{"id":"cc72c969-7f11-4868-ade7-76e1f021a214","arxiv_id":"2608.04147","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LiNC learns per-sample trust parameters during standard training, separates clean, ambiguous, and noisy labels with a 3-component GMM, and corrects only the noisy ones, improving last-epoch accuracy under symmetric label noise.","lead":"This paper introduces LiNC, a training procedure that learns a trust value for each image-label pair, then uses a Gaussian mixture model to identify and correct likely wrong labels during training. It reports accuracy gains on ten medical image datasets with synthetic label noise, plus a strong mislabel-detection score.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The separability claim in Theorem 3.1 requires clean samples to satisfy NLL < H, but the α update's strong weight decay (λ=0.1 per step) makes the real fixed point NLL ≤ H−0.1α; hard-but-clean samples can be pushed into the noisy component and relabeled, consistent with the PathMNIST 10%…","rationale":"The reader identified the early-learning assumption and the fixed 3-component GMM as the weakest assumptions. I agree that the early-learning assumption is the central load-bearing premise, but I found a more specific and testable mechanism: the α update's strong weight decay changes the effective condition for a clean sample to retain high trust from NLL < H to NLL < H − λ_α α. This makes the separability claim more fragile than the paper's sign-only analysis suggests, and it offers a concrete explanation for the one failed setting (PathMNIST 10%). The paper's own Section 5 flags the rare-class risk but does not quantify it or tie it to the reported degradation. The GMM over-splitting concern is real but secondary; the primary issue is which samples land in the noisy component. I therefore recommend keeping the verdict CONDITIONAL, contingent on a diagnostic check of noisy-component precision and a correction-ablation study. The paper's honest acknowledgment of limitations and the generally positive results across 49/50 conditions prevent a stronger verdict change.","tokens_in":9855,"tokens_out":17916,"duration_ms":154852,"concrete_test":"On PathMNIST at 10% symmetric noise, after the 5-epoch warmup, record for every training sample the value of −log p_i,y_i − H(p_i) + λ_α α_i and the GMM component assignment. Then (a) compute the precision of the noisy component (fraction of members whose observed label is actually clean), and (b) rerun LiNC with the hard-correction step disabled, keeping only soft-target training. If the noisy-component precision is below ~80%, or if disabling hard correction restores last-epoch accuracy to the standard-training level (0.9536), the trust-separation assumption is violated in exactly the paper's reported failure setting.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Eq. (5) gives ∂L_i/∂α_i = −log p_i,y_i − H(p_i), so α increases only when −log p_i,y_i < H(p_i). However, Algorithm 1 line 19 updates α_i ← α_i − η_α(∂L/∂α_i + λ_α α_i) with η_α=1 and λ_α=0.1 per optimizer step. The fixed point is NLL − H = −0.1 α_i: with α near 1, a clean sample must sit at least 0.1 nats below the entropy just to avoid decay, and samples with ∇L ≈ 0 lose 10% of their trust per step. Over a 5-epoch warmup this drives any sample whose observed-label probability is near the geometric mean of p_i toward the low-trust component, even if the label is correct. The paper only analyzes the sign of the gradient, not this equilibrium, and does not report sensitivity to λ_α. The result is that hard clean examples, rare-class samples, or classes with systematic spurious correlations are candidates for the 'noisy' GMM component and are hard-corrected to the model's current argmax. Section 5 acknowledges the rare-class risk but neither measures it nor links it to the one reported failure: PathMNIST at 10% noise, where LiNC's last-epoch accuracy (0.9154) is 3.8 points below standard training (0.9536).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LiNC, a method that augments standard supervised training with a per-sample trainable trust parameter α_i. The training target q_i is a convex combination of the observed one-hot label and the model's own predictive distribution, q_i = (1−α_i) stopgrad(p_i) + α_i e_{y_i}. The authors derive the gradient of the loss with respect to α_i (Eq. 5) and argue from the early-learning phenomenon that this gradient pushes α_i upward for clean samples (where the negative log-likelihood of the observed label is below the prediction entropy) and downward for mislabeled samples (where it is above). After a five-epoch warmup, a 3-component Gaussian Mixture Model is fit to the trust values, and samples in the