{"id":"ef5c99cc-a26a-4de0-a813-6970bf5802d4","arxiv_id":"2506.02334","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"RLCD combines a reciprocal two-branch learning framework with a class-wise distribution regularizer to improve both base and novel class accuracy in Generalized Category Discovery.","lead":"This paper introduces RLCD, a training framework for Generalized Category Discovery that adds an auxiliary token specialized in base-class recognition, plus a class-wise distribution regularizer that encourages confident and balanced predictions. On seven benchmarks it reports consistent gains over prior parametric GCD methods, especially for fine-grained datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CDR loss in Eq. (8) is computed from a single batch's probability matrix, and the stated one-hot optimum of Theorem 3.2 is unreachable while Eq.","rationale":"The reader's weakest_assumption focuses on the pseudo-base filter reliability. That is a reasonable concern about the RLF loop, but the paper provides substantial evidence that the filter is not catastrophically wrong: Table 4 shows high base pseudo-label accuracy (e.g., 98.6, 87.1, 96.2, 88.3, 91.4, 75.9, 76.4 across datasets for AUX), Fig. I shows the predicted base count is close to ground truth, and Table 3's ablations show that removing Distill or AUX decreases performance but does not collapse. The reciprocal loop is adequately supported. By contrast, the CDR component's theoretical justification contains a genuine mathematical error (Theorem 3.2 as stated is false), and the batch aggregation is not precisely specified, yet CDR is credited with large novel-class improvements. This is a more concrete and documentable weakness. The concern is routed through internal consistency and correctness risk, not through the pseudo-base loop. I could not fully agree with the reader's choice of weakest assumption because the empirical evidence in the paper mitigates the pseudo-base filter risk, while the CDR theorem issue is an actual flaw in the stated argument. The verdict remains CONDITIONAL: the empirical results are plausible and the ablations support the main claim, but the CDR mechanism should be clarified and re-tested, and the theorem corrected, before accepting the claimed explanation. The test proposed is one concrete experiment plus a trivial analytical counterexample; the latter definitively settles the mathematical issue regardless of implementation details.","tokens_in":19558,"tokens_out":2524,"duration_ms":20532,"concrete_test":"Re-run the CUB200 and CIFAR100 ablations from Table 3, replacing the CDR loss (Eq. 8-9) with a simple per-sample confidence-sharpening loss (e.g., L = -sum_k p_i(k) log p_i(k), the negative entropy of the main branch probabilities) while keeping RLF fixed. If the sharpening baseline matches or exceeds the CDR gains, then the class-wise matrix aggregation in Eq. (7) is not the causal mechanism. Additionally, directly verify Theorem 3.2: construct a two-class example with m_k = m'_k = [0.5,0.5]; this gives L_CDR = 0.25 but m_k not one-hot, falsifying the claimed iff. If the practical training loss is computed with a different normalization than written, the authors should state the exact computation and re-derive the theorem under that definition.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim is that RLCD outperforms prior GCD methods via two mechanisms: the reciprocal pseudo-base filter/distillation (RLF) and Class-wise Distribution Regularization (CDR). The RLF mechanism is plausible and well supported by ablations (Table 3: adding Distill improves base accuracy) and by the pseudo-label accuracy analysis (Table 4). The load-bearing weakness is in the CDR component. Eq. (7) defines m_k = (1 / sum_i p_i^{(k)}) * sum_i p_i^{(k)} p_i, a K x K matrix whose k-th row is a convex combination of per-sample distributions. Eq. (8) then computes L_CDR^(k) = 1 - <m_k, m'_k>, and Theorem 3.2 claims L_CDR^(k) = 0 iff m_k = m'_k is one-hot. However, a one-hot m_k would require all samples in the batch to concentrate all probability mass on a single class k, which contradicts both the requirement that m_k is normalized across the batch (Theorem 3.1) and the fact that each row m_k is a probability distribution over all classes with mass shared across samples. The proof of Theorem 3.2 implicitly treats m_k and m'_k as arbitrary probability vectors and applies Cauchy-Schwarz, but the equality case a_i = b_i for all i plus the \"sum of squares <= 1\" step does not force one-hot; it only forces the two vectors to be identical and each to have l2-norm 1, and the l2-norm-1 condition for probability vectors does hold for any one-hot vector, but the reverse implication (a_i in {0,1}) uses the wrong direction of the inequality (sum a_i^2 <= sum a_i does not imply each a_i is 0 or 1; e.g., a = [0.5,0.5] gives 0.5 <= 1). Thus the theoretical claim in Theorem 3.2 is incorrect as stated. The practical impact is that the CDR loss's aggregation over a batch is underspecified: the paper states Eq. (7) uses 'N is the batch size', but it is not clear whether m is computed over the whole mini-batch or per-class accumulations, whether the normalization is applied per row or globally, and how the two views' matrices are matched.