{"id":"a78f52c3-c455-46f3-81fe-27f84f12d485","arxiv_id":"2412.11620","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Collaborative Cross Learning reduces Semantic Contamination in noisy-label training by aligning embeddings across views and models, improving accuracy on CIFAR and real-world noisy datasets.","lead":"This paper introduces Semantic Contamination, a failure mode where label-refurbishment training makes models link unrelated classes, such as cats and airplanes. It proposes Collaborative Cross Learning, a two-module method that mines semantic structure from embeddings across views and models, and reports gains on several noisy-label benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Own Table 1 contradicts the abstract's 'outperforms existing approaches' claim: CrossSplit beats the method on CIFAR-10 80% symmetric and CIFAR-100 20% symmetric, and RankMatch beats it on CIFAR-100 50% symmetric.","rationale":"I read the paper as making a broad empirical claim: Collaborative Cross Learning outperforms existing methods on synthetic and real-world noisy benchmarks. The condition most needed for that claim to hold is that the reported comparisons support it in the claimed settings. Table 1 immediately violates this condition in several cells, so the headline claim is overstated as written. The reader's weakest assumption about the gap between the individual-positive-pair proof and the same-class positive loss in Eq. (9) is a real theoretical weakness, and I agree it should be fixed; however, the strongest claim in the abstract is empirical, so the Table 1 contradiction is more load-bearing for that claim. I would keep the reader's CONDITIONAL verdict: the method still looks competitive in many settings and the issues are correctable by restating the claim and reporting variance. I see no basis for rejection, and the overclaim does not appear to be an internal inconsistency in the method itself.","tokens_in":16635,"tokens_out":6570,"duration_ms":65501,"concrete_test":"Re-run the contested settings (CIFAR-10 Sym80, CIFAR-100 Sym20, CIFAR-100 Sym50) for Ours, CrossSplit, and RankMatch under identical code, augmentation pipeline, and at least 5 random seeds, reporting mean±std; if either baseline remains at or above Ours by more than one standard error in those cells, the abstract and §5.2 must be weakened to per-benchmark comparisons.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: the method 'outperforms existing approaches' on synthetic and real-world noisy datasets. The paper's own Table 1 shows this is literally false as stated. In §5.2 the text says the method 'outperforms the state-of-the-art models across all noisy levels on CIFAR-10/100 with different types of synthetic noise,' yet Table 1 reports CrossSplit at 95.4 vs. Ours at 94.6 on CIFAR-10 with 80% symmetric noise, CrossSplit at 79.9 vs. Ours at 79.5 on CIFAR-100 with 20% symmetric noise, and RankMatch at 77.6 vs. Ours at 77.4 on CIFAR-100 with 50% symmetric noise. The method is still competitive in most cells, but the blanket superiority claim is contradicted by the paper's own evidence. This is more directly load-bearing than the theory gap flagged by the reader: even if the mutual-information proof for Eq. (9) were repaired, the empirical overstatement would remain. The absence of error bars or confidence intervals in Tables 1–4 makes the 0.1–0.4% gaps impossible to interpret; the appendix only states that results are averaged over 3 trials. A precise, per-benchmark claim with variance reporting would salvage the empirical contribution.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a failure mode it calls Semantic Contamination (SC) in label refurbishment methods for learning with noisy labels: refurbishing with view logits is said to produce class-imbalanced semantic structure, while refurbishing with model logits fails to maintain semantic consistency across models. After analyzing RoLR as a representative method, the authors propose Collaborative Cross Learning, composed of Semantic-wise Decoupling with Confident Learning (SDCL) for cross-view learning and Embedding-based Interactive Alignment (EIA) with Collaborative Contrastive Learning on Refurbished Labels (CCLRL) for cross-model learning. A mutual-information lower bound is claimed for CCLRL. Experiments cover CIFAR-10/100 with synthetic symmetric, pair, and instance-dependent noise, plus CIFAR-N, Animal-10N, and WebVision, together with ablations and sensitivity analyses. The paper claims state-of-the-art results on synthetic and real-world noisy datasets.","tokens_in":16962,"tokens_out":7836,"duration_ms":69327,"significance":"If the claimed results held with the reported breadth, the paper would be a useful contribution: the SC concept gives a plausible explanation for a known failure mode of label