{"id":"bff972b0-8f94-47d5-a632-f865cdf27aa6","arxiv_id":"2608.09052","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A confidence-routed triple-LoRA expert framework reduces pseudo-label noise interference and improves semi-supervised accuracy when adapting frozen vision foundation models.","lead":"TriNoL splits unlabeled images into three confidence groups and trains a separate LoRA adapter for each group, so reliable, ambiguous, and noisy pseudo-labels no longer interfere in one low-rank update. The method gives small, consistent accuracy gains on image classification with very few labels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gain rests on a single low-confidence FOOD-101 N2 comparison; the routing premise itself is not independently tested.","rationale":"The paper is honest and clearly motivated, and the empirical gains are real but modest. The main problem is that the central attribution - confidence-aware expert specialization, not just capacity - rests on a single rank-matched comparison on one dataset (FOOD-101 N2), with all other ablations using a rank-8 single LoRA that is not parameter-matched. The reader's weakest_assumption (routing reliability, Assumption 2) is exactly the same concern: the whole method reduces to whether confidence is a trustworthy routing signal. Since the paper concedes in Limitations that miscalibration or domain shift would break it, and provides no calibration analysis or empirical measure of the routing error bound, the concern is load-bearing rather than stylistic. This does not change the conditional verdict: the method is plausible, the theory is internally consistent as a conditional statement, and the ablations are directionally supportive. It does mean acceptance should be conditioned on either releasing code or supplying the calibration and rank-matched control evidence described above. I agree with the reader that this is the weakest point, and I do not see a stronger concern that would move the verdict to rejection.","tokens_in":15386,"tokens_out":1447,"duration_ms":12706,"concrete_test":"Run the exact FOOD-101 N2 protocol and report (a) reliability-diagram calibration error of the Positive Expert's pseudo-label confidence, (b) the empirical value of ||E_r|| / ||G_a + G_-|| from Eq 12 over training, and (c) the full Table 4 ablation with a rank-24 single LoRA control (0.962M params, 3.0x cost) so every component gain is measured against the same capacity. If either the calibration error is high or the empirical rho exceeds 1 at any point, or the rank-matched control shrinks the TriNoL gain below noise, the specialization claim is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that confidence-aware expert specialization drives the improvement is supported by one comparison: TriNoL 88.42% versus rank-matched Single LoRA rank 24 87.73% on FOOD-101 N2 (Table 2, Section 6). All other ablations (Tables 3-4) compare against Single LoRA rank 8 with 0.372M parameters, which does not isolate specialization from capacity. The component ablation (Table 3) shows Positive+Alignment at 88.02% and Positive+Negative at 87.94%, so the marginal contribution of the Negative Expert over Positive+Alignment is only 0.40%, and the confidence-routing-plus-shared-CE ablation (Table 4) is at 87.84%, leaving the unique benefit of the negative loss at 0.58% on this single dataset. None of these gaps are tested for significance or under label corruption; no calibration analysis is reported. Thus the load-bearing premise in Assumption 2 (Eq 12) that confidence routing yields rho < 1 is assumed, not established. The Limitations section explicitly concedes this dependence on miscalibrated predictions or strong domain shift, and the paper provides no calibration curves, reliability diagrams, or temperature scaling checks to show the routing signal holds in the tested noise regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TriNoL, a semi-supervised adaptation method for frozen vision foundation models (VFMs) that replaces a single LoRA adapter with three expert LoRA branches. Unlabeled samples are routed by pseudo-label confidence into a Positive Expert (high confidence, hard pseudo-label cross-entropy), an Alignment Expert (medium confidence, soft KL alignment), and a Negative Expert (low confidence, a negative-learning loss that suppresses commitment to the top pseudo-label). The backbone remains frozen; only the LoRA experts and classifier head are trained. The paper includes a theoretical analysis of gradient contamination and low-rank subspace stability, plus experiments on CIFAR-100, FOOD-101, Semi-Aves, and ImageNet. The authors report that TriNoL improves over several SSL/PEFT baselines in low-label and noisy pseudo-label settings, and they include an efficiency