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RLCD, a parametric generalized category discovery method combining a Reciprocal Learning Framework (RLF) and Class-wise Distribution Regularization (CDR). RLF inserts an AUX token and a base-only classifier into a ViT; the main branch filters pseudo-base samples for the auxiliary branch, whose soft labels are distilled back into the main branch. CDR computes class-wise expected distributions from two views of a mini-batch and encourages consistency and confidence. The authors introduce an Oracle Base accuracy metric, report state-of-the-art results on seven GCD benchmarks with both DINO and DINOv2 backbones, and provide ablations, hyper-parameter sensitivity, robustness analyses, and an appendix with theoretical support.","tokens_in":19963,"tokens_out":3305,"duration_ms":34191,"significance":"If the empirical claims hold, RLCD is a useful contribution: the AUX-token reciprocal design is simple, adds negligible inference cost, and the reported gains on fine-grained benchmarks (3.6% average improvement over prior methods on DINO) are practically meaningful. The paper also ships code, defines a new evaluation metric (OB) that directly targets base-class discrimination, and provides extensive comparisons and ablations. The main weaknesses are that the theoretical justification for CDR (Theorem 3.2) is incorrect as stated, and a key design choice (removing InfoNCE) is not ablated in isolation, so the attribution of gains to the proposed components is not fully established.","major_comments":[{"comment":"The proof of Theorem 3.2 is invalid. From the Cauchy-Schwarz equality one obtains a = b, but the subsequent step \"sum_i a_i^2 <= sum_i a_i = 1 implies a_i in {0,1}\" is incorrect: the inequality holds for every probability vector and does not force equality, so it does not imply one-hotness. Moreover, the claimed one-hot optimum is unreachable for the quantity defined in Eq. (7): m_k is a convex combination of per-sample distributions p_i weighted by p_i^(k), so m_k can be one-hot only if every sample with positive p_i^(k) has all its probability mass on class k, which cannot hold for all k simultaneously in a batch containing multiple classes. This is load-bearing because the paper uses Theorem 3.2 to claim that CDR 'increases prediction confidence' and converges to one-hot distributions. The theorem should either be corrected to a valid statement (e.g., an inequality relating L_CDR to a proper scoring rule or to the variance of predictions) or removed and replaced by empirical evidence.","section":"Section 3.3, Theorem 3.2 and Appendix A"},{"comment":"The removal of the InfoNCE loss is described as a deliberate choice ('we chose to remove InfoNCE in our approach'), but no ablation isolates this change. Since SimGCD's objective includes InfoNCE, and the paper argues that InfoNCE 'conflicts with the SupCon loss and impairs feature discrimination,' the reported gains over SimGCD could be substantially due to this removal rather than to RLF or CDR. Please add an ablation that runs the full RLCD pipeline with and without InfoNCE, and ideally also runs a SimGCD baseline with InfoNCE removed, so the contribution of each component is identifiable.","section":"Section 3.1 and Table 3"},{"comment":"No standard deviations or multiple seeds are reported. Some improvements are small relative to likely run-to-run variation (e.g., CIFAR10 with DINOv2: 99.0 vs 99.0; several fine-grained numbers), and the claim of 'consistently outperforming' all methods would be strengthened by reporting means and standard deviations over at least three seeds. This is particularly important because GCD training involves stochastic pseudo-label filters and entropy regularization.","section":"Tables 1, 2 and all experimental tables"},{"comment":"The reciprocal learning mechanism assumes that the main branch's pseudo-base filter is reliable enough that feeding filtered pseudo-base samples to the auxiliary branch creates a 'virtuous cycle.' Table 4 reports final pseudo-label accuracy of the AUX branch, but it does not measure the precision of the filter during training, especially in the early epochs when many novel samples may be misrouted into the base set. Please report the precision and recall of the pseudo-base filter as a function of training epoch, or include an upper-bound ablation that uses the ground-truth base mask instead of the pseudo-base filter, to quantify how much performance depends on filter quality.","section":"Section 3.2 and Table 4"}],"minor_comments":[{"comment":"The total CDR