalance in multimodal medical diagnosis is best handled not by resampling rare samples or reweighting their losses, but by controlling the order in which the model learns them. It proposes a curriculum-learning framework, CLIMD, that scores each training sample by how uncertain the individual modalities are and by how much the modalities complement each other, then feeds samples in an order that gradually exposes the model to the true imbalanced class distribution. The authors argue this avoids the overfitting and underfitting that plague resampling and loss reweighting, and that it captures cross-modal interactions that those methods ignore. They report that on multiple multimodal medical datasets, CLIMD outperforms existing approaches across several metrics and can be plugged into other models without changing their architecture.","feed_headline":"Order training by confidence and complementarity to beat class imbalance","feed_subtitle":"A curriculum scheduler for multimodal medical data lifts minority-class diagnosis without resampling.","key_machinery":"The load-bearing machinery is the multimodal curriculum measurer paired with the class distribution-guided scheduler. The measurer assigns each sample a difficulty score from two indicators: intra-modal confidence, how certain each modality branch is about the sample, and inter-modal complementarity, how much the modalities add to each other's evidence. The scheduler uses that score to present samples from easy to hard while gradually exposing the model to the imbalanced class distribution, so that rare classes receive appropriately timed emphasis.","core_discovery":"The central claim is a training-ordering rule: a model should learn imbalanced multimodal data by progressing from easy to hard samples, where hardness is determined by a combination of each modality's own confidence and the information the modalities contribute to one another. A class distribution-guided scheduler paces this progression so that minority classes are not abandoned early and not forced too aggressively. The paper reports that this ordering improves diagnostic performance on imbalanced multimodal medical datasets compared with state-of-the-art methods, and that the framework is plug-and-play, meaning existing multimodal models can adopt it without redesign.","pith_inferences":["The difficulty score could be reused as a signal for active learning or noisy-label detection, since the curriculum marks exactly the samples the model is least certain about.","Because the scheduler is separate from the loss function, CLIMD could potentially be combined with loss reweighting or focal loss rather than replacing them.","A clean isolation experiment would be CLIMD with the class-distribution scheduler removed, which would reveal whether gains come from sample ordering or from the pacing of class exposure.","One could test whether the same curriculum principle helps when a third, unreliable modality is added, since the complementarity term should make the model learn to discount it."],"forward_implications":["Existing multimodal diagnostic models can adopt CLIMD without architectural changes, so the benefit is not tied to one particular fusion design.","Minority classes receive a training regime that avoids the overfitting of naive resampling and the underfitting of simple loss reweighting.","Cross-modal interactions are exploited during sample ordering, not only during fusion, so disagreement between modalities becomes a training signal.","The same ordering principle should transfer to other imbalanced multimodal tasks whenever per-modality confidence and complementarity can be computed."],"supporting_citations":[],"fun_headline_variants":["Easy-to-hard multimodal training improves imbalanced diagnosis","Curriculum ordering by inter-modal complementarity beats class imbalance","Confidence and complementarity guide curriculum for imbalanced diagnosis","Multimodal curriculum: confidence plus complementarity beats imbalance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach rests on the assumption that a model's confidence in each modality and the agreement between modalities reliably say which samples are easier to learn; if those scores mislead early in training, the curriculum could train on the wrong samples and worsen the imbalance.","fun_headline_variants_meta":{"raw":{"variants":["Easy-to-hard multimodal training improves imbalanced diagnosis","Curriculum ordering by inter-modal complementarity beats class imbalance","Confidence and complementarity guide curriculum for imbalanced diagnosis","Multimodal curriculum: confidence plus complementarity beats imbalance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000939,"raw_usage":{"total_tokens":3986,"prompt_tokens":887,"completion_tokens":3099,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":3035}},"tokens_in":503,"tokens_out":3099,"duration_ms":24383,"temperature":1.0,"reasoning_tokens":3035,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:29:33.959821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CLIMD and an otherwise identical baseline with fixed random sample order on the same imbalanced multimodal dataset, using the same epochs and seeds; if the curriculum does not improve minority-class recall over that baseline, the claimed benefit of difficulty ordering is not present.","supporting_citations":[],"review_version":1}