acceptodds
Under review as a conference paper at ICLR 2027

ProtoTail: Prototype-Guided Representation and Pseudo-Labeling for Long-Tailed Fine-Grained Generalized Category Discovery

Abstract

Generalized Category Discovery (GCD) in fine-grained recognition is challenging because categories often differ only in subtle local cues, while long-tailed training provides uneven evidence for learning category-specific representations. When category support is limited, flexible local representations can introduce weakly supported cues, and unreliable predictions may be reinforced through pseudo supervision. We propose \sc ProtoTail, a part-aware GCD framework that couples support-aware local representation learning with reliability-aware pseudo-label selection. ProtoTail uses class-conditioned prototypes over shared latent parts to capture category-specific local evidence, while category-wise gating regularizes effective local-part capacity according to estimated category support. In parallel, separate known- and novel-category selection rules progressively incorporate reliable unlabeled samples using the validation signals available to each group. The two components complement each other: improved local representations support more reliable sample selection, while accepted samples further refine category-specific representations. Under long-tailed training, ProtoTail achieves \bf 75.42% All accuracy on CUB and \bf 68.29% on Stanford Cars, with its clearest benefits on low-frequency categories; under balanced training, it reaches \bf 79.58% and \bf 76.30%, respectively. These results support support-aware local regularization and progressive pseudo supervision as complementary mechanisms for fine-grained GCD under uneven category support. Our code is available at: https://anonymous.4open.science/r/ProtoTail-FC6E

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Generalized Category Discovery, Long-Tailed, Pseudo-Labeling, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8S6uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes ProtoTail for long-tailed fine-grained GCD, combining prototype-guided representations and pseudo-label selection. However, limited novelty and poor presentation undermine its contributions.

Critical strengths

The paper addresses the practical challenge of long-tailed fine-grained GCD and reports improvements on CUB and Stanford Cars.

Critical weaknesses
  1. Poor presentation and figure quality. Figure 1 is excessively dense, with extremely small text and poorly organized modules.
  2. The methodology relies heavily on generic prototype learning, gating, and pseudo-label filtering. It does not adequately explain how these mechanisms address the distinctive challenges of long-tailed fine-grained category discovery beyond existing part-based approaches.
  3. The long-tailed experiments cover only CUB and Stanford Cars with a single imbalance ratio of 10. Comparisons against recent long-tailed GCD methods are insufficient, and the balanced results do not outperform the strongest competitor in average accuracy.
Questions for authors

see weaknesses

1 comment
AuthorsOpenReview ↗

Responses to Reviewer 8S6u

We thank the reviewer for the feedback.

W1: Presentation and Figure 1. We agree that Figure 1 is too dense. We will redesign it as a simplified vector diagram with two clearly separated modules, larger text, and notation consistent with Section 3.

W2: Addressing long-tailed fine-grained GCD specifically. The specific difficulty we target is that fine-grained discrimination depends on category-specific local evidence, and under a long tail that evidence is estimated from very different numbers of samples. Existing part-based GCD methods give every category the same local flexibility. ProtoTail instead constrains how many local parts each category can effectively use (gate mass κ_c) according to its estimated support, so tail categories do not overfit weakly supported cues. The effect appears where it is designed to act: adding the tail-aware term to the full model raises K-Few from 68.90% to 75.52% on CUB and U-Few from 52.00% to 61.36% on Stanford Cars (Table 2). The same trend holds on CIFAR100-LT (U-Few 60.20 → 82.20) and on Herbarium 2019 (K-Few +3.98). A second point specific to this setting is that pseudo-label selection alone hurts a long-tailed base model, and becomes useful only together with the support-aware representation (Table 1).

W3: Long-tailed evaluation and balanced comparison. We extended the evaluation in two ways: stronger imbalance on CIFAR100-LT (IR=20/50; All +3.92 / +4.05, Table 3) and the naturally long-tailed Herbarium 2019 (All 40.25 → 44.50, Table 4). Our base method in all long-tailed experiments, BaCon, is itself a long-tailed GCD method. We agree that ProtoTail is not the best on average balanced accuracy: SEAL reaches 76.3 vs. our 76.1 (Table 5), and PartCo and EAGC report larger gains on some datasets (Table 8). We state this explicitly. Our main target is the long-tailed regime, where the method gives its clearest benefits on low-frequency categories.