loss L_CDR is used in Eq. (9) but never explicitly defined. Please specify whether it is the sum or the mean of L_CDR^(k) over classes, since this affects the effective gradient scale and the interpretation of the hyper-parameter beta.","section":"Equation (9)"},{"comment":"The notation max(p_aux_b,i) should be clarified: it is the maximum over base-class probabilities of the auxiliary branch for sample i. Please also state explicitly that the maximum is taken over the base-class output dimension.","section":"Equation (5)"},{"comment":"The comparison of parameter counts removes the projector in RLCD while SimGCD and LegoGCD keep it. This is a fair design choice, but the text should note that the projector removal is an additional difference from SimGCD, not a consequence of the proposed AUX token, so readers can separate its effect from the proposed components.","section":"Appendix D, Table 8"},{"comment":"There are two typos: 'Semantic Shit Benchmark' should be 'Semantic Shift Benchmark,' and 'GCP' in Appendix G should be 'GPC.'","section":"Introduction and Appendix G"},{"comment":"The hyper-parameter sensitivity plots report All/Base/Novel accuracy for alpha and beta but do not include error bars or repeated runs, so the apparent flatness or peaks may not be significant. Adding multiple seeds would make the recommended default (alpha=beta=0.5) more convincing.","section":"Figure 6"},{"comment":"The supervised cross-entropy (SupCE) loss curves are presented as evidence of reduced label noise, but the plot reports only the training loss, not label accuracy. Reporting the pseudo-label accuracy of the main branch over time (or the fraction of samples whose soft label agrees with the ground-truth base label) would more directly support the claim.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper's central empirical claim—that RLCD achieves state-of-the-art GCD accuracy with negligible inference overhead—is credible and well supported by the experimental breadth. The main blocker is the incorrect Theorem 3.2, which is used to justify a core design element; this is fixable by re-stating the theorem or replacing it with empirical analysis, but it needs to be addressed before publication. The missing InfoNCE-removal ablation is also important for attributing the gains. I would not reject, but I would require a revised version that addresses these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, useful empirical result for parametric Generalized Category Discovery. The authors add one AUX token, a base-only auxiliary classifier, and two loss terms to SimGCD, and obtain consistent state-of-the-art numbers on seven datasets across DINO and DINOv2 backbones. The claimed inference overhead is negligible, and the ablations support each component's contribution. The low-label-ratio and estimated-K experiments strengthen the case. This is a legitimate contribution, not a repackaging.\n\nThe genuinely new piece is the Reciprocal Learning Framework: the main branch filters pseudo-base samples to the AUX branch, and the AUX branch distills its more confident base-class soft labels back to the main branch. Table 4 shows the AUX branch really does produce better base pseudo-labels, which directly addresses the obvious worry that the filter amplifies early bias. The CDR regularizer is explicitly borrowed from Zhang et al. 2024, credit given, and its use here to counter base/novel imbalance is a fair extension.\n\nSoft spots, in order of real concern. First, no error bars or multiple seeds anywhere. Reported gains are large, so variance matters. Second, the proof of Theorem 3.2 is written incorrectly: the final step claims sum a_i^2 <= sum a_i = 1 forces a_i in {0,1}, which is false as stated. The theorem itself is true, because equality in the chain requires sum a_i^2 = 1, which does force one-hot. It is a presentation flaw, not a conceptual one, but it should be fixed. Third, Eq. (9) never says whether the CDR loss is summed or averaged over classes and branches; presumably a mean, but say it. Fourth, InfoNCE removal is explained verbally but never ablated; a quick experiment isolating it would add rigor.\n\nI disagree with the stress-test note's claim that Theorem 3.2 is false. The proof is sloppy; the theorem holds. The one-hot optimum being unreachable in practice is irrelevant because CDR is a regularizer, not a constraint.