comparison against a rank-matched Single LoRA baseline.","tokens_in":15858,"tokens_out":5353,"duration_ms":51597,"significance":"If the empirical claims are reliable, the core idea—using pseudo-label confidence to control the adaptation pathway rather than only the loss weight—is a timely and potentially useful contribution for parameter-efficient SSL with foundation models. The paper has several strengths: it honestly reports settings where TriNoL does not win (CIFAR-100 N25/N100, ImageNet 10%); it includes a parameter-matched efficiency comparison (Table 2); and its routing ablation (Table 4) separates the effect of confidence routing from the use of multiple experts. The theoretical bounds are clearly stated, even though they are conditional on assumptions. However, the empirical support for the central mechanism is thinner than the claims require: the key rank-matched comparison is reported on only one dataset setting, several improvements are within one standard deviation, ImageNet results lack error bars and significance information, and the sensitivity analysis suggests hyperparameters were selected on the evaluation datasets. The paper's theoretical Section 4 also partly restates the desired conclusion as an assumption.","major_comments":[{"comment":"The claim that TriNoL's improvement comes from confidence-aware expert specialization rather than from increased LoRA capacity rests on a single comparison: FOOD-101 N2, where TriNoL (88.42) beats Single LoRA rank 24 (87.73). No rank-matched Single LoRA rank 24 results are reported for CIFAR-100 N4, Semi-Aves (in-distribution or OOD), ImageNet 1%, or any other configuration. Without such controls, the alternative explanation—that three independent low-rank updates are beneficial even without confidence-aware routing—is not excluded. Please add rank-matched comparisons for at least the other main datasets, or explicitly restrict the specialization claim to FOOD-101 N2.","section":"Section 6, Table 2"},{"comment":"The sensitivity analysis in Table 5 appears to select the confidence thresholds (tau_minus=0.3, tau_plus=0.7) and loss weights (lambda_a=1.0, lambda_n=0.1) on the same datasets that are then used for the main results, FOOD-101 N2 and Semi-Aves. If no held-out validation split was used, the reported accuracies are the result of test-set tuning, which can inflate the apparent improvement over baselines that were not tuned in the same way. The paper should state explicitly which split was used for hyperparameter selection and, ideally, evaluate the final configuration on a separate test set or use nested validation.","section":"Section 5, Table 5"},{"comment":"Theorem 1 (Eq. 13) is essentially a restatement of Assumption 2 (Eq. 12): assuming ||E_r|| <= rho ||G_a + G_-|| with rho < 1 is exactly the claim that the confidence routing already reduces the contaminating gradient norm entering the Positive Expert. The theorem therefore provides no independent theoretical justification for TriNoL. Moreover, the paper does not empirically verify Assumption 2—for example, it does not measure the residual routing error E_r, compare ||E_r|| to ||G_a + G_-||, or provide calibration curves/reliability diagrams showing that the confidence ordering is reliable in the tested noise regimes. The Limitations section concedes this dependence, but the main text should either weaken the theoretical claim or supply direct evidence for the assumption.","section":"Section 4.2, Assumption 2 and Theorem 1"},{"comment":"Several of the headline improvements are small relative to the reported variation. For example, on CIFAR-100 N4 TriNoL is 80.92 +/- 0.19 vs FineSSL 80.66 +/- 0.24; on Semi-Aves mixed OOD it is 61.82 +/- 0.20 vs FineSSL 61.30 +/- 1.19. The ImageNet 1% and 10% rows have no standard deviations at all, so the reader cannot tell whether the 74.72 vs 74.46 difference on ImageNet 1% is meaningful. The component ablation in Table 3 also shows that adding the Negative Expert over Positive+Alignment gives only +0.40 on FOOD-101 N2 and +0.31 on Semi-Aves mixed, which are within one standard deviation of the component settings (e.g., 88.02 +/- 0.35 vs 88.42 +/- 0.33). Please report error bars for ImageNet, run more seeds where practical, and provide a simple significance test (e.g., paired bootstrap or repeated-measures test) for the key comparisons and ablations.","section":"Section 5 and Section 7, Tables 1, 3, and 4"}],"minor_comments":[{"comment":"Figures 1 and 2 are described as supporting the observation that low-confidence samples have higher error rates and that a single adapter shows larger gradient deviation, but the figures are not accompanied by quantitative captions or axis labels in the text; please report the actual error rates and gradient cosine values so the reader can