refurbishment, and the decomposition into prediction guidance versus semantic smoothing, and view alignment versus model alignment, is informative. The empirical study is broad, and the appendix's use of the external class-taxonomy metric LCA is a step toward non-circular validation of the semantic-contamination claim. The paper is also candid about the simplifications in its proof. However, the central empirical claim is stated more strongly than the paper's own Table 1 supports, and the proof of Eq. (11) covers a simplified setting rather than the loss actually optimized in Eq. (9). Both issues need to be repaired before the claimed results can be accepted.","major_comments":[{"comment":"The sentence in Section 5.2 stating that the method 'outperforms the state-of-the-art models across all noisy levels on CIFAR-10/100' is contradicted by Table 1: CrossSplit reports 95.4 versus Ours 94.6 on CIFAR-10 with 80% symmetric noise, CrossSplit reports 79.9 versus 79.5 on CIFAR-100 with 20% symmetric noise, and RankMatch reports 77.6 versus 77.4 on CIFAR-100 with 50% symmetric noise. The Abstract's 'outperforms existing approaches' is therefore too strong as written. Because the paper's central claim is empirical superiority, the authors should replace blanket superiority statements with per-benchmark claims and explicitly acknowledge the cells in which the method is not the best.","section":"Section 5.2, Table 1"},{"comment":"None of the tables reports standard deviations or confidence intervals, although the appendix says that results are averaged over 3 trials. Many of the margins over the strongest baselines are 0.1-0.4 percentage points (e.g., 97.0 vs. 96.9 on CIFAR-10 with 20% symmetric noise, and 96.5 vs. 96.3 on CIFAR-10 with 50% symmetric noise), which is within typical run-to-run variation for these benchmarks; conversely, the losses identified in Major Comment 1 may also be within noise. Please report per-seed results and mean +/- std for all numbers produced by the authors, and clearly distinguish numbers copied from published papers from numbers reproduced by the authors, since Table 1 mixes both sources.","section":"Tables 1-4; Appendix Implementation Details"},{"comment":"The proof leading to Eq. (11) assumes a single positive pair consisting of the same input sample under the two models, and the appendix explicitly states that it 'exclusively focus[es] on individual positive sample pairs' and ignores view differences. However, the loss in Eq. (9) that is actually minimized treats all same-class samples in the batch as positives and uses strong augmentations from one model against weak augmentations from the other. The claimed bound I(f\\theta_m(x), f\\theta_{1-m}(x)) >= log(N) - E[L_CCLRL] therefore does not apply to the deployed objective in any straightforward way. Either the proof needs to be extended to same-class positive pairs and cross-view pairs, or the mutual-information statement should be reframed as motivation rather than as a theorem for CCLRL.","section":"Appendix, Proof of maximizing the lower bound of the mutual information; Eq. (9)"},{"comment":"The main-text evidence that the method reduces Semantic Contamination (Table 6) is partly self-confirming: the reported cosine similarities come from embeddings produced by a model whose training objective is explicitly designed to shape the embedding geometry through the contrastive terms in Eqs. (5), (8), and (9), so favorable geometry is in part a direct optimization outcome rather than an independent measurement. The LCA metric in Table 8 is more independent because it uses an external class taxonomy; please move that analysis (with error bars and more samples) to the main text and present Table 6 as an illustrative example rather than as standalone validation.","section":"Table 6; Appendix Table 8"}],"minor_comments":[{"comment":"There are several typos in the text ('Aniaml-10N', 'Emebdding-based', 'Pesudo-code', 'shwon'); please proofread the manuscript carefully.","section":"Abstract and Section 5.2"},{"comment":"The subscript q in f\\theta_q is undefined; it should presumably be v, i.e., 1-m, matching the notation used elsewhere.","section":"Eq. (11)"},{"comment":"The appendix says 'For CIFAR-like datasets, we consider the first 30 epochs as warm-up', but the Implementation Details say 'warm-up epoch is 10 epochs for CIFAR-10 and 30 epochs for CIFAR-100'; please reconcile this inconsistency.","section":"Appendix Details of our method; Implementation Details"},{"comment":"The experimental setup says synthetic noise rates tau_0 in {0.2, 0.4, 0.5, 0.8}, but Table 1 reports only 20%, 50%, and 80% for symmetric and pair noise and 40% for instance-dependent noise; please clarify the exact protocol.","section":"Experimental Setup"},{"comment":"The caption says 'variance of hyperparameters' but appears to mean sensitivity to hyperparameter values, and the y-axis of the plots is not labeled.","section":"Figure 6"},{"comment":"Table 6 reports a single illustrative example without error bars or a description of how many samples or batches were used; please add this information or move the example to the appendix.","section":"Table 6"},{"comment":"No code release is mentioned; making the code available would substantiate the reproducibility of the averaged results.