\n\nBottom line: this paper deserves serious peer review, and by the venue stamp it got it. For GCD researchers it is a new strong baseline worth citing. For others, it is a clean example of a well-scoped plug-in improvement. I would engage with it.","headline":"A strong, honest empirical GCD paper: RLF plus CDR gives a real plug-in gain, with a sloppy but salvageable theorem proof and missing error bars.","tokens_in":20607,"tokens_out":3519,"would_cite":true,"duration_ms":33179,"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":"Adding an auxiliary base-class branch and a class-wise distribution loss improves generalized category discovery.","keywords":["Generalized Category Discovery","parametric clustering","reciprocal learning","auxiliary token","class-wise distribution regularization","pseudo-label distillation","base class discrimination","novel class discovery"],"falsifier":"On CUB200, replace the pseudo-base filter with a random subset of equal size during training; if all-class accuracy does not drop well below the reported 70.0, the reciprocal filter is not the mechanism behind the gain. A complementary check: measure the auxiliary branch's base pseudo-label accuracy in the first 50 epochs; if it is not higher than the main branch's, distillation is not supplying more reliable labels.","tokens_in":19351,"feed_emoji":"🎯","tokens_out":7561,"duration_ms":65583,"temperature":0.7,"pith_summary":"The paper claims that parametric Generalized Category Discovery (GCD) methods, which cluster unlabeled data that mix known and unknown classes, fail mainly because their base-class discrimination is weak. It proposes RLCD, a two-branch training setup: a shared encoder produces a main all-class classifier and an auxiliary base-only classifier, and the two teach each other so that reliable base soft labels circulate back to the main branch. A class-wise distribution regularization term then sharpens predictions on unlabeled data, which pushes more novel samples to be predicted as novel. Across seven GCD benchmarks, this combination raises all-class accuracy, including base and novel accuracy separately, while keeping inference cost nearly unchanged because the auxiliary branch is dropped at test time. This matters because GCD is a practical open-world setting where unlabeled data contains both familiar and never-seen categories.","feed_headline":"Auxiliary branch adds 3.6% to fine-grained category discovery","feed_subtitle":"A token plus a class-wise distribution loss raises base and novel accuracy with no extra inference cost.","key_machinery":"The load-bearing objects are the AUX token and the CDR loss. The AUX token is a single trainable parameter inserted before the last block of the ViT; its output is dedicated to a base-only classifier, giving the model a specialist module for known classes. The reciprocal loop consists of a pseudo-base filter (the main branch selects samples it predicts as base) and cross-branch distillation, with KL divergence weighted by the auxiliary branch's maximum probability. The CDR loss computes, for each category k, a class-wise expected distribution m_k over the mini-batch and maximizes the inner product between the two augmented views' distributions, which the authors prove equals one only when both views agree and are one-hot; the loss thus drives predictions toward confident, view-consistent distributions and treats every class equally, counteracting the base-class bias introduced by the reciprocal loop.","core_discovery":"The central discovery is that the bottleneck in current GCD models is unreliable soft labels from self-supervision, which hurt base-class discrimination. RLCD addresses this with a Reciprocal Learning Framework: a trainable AUX token is inserted before the final transformer block; its output feeds a base-only classifier, while the CLS token continues to feed the all-class classifier. During training, the main branch filters samples it predicts as base classes and sends them to the auxiliary branch; the auxiliary branch, in turn, distills its base-class distribution back to the main branch, weighted by the auxiliary confidence. This creates a virtuous cycle because the auxiliary branch specializes in base classes and provides more accurate soft labels than the main branch's own self-supervision. On top of this, Class-wise Distribution Regularization (CDR), which enforces consistency between class-wise expected predictive distributions of two augmented views, increases prediction confidence and mitigates a learned bias toward base classes, improving novel-class accuracy. After training only the main branch is kept, so the extra cost during inference is negligible.","pith_inferences":["We infer that the same reciprocal structure could apply to other token-based classifiers, such as multi-modal encoders, where a dedicated known-class token might guard against open-world drift.","We infer that the fixed CDR weight likely needs rescaling with batch size and class imbalance, and a batch-size-adaptive beta would be a natural testable variant.","We infer that the larger gains observed at 10% labeled data indicate the method acts as a form of self-distillation that is most valuable under scarce supervision.","We infer that because the paper removes InfoNCE alongside adding RLF and CDR, the isolated contribution of each change is not fully separated; a two-by-two ablation of InfoNCE removal and RLF+CDR would test whether the reciprocal loop alone drives the