judge the magnitude of these effects.","section":"Section 5, Figures 1 and 2"},{"comment":"Figure 6 refers to 'injected pseudo-label corruption' but the paper does not specify how corruption was injected, at what rate, or on which dataset; please define the protocol so the noise-robustness claim can be reproduced.","section":"Section 7, Figure 6"},{"comment":"The text says lambda_pos is 'usually larger,' but Table 5 reports sensitivity only for lambda_a and lambda_n; please state the fixed value of lambda_pos used in all experiments and whether it was tuned.","section":"Section 3.5, Eq. (7)"},{"comment":"The KL divergence is written as KL(sg(p^w_i) || softmax(z^a(u^s_i))), which is KL(weak || strong); the surrounding text says 'matching the strong-view prediction to the detached weak-view probability,' which suggests the opposite direction. Please clarify the intended direction or correct the notation.","section":"Section 3.5, Eq. (5)"},{"comment":"Theorem 3 assumes local strong convexity and smoothness of the 'reliable objective' L_rel, which is not an object that is directly optimized by TriNoL; the connection between the idealized L_rel and the actual loss in Eq. (7) should be stated more concretely.","section":"Section 4.5, Theorem 3"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible but currently supported by a narrow set of comparisons, and the test-set tuning concern in Section 5 is serious. The reference list also contains a very large number of self-citations by one of the authors (e.g., [13]–[23] in the related work and elsewhere); the editor may wish to verify that all of these are necessary and relevant to the present paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is a modest but genuine increment in semi-supervised PEFT for frozen vision foundation models, and it is honest about where it does not win. The new thing is routing pseudo-labeled samples into three confidence regions, each handled by its own LoRA expert with a different objective - hard CE for high-confidence, soft KL for medium-confidence, and a negative suppression loss for low-confidence. That specific combination is not in the V-PET or FineSSL baselines, and the authors do include a rank-matched comparison: three rank-8 experts versus one rank-24 LoRA at equal parameter count. The component and routing ablations support the claim that the gain comes from confidence-aware specialization rather than just capacity. Credit is earned for the structure and honesty of the empirical work: the limitations section concedes dependence on calibrated confidence, and the paper openly reports settings where TriNoL does not win.\n\nSoft spots, in proportion. First, the theory is much weaker than it looks. Theorem 1 is essentially an assumption in disguise: Assumption 2 asserts that routing already reduces the contaminating gradient norm, and Theorem 1 restates that. The later subspace and stability results are standard low-rank perturbation bounds. I would not use the theory as a reason to accept; the empirical case has to stand alone. Second, the flagship evidence for expert specialization is one comparison on FOOD-101 N2 - 88.42% versus 87.73% for rank-24 single LoRA. The other ablations compare against rank-8 single LoRA, which does not isolate specialization from capacity. The marginal contribution of the Negative Expert over Positive+Alignment is on the order of 0.3-0.6 points, and nothing is significance-tested. Third, no code is released, and ImageNet numbers have no error bars. Fourth, the confidence thresholds and loss weights were selected on the evaluation datasets through the sensitivity analysis, so the headline results are post-selection. That is common in this area, but it should be disclosed more prominently.\n\nNone of this kills the paper. The core idea is coherent, the ablations are directionally consistent, and the gains are plausible for the low-label/noisy regime. I would ask the authors for code, for error bars or a clear statement about single runs on ImageNet, and for a direct test of the routing premise - calibration curves or reliability diagrams, and an analysis showing accepted-but-wrong rates actually differ across the three experts (Figure 6 partly addresses this). I would also ask them to discuss the relationship to their own prior multi-expert SSL papers so the novelty claim is clean.