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely salvageable. The empirical overclaim is fixable by rewording the summary claims and reporting variance, and the theory gap is fixable by extending the proof to the actual loss or softening the theoretical claim. The reported numbers, if accurate and reproducible, position the method as competitive or state-of-the-art in most benchmark settings. I would not reject on the current evidence, but I would not accept before the issues in Major Comments 1-3 are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuinely useful noisy-label paper, but the headline claim needs to be walked back. The Semantic Contamination framing is the real contribution, not the SOTA claim.\n\nWhat's new: they identify a failure mode in label refurbishment—models learn corrupted semantic associations (e.g., cat closer to airplane than dog) even when predictions are accurate—and they show why RoLR suffers from it: view-based alignment gives class imbalance, model-based alignment gives cross-model inconsistency. The diagnostic metrics (entropy of class variances, Membed/Mlogit, and the LCA taxonomy metric in the appendix) support the story. The proposed CCLRL loss, which aligns contrastive distributions across models on refurbished labels, is a reasonable new objective, and the ablations show each component contributes. Real-world results on CIFAR-N, Animal-10N, and WebVision are competitive and often best.\n\nSoft spots, in order of severity. First, the empirical overclaim. The abstract says the method outperforms existing approaches on synthetic and real-world sets, and §5.2 says 'outperforms the state-of-the-art models across all noisy levels on CIFAR-10/100.' Table 1 shows CrossSplit beats it on CIFAR-10 80% symmetric (95.4 vs 94.6) and CIFAR-100 20% symmetric (79.9 vs 79.5), and RankMatch on CIFAR-100 50% (77.6 vs 77.4). Those are small gaps, but the blanket claim is false, and with three trials and no variance reported, even the close wins are hard to interpret. That's fixable: report per-benchmark numbers with standard deviations and soften the conclusion.\n\nSecond, the theory gap. The proof in the appendix only handles individual positive pairs—same sample seen by two models—but Eq. (9) treats all same-class samples in the batch as positives. The appendix acknowledges this simplification but never shows the mutual-information bound survives the change. Since the loss actually deployed is the same-class version, the theoretical grounding as written does not apply to the deployed objective. This is not fatal to the empirical method, but it should be stated as a proof sketch with an explicit caveat, not a full derivation.\n\nThird, Table 6, which shows their embeddings put automobile closer to truck than to cat, is partly self-confirming because CCLRL is trained to pull same-class embeddings together. The LCA metric based on an external taxonomy partially answers this; I'd make that the primary SC evidence.\n\nAll in all: the paper deserves a serious referee. The framing and analysis are worth engaging with even if the SOTA claim needs tempering. I would send it out; a good referee could push for variance reporting and a repaired or honestly-scoped theory section.","headline":"Useful diagnostic framing and solid empirical method, but the 'outperforms SOTA' claim is contradicted by its own Table 1, and the theory covers a simplified loss, not the one actually run.","tokens_in":17442,"tokens_out":2175,"would_cite":false,"duration_ms":16591,"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 label refurbishment methods fail because they align logits instead of embeddings, and that its proposed Collaborative Cross Learning method, which learns semantic associations contrastively across views and models…","keywords":["learning with noisy labels","label refurbishment","semantic contamination","contrastive learning","mutual information","semi-supervised learning","representation learning"],"falsifier":"Train the exact CCLRL loss (same-class positives) on CIFAR-10 with 80% symmetric noise, then compute the empirical mutual information $I(f_0(x), f_1(x))$ over a held-out split; if the inequality $I \\geq \\log(N) - \\mathbb{E}[L_{CCLRL}]$ fails by a large margin, the theoretical grounding does not describe the deployed objective. Alternatively, a baseline that aligns