lowest-mean component are corrected, first via soft targets from the model and later via hard argmax labels. Experiments on ten MedMNISTv2 datasets with symmetric label noise up to 50% report mean last-epoch accuracy improvements of 21.41 percentage points at 50% noise relative to standard training, AUC 0.9837 for mislabel detection on OrganSMNIST, and O(N) additional memory with negligible asymptotic runtime overhead.","tokens_in":10193,"tokens_out":10877,"duration_ms":93737,"significance":"If the empirical claims hold, LiNC is attractive because it is conceptually simple, requires no second network, no clean validation set, and no manually tuned threshold, and it produces a per-sample trust score that could be used for dataset auditing. The gradient identity in Theorem 3.1 is correctly derived, and Figure 2 gives visual evidence that the trust values separate under symmetric noise on OrganSMNIST. The method's asymptotic overhead is genuinely small, and the authors are explicit about several limitations, including the rare-class risk and the fixed K=3 GMM. However, the strength of the conclusions is limited by the experimental protocol (single runs, no error bars), the single-dataset detection comparison, and a theoretical analysis that omits the effect of the trust weight decay. These issues are load-bearing for the headline claims, so the paper currently falls short of the standard for acceptance, but the core idea is promising and worth revising.","major_comments":[{"comment":"The sign analysis of the trust gradient considers only ∂L/∂α_i = NLL − H, but the actual trust update in Algorithm 1 line 19 includes weight decay: α_i ← α_i − η_α(NLL − H + λ_α α_i). With η_α=1 and λ_α=0.1, the fixed point is NLL − H = −0.1 α_i, not NLL − H = 0. Consequently, a clean sample with NLL − H slightly negative (but larger than −0.1 α_i) will have its trust decay at every optimizer step, even though the paper's clean-label condition NLL < H is satisfied. Over the five-epoch warmup, this can push hard-but-clean samples into the low-trust GMM component and cause them to be hard-corrected to the model's current argmax, which may be wrong. This provides a plausible mechanism for the PathMNIST 10% noise result in Table 3, where LiNC's last-epoch accuracy (0.9154) is 3.8 points below standard training (0.9536). The manuscript should analyze the equilibrium including λ_α, report the sensitivity of the results to λ_α, or modify the trust update so that the clean/noise boundary is not shifted.","section":"§3.3, Eq. (5); Algorithm 1 line 19"},{"comment":"All numerical results are reported as single runs with no repeated seeds, standard deviations, or confidence intervals. The headline improvement of 21.41 percentage points in mean last-epoch accuracy at 50% noise is a point estimate from one training run per configuration; on the smaller MedMNISTv2 datasets, such as BreastMNIST and PneumoniaMNIST, single-seed accuracy differences can easily exceed several points. The conclusions 'LiNC achieves higher last-epoch accuracy in 49 of the 50 dataset-noise combinations' and 'consistent gains' are therefore not supported by the evidence as reported. Please provide means and standard deviations over at least three seeds, or justify that the experimental setting is deterministic and that the reported numbers are stable.","section":"§4.2, Table 3"},{"comment":"The mislabel detection comparison in Table 1 is restricted to a single dataset (OrganSMNIST at 20% noise) and provides no experimental details for the seven baselines. The paper does not state how AUM, DataMaps, EL2N, Forgetting, CNLCU-S, or V oG were implemented, which underlying scores they used, how hyperparameters were selected, or whether the reported AUC values are averaged over runs. Without this information, the claimed improvement (LiNC 0.9837 vs V oG 0.9126) is not reproducible. Please add implementation details for all baselines, report results on more than one dataset, and include error bars or at least multiple runs to substantiate the superiority claim.","section":"§4.3, Table 1"},{"comment":"The hard-correction stage replaces the label of every sample in the lowest-trust GMM component with the model's current argmax, regardless of the model's confidence on that sample. If a non-negligible fraction of that component is actually clean—as the weight-decay equilibrium discussed in the first major comment suggests—then the method is actively replacing correct labels with incorrect ones. The paper should quantify this risk by reporting the precision and recall of the GMM 'noisy' component relative to the true injected noise, as well as the accuracy of the new hard labels compared to the original labels. This analysis is necessary to