Reviewer_KtYNOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper focuses on long-tailed fine-grained generalized category discovery and proposes the ProtoTail approach. The proposed ProtoTail mainly contains two modules: (i) prototype-guided local representation learning, which introduces class-conditioned prototypes over shared latent parts, together with category-specific gating and tail-aware capacity regularization; and (ii) dual-branch pseudo-label selection, which employs separate pseudo-label selection strategies for known and novel categories, using classification confidence and clustering-based reliability measures, respectively. ProtoTail is evaluated on CUB and Stanford Cars under both balanced and artificially constructed long-tailed settings. It improves the overall accuracy over Baseline by 1.7% on CUB and 1.43% on Stanford Cars under the long-tailed setting.

Critical strengths
  • The idea of tail-aware representation capacity is interesting. By adjusting the effective number of local parts according to category support, the method aims to reduce overfitting for underrepresented categories while allowing more flexible representations for categories with sufficient training samples.
Critical weaknesses
  • My main concern is the experimental evaluation. (i) From my perspective, this paper mainly focuses on the long-tailed setting. However, the authors only evaluate ProtoTail on artificially constructed long-tailed splits of CUB and Stanford Cars, which may not adequately reflect real-world long-tailed distributions. Since iNaturalist is a naturally long-tailed fine-grained dataset, I believe it would be a more convincing benchmark for validating the proposed method. (ii) Moreover, under the long-tailed setting, ProtoTail is mainly compared against the BaCon baseline, which is insufficient to demonstrate its effectiveness. (iii) For the balanced setting, comparisons with recent GCD methods, such as [a], [b], and [c], are missing. All three methods report better performance on some of the same fine-grained GCD benchmarks. Including these methods in the comparison would help better establish the competitiveness of the proposed method. Overall, I believe a more comprehensive experimental evaluation, including additional datasets and stronger baselines, is necessary to support the paper's claims.
  • The tail-aware capacity regularization is one of the main contributions of this paper. However, the authors only compare the results with and without this component. I am not convinced that the improvements are specifically due to frequency-aware capacity allocation, rather than simply regularizing the number of effective local parts. A comparison with uniform capacity regularization would be helpful. Also, I wonder why the authors still use tail-aware regularization in the balanced setting, where class frequencies are relatively uniform.

[a] The Devil Is in Gradient Entanglement: Energy-Aware Gradient Coordinator for Robust Generalized Category Discovery, CVPR 2026. [b] Fourier Self-Supervision for Fine-Grained Generalized Category Discovery, ECCV 2026 [c] BURES-ISOTROPY ALIGNMENT: MANIFOLD LEARNING OF GENERALIZED CATEGORY DISCOVERY, ICLR 2026.

Questions for authors
  • First, I would appreciate it if the authors could address my main concerns raised in the Critical Weaknesses section.
  • How accurately can the method estimate the category frequencies, especially for novel categories without ground-truth labels? Since these estimates are used to determine the representation capacity, I wonder how sensitive the method is to errors in frequency estimation.
  • The novelty of the overall framework is somewhat limited. The shared latent-part representation builds on existing part-based GCD approaches, while prototype matching, gating, and confidence-based pseudo-labeling are established techniques. It would be better to clarify what they consider the main technical contribution of this work and how it differs from existing methods.
1 comment
AuthorsOpenReview ↗

Responses to Reviewer KtYN

We thank the reviewer for the constructive and specific suggestions.

Q1(i): Natural long-tailed benchmark. We agree that synthetic splits alone are insufficient. We added Herbarium 2019, a naturally long-tailed fine-grained benchmark widely used in GCD (Table 4). With BaCon, ProtoTail improves All accuracy from 40.25% to 44.50% and Known-Few from 21.95% to 38.96%. Comparing Full (+Tail) with Full (w/o Tail) isolates the tail-aware term on this natural distribution: All +0.67, Known-Few +3.98, Novel-Few +1.56. We also added stronger synthetic imbalance on CIFAR100-LT (IR=20/50, Table 3). We consider iNaturalist a valuable next benchmark.

Q1(ii): Long-tailed baselines. BaCon is a dedicated long-tailed GCD method. ProtoTail is built on it, and all long-tailed comparisons use identical splits. We agree that more long-tailed GCD baselines would strengthen the evaluation.