logits but adds heavy embedding regularization would test whether the benefit comes specifically from decoupling semantic learning from logit alignment.","tokens_in":16433,"feed_emoji":"🧠","tokens_out":6316,"duration_ms":49484,"temperature":0.7,"pith_summary":"This paper identifies a failure mode it calls Semantic Contamination: when learning with noisy labels, label-refurbishment methods that align model predictions (logits) can make models learn corrupted semantic relationships—e.g., placing 'cat' closer to 'airplane' than to 'dog'—which degrades representations even when predictions are accurate. The authors analyze RoLR, a representative refurbishment method, and trace the contamination to two sources: view-based alignment causes semantic imbalance among classes, and model-based alignment breaks semantic consistency across models. To avoid this, they propose Collaborative Cross Learning, which decouples the predicted class from the semantic concept and learns semantic associations through contrastive losses on embeddings, both across augmented views and across two models. The theoretical component shows that the cross-model contrastive loss lower-bounds the mutual information between the two models' embeddings. Experiments on CIFAR, CIFAR-N, Animal-10N, and WebVision report state-of-the-art accuracy while improving semantic-consistency metrics.","feed_headline":"Contrastive embeddings stop noisy labels from corrupting semantics","feed_subtitle":"Collaborative cross learning beats prior methods on synthetic and real-world noisy-label benchmarks.","key_machinery":"The load-bearing object is the contrastive embedding distribution $q^{s\\to w}$—the distribution over weak-augmentation embeddings given a strong-augmentation anchor—which carries semantic associations without invoking logits. Cross-view learning (CVL) combines a confident-learning prediction term $L_{PG}$ (cross-entropy only when the top predicted probability exceeds threshold $c$), an augmentation-wise contrastive loss $L_{ACL}$ that aligns strong and weak views of the same sample, and a view-wise mimicry KL $L_{VM}$ that makes the two directional contrastive distributions agree. Cross-model learning (CML) adds Collaborative Contrastive Learning on Refurbished Labels (CCLRL), whose positive pairs are samples sharing a refurbished label $y'$ obtained by mixing the noisy label with the peer model's prediction, and a model-wise mimicry KL $L_{MM}$. The theoretical anchor is the inequality $I(f_{\\theta_m}(x), f_{\\theta_{1-m}}(x)) \\geq \\log(N) - \\mathbb{E}[L_{CCLRL}]$, which connects the practical loss to mutual information between the two models' embeddings.","core_discovery":"The central claim is that the logits used in label refurbishment are the wrong carrier of semantic information: aligning logits across views over-concentrates confidence in some classes (semantic imbalance), and aligning logits across models does not produce consistent embedding relationships (semantic inconsistency). Collaborative Cross Learning therefore keeps the refurbished-label supervision for the predicted class—gated by a confidence threshold—but routes all semantic information through contrastive learning on embeddings. Cross-view, it uses an augmentation-wise contrastive loss and a view-wise mimicry KL term; cross-model, it uses Collaborative Contrastive Learning on Refurbished Labels (CCLRL), where positive pairs are same-class samples according to labels refurbished by the peer model, plus a model-wise mimicry term. The paper proves that minimizing CCLRL maximizes a lower bound on the mutual information between the two models' embeddings, and shows empirically that the method outperforms prior art on synthetic and real-world noise benchmarks and achieves better semantic consistency by embedding- and taxonomy-based metrics.","pith_inferences":["Semantic Contamination may be more general than the noisy-label setting: any training signal that aligns outputs without aligning latent spaces, such as distillation or ensemble agreement, could produce similar corrupted semantic ordering; the embedding-based decoupling proposed here is a candidate remedy.","The CCLRL mutual-information bound is derived for individual positive pairs; whether it holds for the same-class positive loss actually minimized is an open question that could be tested by measuring $I(f_{\\theta_m}, f_{\\theta_{1-m}})$ directly on a trained model.","The contrastive distribution mimicry (KL between $q^{s\\to w}$ and $q^{w\\to s}$) could serve as a generic regularizer for representation learning beyond noise-robust classification."],"forward_implications":["If the claim holds, label refurbishment methods should abandon direct logit alignment and instead mine semantic associations in embedding space to avoid Semantic Contamination.","The CCLR losses