demonstrate that the correction step is net beneficial, and it would also clarify the PathMNIST 10% failure case that the authors themselves flag in Section 5.","section":"§3.5, Algorithm 1 lines 10–14; §5"}],"minor_comments":[{"comment":"The theorem statement says that 'if the model agrees with the observed label, then ∂L/∂α_i < 0 and if the model disagrees ... ∂L/∂α_i > 0.' This is not a mathematical consequence of the gradient formula alone; it follows only under the empirical early-training assumption that NLL < H for clean samples and NLL > H for mislabeled samples. Please rephrase the theorem to state the gradient identity and the sign condition separately, and make explicit that the clean/noisy mapping rests on the early-learning hypothesis.","section":"§3.3, Theorem 3.1"},{"comment":"Equations (4) and (5) are redundant derivations of the same identity; keeping only one of them would reduce clutter without changing the content.","section":"§3.3, Eqs. (4) and (5)"},{"comment":"In Figure 2, overlaying the GMM decision boundaries and reporting the fraction of true clean samples assigned to the noisy component would make the visual claim of separation more quantitative and easier to check.","section":"Figure 2"},{"comment":"Table 2 reproduces baseline numbers from reference [26] and not from experiments in this paper. Please state this explicitly in the caption and text so readers do not mistakenly attribute those results to the authors' own runs.","section":"§4.2, Table 2"},{"comment":"The sentence 'The optimal hyperparameters without LiNC are still optimal with LiNC' is made without any ablation or sensitivity study. Either add supporting evidence or soften the claim.","section":"§4.2"},{"comment":"After the hard-correction step, the trust parameters are frozen. The paper should briefly explain why α is no longer updated or used, since the method description earlier implies continuous trust learning.","section":"Algorithm 1 line 26"},{"comment":"Given that the fixed K=3 GMM splits into three components even when the data are clean or lightly noisy, the paper's claim of 'consistent gains' is strained by the PathMNIST 10% result. A simple abstention rule (e.g., skip hard correction when the trust distribution is unimodal) might mitigate this, and an experiment along these lines would strengthen the discussion in Section 5.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The authors are honest about their limitations, which I appreciate, and the gradient identity is a useful clarifying contribution. However, the experimental evidence is not yet at the level needed for the paper's claims: single-seed results, a one-dataset detection comparison with undocumented baselines, and a trust-update analysis that ignores the substantial weight decay. The PathMNIST 10% failure is a concrete sign that the correction mechanism can hurt, and the weight-decay equilibrium in the trust update is a likely contributor. I would be willing to review a revised manuscript that adds repeated runs, documents the detection baselines, and addresses the weight-decay analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear —,\n\nShort version: this paper earns a serious look. The idea is lightweight in the best sense—one trainable scalar per sample, an interpolated target, a 3-GMM, and a staged correction schedule. They derive the trust gradient correctly and show empirically on ten MedMNISTv2 datasets that last-epoch accuracy improves consistently, with a big mean gain at 50% noise. The related work is honest and the complexity analysis is clear.\n\nWhat I’d push back on, in order of importance. First, no error bars, no repeated seeds, no code. That’s a real problem for a paper whose headline claim is an average improvement. Second, their Algorithm 1 adds weight decay to the alpha update but the analysis in Theorem 3.1 only considers the gradient sign, not the fixed point. Because λ=0.1, a clean sample's NLL must sit at least 0.1 nats below the entropy just to keep alpha from falling. So hard clean examples are candidates for the \"noisy\" component. The PathMNIST 10% result—where LiNC is 3.8 points below plain training at the last epoch—is exactly the failure you'd expect from that mechanism. The authors acknowledge the general risk in Section 5 but don’t connect it to the alpha update or report sensitivity to λ. Third, the mislabel-detection comparison is only one dataset, and the baseline setup isn't documented.\n\nI still think the method deserves a referee. The problems are fixable with more experiments and a proper analysis of the update rule. I’d ask for code, multiple seeds, standard deviations, and a sensitivity sweep on λ and the warmup length. If those come out clean, the method has a real niche.