Q1(iii): Recent GCD methods [a-c]. Appendix A.5 (Table 8) compares improvements over SimGCD (DINOv2) as reported by each method's authors:

MethodCUB All (gain)Cars All (gain)
ProtoTail79.58 (+8.08)76.30 (+4.80)
PartCo81.1 (+9.6)78.9 (+7.4)
EAGC77.9 (+6.4)81.5 (+10.0)
AllGCD78.4 (+6.9)76.2 (+4.7)

ProtoTail is competitive but not the best in the balanced setting, and we state this explicitly. Our primary target is the long-tailed regime. EAGC and BIA are plug-and-play modules that act on gradients and on representation geometry, respectively, which is a different level from our capacity regularization and pseudo-label selection. Combining them is a promising direction.

W2: Is the gain from frequency-aware allocation or from capacity regularization alone? This is a fair point. The current evidence isolates the tail-aware term, and its effect is concentrated where support is low:

  • Full, w/o → +Tail: K-Few 68.90 → 75.52 (CUB) and 54.80 → 59.38 (Cars); U-Few 52.00 → 61.36 (Cars).
  • Prototype, w/o → +Tail: U-Few 68.60 → 76.13 (CUB) and 50.70 → 60.50 (Cars).
  • CIFAR100-LT, IR=10 (Table 7): U-Few 60.20 → 82.20.
  • Herbarium 2019: K-Few +3.98.

A regularizer that limited capacity uniformly would not, in general, help mainly the rare categories. This pattern is consistent with support-dependent allocation. We agree, however, that a uniform-capacity control with the same average budget would be the most direct test, and we will add it.

W2b: Why use tail-aware regularization in the balanced setting? We keep one configuration for both settings to avoid tuning per setting. The support statistic is never the true class frequency. It is estimated by clustering (BaCon's estimator), and for novel categories it reflects how well each category has been discovered so far. Even when the true distribution is balanced, the estimates are not uniform, and categories with less reliable evidence receive a tighter capacity target. When support is near-uniform, the targets become less spread out and the regularizer acts as a mild constraint.

Q2: Accuracy of frequency estimation. We use BaCon's distribution estimator, refreshed every 10 epochs. Known categories take the sizes of their aligned clusters, and the remaining sizes are assigned to novel categories. Two design choices limit sensitivity to estimation error:

  • The mapping to capacity uses a log-scale normalization clipped to [1, M], so it depends mainly on the ordering of the estimated support rather than on exact counts.
  • Capacity is regularized softly toward the target; it is not fixed.

We agree that a direct comparison between estimated and true category sizes, and a test with noise added to the estimates, would make this explicit. We will add both.

Q3: Main technical contribution. We agree that prototypes, gating and confidence-based pseudo-labeling are established techniques, and we do not claim them individually. Our contributions are:

  1. Support-dependent local capacity. Part-based GCD methods (Dai et al., 2025; Wang et al., 2026) share latent parts across categories but give every category the same local flexibility. ProtoTail introduces a per-category gate mass κ_c, which measures the effective number of local parts a category uses, and regularizes it toward a target set by the category's estimated support. Categories with fewer observations are restricted to fewer category-specific local degrees of freedom.
  2. Selection rules split by available validation signal. Known categories are validated against labeled data. Novel categories, which have no labels, are validated through clustering stability, separation, prediction agreement and temporal consistency.
  3. Coupling between representation and selection. Under the long-tailed BaCon setting, selection alone reduces accuracy (CUB 73.72 → 73.40; Cars 66.86 → 65.70). Combined with the support-aware representation, it produces the best results (75.42 / 68.29). The two components are complementary, not simply additive.
Reviewer_iCcZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the Generalized Category Discovery task under the long-tail distribution. The authors present ProtoTail, a framework that combines class-conditioned prototypes with a category-wise gate to allocate local representation capacity based on estimated category support. It also introduces pseudo-label selection branches for known and novel category learning. Experiments on fine-grained benchmarks report gains in both balanced settings and long-tailed settings.