\n\nWho this is for: practitioners doing low-label PEFT-SSL with frozen backbones, and researchers working on pseudo-label noise in foundation-model adaptation. It deserves a serious referee. I would conditionally accept with the above requests.","headline":"Modest but genuine gain from confidence-routed triple-LoRA experts; the paper is honest and well-ablated, but the theory is largely assumed and the flagship gain rests on one dataset.","tokens_in":16191,"tokens_out":2283,"would_cite":true,"duration_ms":22972,"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":"Pseudo-label confidence should choose which LoRA expert learns, not just how much each sample is weighted, and splitting the three confidence regions into separate experts makes semi-supervised vision foundation model adaptation more…","keywords":["semi-supervised learning","vision foundation models","LoRA","pseudo-label noise","confidence-based routing","parameter-efficient fine-tuning","expert specialization","image classification"],"falsifier":"Run TriNoL with the routing rule inverted on a held-out noisy setting, sending low-confidence samples to the Positive Expert and high-confidence samples to the Negative Expert; the paper's mechanism predicts accuracy should fall to roughly single-LoRA levels, so if it does not, the gain is not caused by confidence-aware routing.","tokens_in":15224,"feed_emoji":"🎯","tokens_out":9981,"duration_ms":82888,"temperature":0.7,"pith_summary":"TriNoL argues that in semi-supervised adaptation of a frozen vision foundation model, pseudo-labels are not one uniform training signal: high-, medium-, and low-confidence samples play different roles and should be handled by different low-rank adapters. Instead of feeding all pseudo-label gradients into a single LoRA branch, the method routes high-confidence samples to a Positive Expert trained with hard pseudo-labels, medium-confidence samples to an Alignment Expert trained with soft KL alignment, and low-confidence samples to a Negative Expert whose objective suppresses commitment to the top pseudo-label. The paper shows this separation reduces contamination of the reliable update direction, and demonstrates empirically that a triple-expert rank-8 setup beats a rank-matched single LoRA with the same parameter count, so the gain comes from routing rather than capacity. A reader should care because noisy pseudo-labels are the main obstacle to cheap vision foundation model adaptation with few labels, and TriNoL addresses it by changing which adapter learns, not just how much the loss is weighted.","feed_headline":"Three LoRA experts beat one on noisy pseudo-labels","feed_subtitle":"Routing reliable, ambiguous, and noisy pseudo-labels into separate adapters lifts accuracy at equal parameter cost.","key_machinery":"The machinery is a confidence-routed triple-expert LoRA: a set of three independent low-rank updates $\\Delta W_e = B_e A_e$ applied to frozen projection weights, with routing by the max softmax probability of the Positive Expert's weak-view prediction. Two thresholds $\\tau_-$ and $\\tau_+$ define the Positive, Alignment, and Negative regions, and each expert has a matching objective (hard cross-entropy, soft KL alignment with stop-gradient, and negative learning). This object does the work of decomposing the mixed pseudo-label gradient because each expert occupies a separate low-rank subspace, so the paper's Theorems 1 and 2 bound the deviation of the Positive Expert's update from the reliable gradient and show its learned subspace is closer to the reliable target subspace than a single adapter's is.","core_discovery":"The central claim is that pseudo-label reliability should control the adaptation pathway, not only the loss weight. Concretely, with two confidence thresholds $\\tau_-$ and $\\tau_+$, unlabeled samples are split into three regions that feed three LoRA experts sharing a frozen backbone: hard cross-entropy on high-confidence pseudo-labels for the Positive Expert, Kullback-Leibler divergence to the detached weak-view distribution for medium-confidence samples in the Alignment Expert, and a negative-learning loss for low-confidence samples in the Negative Expert. The argument is carried by gradient decomposition: a single adapter receives $G_{\\mathrm{single}} = G_+ + G_a + G_-$, while the Positive Expert receives $G_+ + E_r$, and under a confidence-routing quality assumption $\\|E_r\\|_2 \\le \\rho \\|G_a + G_-\\|_2$ with $\\rho<1$, the reliable update is less contaminated. At inference only the Positive Expert is used; the other two regularize training. The empirical core is the rank-matched comparison: three LoRA experts at rank 8 reach 88.42% on FOOD-101 N2, versus 87.73% for a single LoRA at rank 24 with the same parameter count, which attributes the improvement to confidence-aware specialization.","pith_inferences":["Extending the paper's logic, the same routing principle should transfer to other parameter-efficient modules such as adapters or prompts, because the mechanism is the separation of gradient subspaces rather than a property of LoRA itself.","A natural test the authors leave implicit is to feed the router a calibration-corrected confidence score; if miscalibration