\n\nRead it if you work on noisy labels in medical imaging. I’d bring it to reading group—the fixed-point issue is a nice teaching example.\n\nRecommendation: send to peer review, not desk reject.","headline":"Useful, lightweight label-noise method with a clean trust-gradient derivation, but the alpha weight-decay term and missing error bars undermine the central separability claim; worth refining and re-reviewing.","tokens_in":10723,"tokens_out":3783,"would_cite":false,"duration_ms":33195,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A per-sample trust parameter learned during ordinary training separates clean from noisy labels, and a Gaussian mixture model then corrects only the low-trust samples.","keywords":["label noise","noisy labels","per-sample trust","Gaussian mixture model","label correction","medical imaging","MedMNIST","early learning"],"falsifier":"Train LiNC on a dataset where one class has all its labels systematically flipped to a single incorrect class; if those samples receive high trust because the model confidently agrees with the wrong label, the class-conditional correction rate and accuracy will reveal the failure. A second test is to run LiNC on a fully clean dataset and check whether the fixed three-component GMM still sends a nonzero fraction of true labels to hard correction.","tokens_in":9648,"feed_emoji":"🩺","tokens_out":7188,"duration_ms":58449,"temperature":0.7,"pith_summary":"LiNC claims that a single trainable trust parameter per training sample can separate correct from corrupted labels during the first epochs of ordinary training, without an extra model, a clean validation set, or a manually chosen threshold. The paper derives the gradient of the training objective with respect to that trust parameter and shows it pushes trust upward when the model agrees with the observed label and downward when it disagrees. A three-component Gaussian mixture model then groups trust values into noisy, ambiguous, and clean cases, and only the noisy group is relabeled, first softly and then hard. Across ten 2D medical imaging datasets with symmetric label noise up to 50%, the method raises mean last-epoch accuracy by 21.41 percentage points at the highest noise rate and provides a per-sample trust score that can also be used to audit datasets.","feed_headline":"One trust value per sample lifts noisy-label accuracy 21 points","feed_subtitle":"A single learned number separates clean from corrupted labels inside the training loop, with no extra model or clean validation set.","key_machinery":"The load-bearing object is the per-sample trust parameter $\\alpha_i\\in(0,1)$, which interpolates between the observed one-hot label and the model's own prediction through the soft target $q_i(\\alpha_i)=(1-\\alpha_i)\\,\\mathrm{stopgrad}(p_i)+\\alpha_i e_{\\tilde y_i}$. The argument turns on the trust-gradient identity $\\partial L_i/\\partial\\alpha_i = -\\log p_{i,\\tilde y_i}-H(p_i)$: in early training the model is expected to fit correct patterns first, so the sign of this gradient separates clean from noisy samples. The second mechanism is a fixed three-component Gaussian mixture model fitted to the trust values, which provides a threshold-free partition into noisy, ambiguous, and clean groups; only the noisy component is later corrected, first with soft targets and then with hard labels.","core_discovery":"The paper's central claim is that the sign of the trust gradient is determined by a comparison between the negative log-likelihood of the observed label, $-\\log p_{i,\\tilde y_i}$, and the prediction entropy $H(p_i)$: during early training, $-\\log p_{i,\\tilde y_i}<H(p_i)$ for clean samples, so $\\alpha_i$ rises, while $-\\log p_{i,\\tilde y_i}>H(p_i)$ for mislabeled samples, so $\\alpha_i$ falls. This produces a separable distribution of trust values without access to ground-truth labels. Fitting a three-component Gaussian mixture model to the trust values, LiNC assigns the lowest-mean component to noisy samples, trains those samples first with the model's soft predictions and then with the hard $\\arg\\max$ label, and leaves the other samples untouched. On ten MedMNISTv2 datasets under symmetric noise, the method reports mean last-epoch accuracy gains of 2.19, 7.86, 12.31, 16.94, and 21.41 percentage points at noise rates of 10% through 50%, and improves mislabel-detection AUC to 0.9837 on OrganSMNIST at 20% noise.","pith_inferences":["A testable extension of this mechanism is that any loss whose gradient separates agreement from disagreement could generate similar trust scores, so the approach may transfer beyond cross-entropy to other training objectives.","Because the separation depends on the early-learning