Critical strengths
  1. The paper focuses on the interaction between fine-grained visual cues, long-tail distribution, and noisy pseudo-labeling, which are all important for balanced GCD.
  2. The proposed components are motivated and complementary to solve the proposed challenges.
  3. The ablation studies validate the effectiveness of the proposed framework.
Critical weaknesses
  1. Fig. 1 seems to be generated by AI, so it is hard to see clearly.
  2. The pseudo-label selection procedure relies on several heuristics and fixed thresholds, but the paper does not establish how sensitive the results are to these choices or how well they transfer to different datasets and imbalance levels.
  3. The main component and long-tailed analyses focus on CUB and Stanford Cars, using an imbalance ratio of 10, which is relatively narrow and cannot reflect the effectiveness of the proposed method on extreme unbalanced situation.
  4. The performance gains are limited on many benchmarks and metrics.
  5. The computational cost is not characterized in enough detail. In particular, the repeated clustering, silhouette computation, and temporal checks may add nontrivial overhead.
  6. Every long-tailed experiment (Tables 1, 2, 4, 5) compares ProtoTail only against the BaCon and SimGCD baselines that the authors re-trained themselves. Not a single long-tailed GCD method is compared under the long-tailed setting.
Questions for authors

See weakness.

1 comment
AuthorsOpenReview ↗

Responses to Reviewer iCcZ

We thank the reviewer for the detailed comments.

W1: Figure 1. We agree that Figure 1 is too dense and that its text is too small. We will redesign it as a simplified vector diagram with two clearly separated modules (local representation and pseudo-label selection), larger labels, and notation that matches Section 3.

W2: Heuristics and fixed thresholds. We agree that the selection procedure has several thresholds. We do not tune them per setting. One fixed set of values (Appendix B) is used for every dataset (CUB, Stanford Cars, FGVC-Aircraft, CIFAR100-LT, Herbarium 2019), both base methods (SimGCD, BaCon), the balanced setting and every imbalance level (IR=10, 20, 50, and the natural long tail). ProtoTail improves All accuracy over its base method in all of these long-tailed settings (Tables 1, 3, 4, 6, 7). This suggests the thresholds transfer reasonably well, although it is not a formal sensitivity analysis. We list this reliance on fixed thresholds as a limitation and will add a one-at-a-time sensitivity study of the main thresholds.

W3: Only IR=10. We added two further evaluations:

  • Stronger imbalance: CIFAR100-LT at IR=20 and IR=50 (Table 3). All +3.92 / +4.05; Known-Few +8.5 / +20.0; Novel-Few +7.0 / +5.4.
  • Natural long tail: Herbarium 2019 (Table 4). All 40.25 → 44.50; Old 45.66 → 62.50; Known-Few 21.95 → 38.96.

The gains are not uniform. New accuracy decreases by 4.14 points at IR=50 on CIFAR100-LT and slightly on Herbarium 2019 (37.81 → 36.39), and we state this in the paper.

W4: Limited gains. The All-accuracy gains on CUB and Stanford Cars at IR=10 are moderate (+1.70 and +1.43). As intended by the design, however, the gains concentrate on low-frequency categories. Known-Few improves by +5.14 on CUB and +5.57 on Stanford Cars, and Novel-Few improves by +8.42 on Stanford Cars (Table 2). Larger gains appear in other long-tailed settings: +7.51 and +3.01 All with SimGCD at IR=10 (Table 6), +6.30 on CIFAR100-LT at IR=10 (Table 7), and +4.25 on Herbarium 2019 (Table 4). The results are single-run, and we will add multi-seed statistics.

W5: Computational cost. The added cost is bounded by design (Appendix B).

  • The novel-category branch runs at most six times, only during epochs 50-100. Each run fits two k-means partitions on at most 15k samples and computes Silhouette scores on a random 2k subset.
  • The known-category branch only ranks existing predictions by confidence.
  • The support statistic reuses BaCon's estimator and is refreshed every 10 epochs.
  • The local branch adds M=3 slot queries, a prototype bank and per-category gate logits.

All selection steps are used only during training. We will add a wall-clock and memory breakdown.

W6: No long-tailed GCD method compared. BaCon (Bai et al., NeurIPS'23) is itself a dedicated long-tailed GCD method. All our long-tailed experiments (Tables 1, 2, 4, 7) use it as the base and evaluate it under identical splits, so every long-tailed comparison is against a long-tailed GCD method. Using the same precomputed splits for every configuration also rules out differences caused by data sampling. We agree that adding further long-tailed GCD methods would strengthen the evaluation, and we will include them.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.