is the bottleneck, accuracy should rise further and the Negative Expert should become less active as training progresses.","If the central claim is right, an oracle router with access to true label correctness is the upper bound of the method, and the gap between TriNoL and that oracle measures how much remains to be gained from better confidence signals."],"forward_implications":["In scarce-label, noisy-pseudo-label regimes such as FOOD-101 with 2 to 10 labels per class and Semi-Aves with mixed out-of-distribution unlabeled data, separating the three confidence regions into specialized adapters yields consistent accuracy gains over the strongest single-path baselines.","The rank-matched efficiency comparison implies the gain is architectural rather than parametric: a single LoRA at rank 24 with equal parameter count is outperformed by three rank-8 experts.","Expert routing is dynamic during training, so the method turns a static confidence threshold into a curriculum that moves samples from uncertain and noisy regions into the reliable region as the model matures.","Because the backbone stays frozen and only the LoRA experts and classifier head are updated, the training cost stays near that of a single adapter, making the robustness cheap to buy.","The method does not dominate everywhere: in more stable settings such as CIFAR-100 N25/N100 and ImageNet 10%, existing baselines remain slightly better, indicating the benefit is specific to low-label and noisy regimes."],"supporting_citations":[{"why":"Supplies the LoRA low-rank update formulation that all three experts are built on.","marker":"[25]"},{"why":"Defines the weak-to-strong pseudo-labeling and confidence-threshold pipeline that TriNoL modifies and must outperform.","marker":"[37]"},{"why":"A strong vision-foundation-model semi-supervised PEFT baseline; TriNoL compares directly against it on noisy and out-of-distribution settings.","marker":"[9]"},{"why":"Represents confidence-based curriculum pseudo-label selection, the family of methods that reweight samples without changing the adaptation pathway.","marker":"[48]"},{"why":"Supplies the frozen vision-language backbone that serves as the pretrained encoder in the adaptation setting.","marker":"[36]"}],"fun_headline_variants":["Three LoRA experts beat one at same cost","Confidence routing gives noisy labels three paths","Route pseudo-labels by reliability, not weight","Split pseudo-label confidence into three LoRAs","Triple LoRA experts lift semi-supervised VFM accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a high max-softmax confidence score genuinely orders pseudo-label correctness; if confidence is miscalibrated or domain shift scrambles that ordering, the Positive Expert still receives contaminated gradients and the whole gain is lost.","fun_headline_variants_meta":{"raw":{"variants":["Three LoRA experts beat one at same cost","Confidence routing gives noisy labels three paths","Route pseudo-labels by reliability, not weight","Split pseudo-label confidence into three LoRAs","Triple LoRA experts lift semi-supervised VFM accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000722,"raw_usage":{"total_tokens":3263,"prompt_tokens":990,"completion_tokens":2273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":2200}},"tokens_in":606,"tokens_out":2273,"duration_ms":16266,"temperature":1.0,"reasoning_tokens":2200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:16:17.314855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TriNoL with the routing rule inverted on a held-out noisy setting, sending low-confidence samples to the Positive Expert and high-confidence samples to the Negative Expert; the paper's mechanism predicts accuracy should fall to roughly single-LoRA levels, so if it does not, the gain is not caused by confidence-aware routing.","supporting_citations":[{"cited_title":"Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, et al","cited_arxiv_id":null,"evidence_quote":"Supplies the LoRA low-rank update formulation that all three experts are built on."},{"cited_title":"Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li","cited_arxiv_id":null,"evidence_quote":"Defines the weak-to-strong pseudo-labeling and confidence-threshold pipeline that TriNoL modifies and must outperform."},{"cited_title":"Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling.Advances in Neural Information Processing Systems, 34: 18408–18419, 2021","cited_arxiv_id":null,"evidence_quote":"Represents confidence-based curriculum pseudo-label selection, the family of methods that reweight samples without changing the adaptation pathway."},{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen vision-language backbone that serves as the pretrained encoder in the adaptation setting."}],"review_version":1}