improvement."],"forward_implications":["Base-class discrimination in GCD can be improved without extra inference cost, because at test time only the main branch and its all-class classifier are used.","The gap between supervised-only base accuracy and GCD base accuracy can be closed or even inverted, with RLCD reporting oracle base accuracy above the supervised reference on CIFAR100 and Stanford Cars.","Class-wise distribution regularization is a label-free consistency loss, so it can be applied to any branch and should transfer to other semi-supervised or open-world clustering settings.","RLCD remains effective when the number of classes is estimated rather than given, and when the labeled fraction is small, as shown by 10% labels on CUB200, suggesting the mechanism does not depend on exact supervision.","Removing either the auxiliary branch or the distillation step drops base accuracy substantially, so the reciprocal loop is a genuinely load-bearing component rather than a side effect of extra parameters."],"supporting_citations":[{"why":"SimGCD, the parametric baseline this method extends and the main comparison, supplies the parametric clustering framework and loss components.","marker":"Wen et al., 2023"},{"why":"Defines the GCD problem, the dataset splits, and the clustering-based baseline that the paper improves upon.","marker":"Vaze et al., 2022b"},{"why":"Source of the class-wise expected distribution definition adopted by CDR, along with the proof that it sums to one.","marker":"Zhang et al., 2024"},{"why":"LegoGCD, a comparison method that addresses base-class forgetting in SimGCD and is used in the oracle base accuracy analysis.","marker":"Cao et al., 2024"},{"why":"DINO, the self-supervised ViT backbone on which the main experiments are conducted.","marker":"Caron et al., 2021"},{"why":"DINOv2, the second backbone that demonstrates the method's gains generalize across pre-trained feature extractors.","marker":"Oquab et al., 2024"},{"why":"CRNCD, the distillation-based NCD method contrasted in Appendix F, justifying the one-stage base-class distillation design.","marker":"Gu et al., 2023"}],"fun_headline_variants":["Reciprocal learning loop sharpens base and novel class accuracy","Auxiliary token and distribution loss lift GCD by 3.6%","GCD via reciprocal learning: virtuous cycle for base classes","RLCD: trainable auxiliary branch boosts category discovery","Base-class filter and soft labels fix GCD self-supervision"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reciprocal loop only helps if the main branch's pseudo-base filter is accurate enough that most samples it sends to the auxiliary branch are genuinely base-class samples; if early misrouting sends many novel samples in, the auxiliary branch will amplify that bias.","fun_headline_variants_meta":{"raw":{"variants":["Reciprocal learning loop sharpens base and novel class accuracy","Auxiliary token and distribution loss lift GCD by 3.6%","GCD via reciprocal learning: virtuous cycle for base classes","RLCD: trainable auxiliary branch boosts category discovery","Base-class filter and soft labels fix GCD self-supervision"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1514,"prompt_tokens":951,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":567,"tokens_out":563,"duration_ms":5505,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:26:07.255193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On CUB200, replace the pseudo-base filter with a random subset of equal size during training; if all-class accuracy does not drop well below the reported 70.0, the reciprocal filter is not the mechanism behind the gain. A complementary check: measure the auxiliary branch's base pseudo-label accuracy in the first 50 epochs; if it is not higher than the main branch's, distillation is not supplying more reliable labels.","supporting_citations":[{"cited_title":"Parametric classification for generalized category discovery: A baseline study","cited_arxiv_id":null,"evidence_quote":"SimGCD, the parametric baseline this method extends and the main comparison, supplies the parametric clustering framework and loss components."},{"cited_title":"Rethinking guidance information to utilize unlabeled samples: A label encoding perspective","cited_arxiv_id":null,"evidence_quote":"Source of the class-wise expected distribution definition adopted by CDR, along with the proof that it sums to one."},{"cited_title":"Solving the catastrophic forgetting problem in generalized category discovery","cited_arxiv_id":null,"evidence_quote":"LegoGCD, a comparison method that addresses base-class forgetting in SimGCD and is used in the oracle base accuracy analysis."},{"cited_title":"DINO v2: Learning robust visual features without supervision","cited_arxiv_id":null,"evidence_quote":"DINOv2, the second backbone that demonstrates the method's gains generalize across pre-trained feature extractors."}],"review_version":1}