assumption, systematically corrupted subgroups, such as a rare class whose instances all receive the same wrong label, could keep high trust and evade correction; the paper itself flags this risk.","The fixed three-component GMM will always find three groups even on clean data, and the paper's PathMNIST 10% result (last-epoch accuracy 0.9154 with LiNC versus 0.9536 without) is consistent with over-correction, suggesting a data-driven component count or an abstention rule would help.","The trust scores could be reused after training as a ranking for expert re-review, subgroup fairness analysis, or sample reweighting, even in settings where label correction itself is not desired."],"forward_implications":["At 50% symmetric label noise, mean last-epoch accuracy across the ten datasets rises from 0.5882 to 0.8023, a gain of 21.41 percentage points, and LiNC beats standard training in 49 of 50 dataset-noise combinations on last-epoch accuracy.","The gap between best-epoch and last-epoch accuracy shrinks from 13.14 to 2.31 percentage points on average, which means the method suppresses late-training memorization of corrupted labels.","At 20% noise on OrganSMNIST, the trust values detect mislabeled samples with AUC 0.9837, better than the strongest listed baseline at 0.9126, so the same learned scores can be used for dataset auditing.","The method adds $O(N)$ memory for trust parameters and keeps the asymptotic training-time complexity at $O(EN)$, so it can be added to a standard fine-tuning loop.","The staged correction schedule, soft warmup, soft correction, then hard correction, limits the risk of reinforcing wrong model predictions when the model itself is still unreliable."],"supporting_citations":[{"why":"Supplies the empirical early-learning observation that networks fit clean patterns before memorizing noise, which is the premise for the trust-gradient sign separation.","marker":"[2]"},{"why":"Provides the small-loss noise-separation baseline and the symmetric-noise evaluation protocol LiNC follows.","marker":"[9]"},{"why":"Motivates the Gaussian-mixture separation of examples and the choice of three components, and is the strongest noisy-label method LiNC is compared against.","marker":"[13]"},{"why":"Supplies the ten MedMNISTv2 datasets and the baseline accuracy table used to compare LiNC with standard training.","marker":"[26]"},{"why":"Provides the expectation-maximization algorithm used to fit the three-component GMM to trust values.","marker":"[4]"}],"fun_headline_variants":["Per-sample trust value separates clean from noisy labels","A single trust number lifts noisy-label accuracy by 21 points","Trust gradient drives clean-vs-noisy split without ground truth","GMM on trust values detects mislabels and corrects them","21-point gain on noisy medical images from one parameter per sample"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the empirical early-learning premise that, in the first epochs, the model assigns higher probability to correct labels than to incorrect ones; if a wrong label is systematically repeated or a correct label is hard, the trust gradient can point the wrong way and the correction will miss or damage those samples.","fun_headline_variants_meta":{"raw":{"variants":["Per-sample trust value separates clean from noisy labels","A single trust number lifts noisy-label accuracy by 21 points","Trust gradient drives clean-vs-noisy split without ground truth","GMM on trust values detects mislabels and corrects them","21-point gain on noisy medical images from one parameter per sample"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000677,"raw_usage":{"total_tokens":3124,"prompt_tokens":1038,"completion_tokens":2086,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":2003}},"tokens_in":654,"tokens_out":2086,"duration_ms":13464,"temperature":1.0,"reasoning_tokens":2003,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:41:42.385064+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LiNC on a dataset where one class has all its labels systematically flipped to a single incorrect class; if those samples receive high trust because the model confidently agrees with the wrong label, the class-conditional correction rate and accuracy will reveal the failure. A second test is to run LiNC on a fully clean dataset and check whether the fixed three-component GMM still sends a nonzero fraction of true labels to hard correction.","supporting_citations":[{"cited_title":"A closer look at memorization in deep networks","cited_arxiv_id":null,"evidence_quote":"Supplies the empirical early-learning observation that networks fit clean patterns before memorizing noise, which is the premise for the trust-gradient sign separation."}],"review_version":1}