can be attached to existing SSL-based LwNL methods such as DivideMix and yield consistent accuracy gains, as shown in the paper's combination experiments.","The mutual-information lower bound provides a principled justification for using peer-refurbished labels in cross-model contrastive learning.","The method reports state-of-the-art accuracy under extreme symmetric noise (80% on CIFAR-100, reaching 70.3%) and on several real-world noise benchmarks, suggesting robustness to high noise rates."],"supporting_citations":[{"why":"RoLR, the representative label refurbishment method whose loss decomposition and failure modes are analyzed and used as the main comparison baseline.","marker":"(Chen et al. 2023)"},{"why":"DivideMix, the semi-supervised framework used as a baseline and as the host method in the combination experiments with the proposed losses.","marker":"(Li, Socher, and Hoi 2020)"},{"why":"Supplies the mutual-information lower-bound derivation technique used to justify CCLRL.","marker":"(Yang et al. 2022)"},{"why":"Contrastive Multiview Coding, the origin of the contrastive mutual-information bound that the proof adapts.","marker":"(Tian, Krishnan, and Isola 2020)"},{"why":"FixMatch, the consistency-and-confidence SSL approach whose strong/weak augmentation and pseudo-labeling ideas inform the confident learning component.","marker":"(Sohn et al. 2020)"},{"why":"RankMatch, a recent SSL- and rank-contrastive-based LwNL method that is a key comparison baseline.","marker":"(Zhang et al. 2023)"},{"why":"Provides the embedding-arithmetic analogy used in the M_embed metric to evaluate semantic consistency across models.","marker":"(Mikolov et al. 2013)"}],"fun_headline_variants":["Collaborative cross learning stops semantic contamination from noisy labels","Cross-model embeddings beat logit refurbishment for label noise","Contrastive learning on refurbished labels purges semantic contamination","Cooperative embeddings fix noisy-label training without semantic drift","Cross-view cross-model contrastive learning stops semantic contamination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical guarantee that CCLRL raises mutual information between the two models is proven only for the simplified case where the positive pair is the same image seen by both models, but the deployed loss treats every same-class sample in the batch as a positive pair; if that simplification does not carry over, the mutual-information bound does not apply to the actual objective.","fun_headline_variants_meta":{"raw":{"variants":["Collaborative cross learning stops semantic contamination from noisy labels","Cross-model embeddings beat logit refurbishment for label noise","Contrastive learning on refurbished labels purges semantic contamination","Cooperative embeddings fix noisy-label training without semantic drift","Cross-view cross-model contrastive learning stops semantic contamination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2413,"prompt_tokens":898,"completion_tokens":1515,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":1436}},"tokens_in":514,"tokens_out":1515,"duration_ms":10906,"temperature":1.0,"reasoning_tokens":1436,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:45:14.686579+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the exact CCLRL loss (same-class positives) on CIFAR-10 with 80% symmetric noise, then compute the empirical mutual information $I(f_0(x), f_1(x))$ over a held-out split; if the inequality $I \\geq \\log(N) - \\mathbb{E}[L_{CCLRL}]$ fails by a large margin, the theoretical grounding does not describe the deployed objective. Alternatively, a baseline that aligns logits but adds heavy embedding regularization would test whether the benefit comes specifically from decoupling semantic learning from logit alignment.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RoLR, the representative label refurbishment method whose loss decomposition and failure modes are analyzed and used as the main comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-information lower-bound derivation technique used to justify CCLRL."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contrastive Multiview Coding, the origin of the contrastive mutual-information bound that the proof adapts."},{"cited_title":"D.; Kurakin, A.; and Li, C","cited_arxiv_id":null,"evidence_quote":"FixMatch, the consistency-and-confidence SSL approach whose strong/weak augmentation and pseudo-labeling ideas inform the confident learning component."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RankMatch, a recent SSL- and rank-contrastive-based LwNL method that is a key comparison baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the embedding-arithmetic analogy used in the M_embed metric to evaluate semantic consistency across models."}],"review_version":1}