acceptodds
Under review as a conference paper at ICLR 2027

DD-FSC: Dirichlet Process-Guided Diffusion Model for Few-Shot Clustering

Abstract

In few-shot deep clustering, partial category coverage and an unknown number of latent categories make it difficult to propagate sparse semantic supervision while reliably recovering the complete clustering structure. To address this challenge, we propose DD-FSC, a nonparametric few-shot clustering framework that jointly models feature-space density and latent cluster structure. DD-FSC characterizes the feature distribution with a diffusion model and employs score-guided Langevin dynamics to propagate limited supervision toward high-density semantic regions, thereby identifying high-confidence same-category samples and constructing density-aware positive pairs. Meanwhile, Dirichlet process (DP) inference models the evolving cluster structure and adaptively determines the effective number of clusters without requiring it to be specified in advance. Theoretically, we show that the deterministic drift of score-guided Langevin dynamics is positively aligned with the modeled log-density gradient, providing a principled basis for density-guided semantic propagation. Through this density-guided semantic expansion and adaptive structure inference, DD-FSC extends sparse supervision beyond the observed categories while recovering the underlying clustering structure. Extensive experiments demonstrate superior clustering performance under both known- and unknown-cluster-number settings, together with accurate estimation of the latent category count.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
few-shot clustering, deep clustering, diffusion models, Dirichlet process, nonparametric clustering

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KEGUOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies transductive few-shot clustering with labels covering only a subset of categories. DD-FSC combines a noise- conditional score model with Dirichlet-process mixture inference. Score-guided feature trajectories refine semantic anchors, expand the labeled groups with confident unlabeled samples, and provide targets for representation learning. The inferred cluster structure feeds back into the trajectories and a prototype-level objective. The framework supports both known and adaptively inferred cluster counts.

Critical strengths
  1. The paper is well written and well structured.

  2. A useful framework for partial category coverage. The method connects sparse semantic supervision, density-based representation refinement, and adaptive cluster discovery. This addresses a meaningful setting where some categories have no labeled examples and the cluster count may be unknown. The feedback between representation learning and cluster inference is a useful aspect of the design.

  3. Broad empirical support for the combined framework. The evaluation covers multiple image benchmarks, both known- and unknown-K settings, and generalized category discovery. The ablations report lower performance when score guidance or DP–diffusion coupling is removed. These results support the practical value of the components within the proposed framework.

Critical weaknesses
  1. The contribution over prior density-based methods needs clearer discussion. DIPO [1] and TAFSSL [2] refine few-shot prototypes using unlabeled neighborhoods. DIPO does so without additional model training. CMS [3] combines mean-shift refinement with contrastive learning for category discovery. HG-VAE [4] uses high- density hub samples to improve feature generation. CMS is cited and evaluated, but the added benefit of learned score guidance remains unclear. The kNN expansion ablation does not isolate this benefit over DIPO-style refinement. A matched comparison would help clarify the performance gains and additional training cost.

  2. The effect of unlabeled class imbalance needs further evaluation. Prior work shows that transductive few-shot gains can depend strongly on balanced query sets [5]. The ImageNet-50/100/200 experiments use balanced subsets. Appendix G.2 varies support counts rather than unlabeled class proportions. The Herbarium19 results are useful, but use the GCD supervision protocol. A controlled imbalance study under the main few-shot setting would help assess whether the gains persist, especially when K is unknown.

[1] Li et al. A Density-driven Iterative Prototype Optimization for Transductive Few-shot Learning (https://www.ijcai.org/proceedings/2024/0488.pdf). IJCAI, 2024.

[2] Lichtenstein et al. TAFSSL: Task-Adaptive Feature Sub-Space Learning for Few-Shot Classification (https://arxiv.org/abs/2003.06670). ECCV, 2020.

[3] Choi et al. Contrastive Mean-Shift Learning for Generalized Category Discovery (https://openaccess.thecvf.com/content/CVPR2024/papers/Choi_Contrastive_Mean-Shift_Learning_for_Generalized_Category_Discovery_CVPR_2024_paper.pdf). CVPR, 2024.

[4] Chen et al. A Novel Perspective on Leveraging Hubness in VAE for Eliminating Representative Shift Vectors in Few-Shot Learning (https://doi.org/10.1109/ICME59968.2025.11209098). ICME, 2025.

[5] Veilleux et al. Realistic Evaluation of Transductive Few-Shot Learning (https://proceedings.neurips.cc/paper_files/paper/2021/file/4d7a968bb636e25818ff2a3941db08c1-Paper.pdf). NeurIPS, 2021.

Questions for authors

Please see Critical Weaknesses.

5 comments

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following, where W denotes a weakness raised by the reviewer and R denotes our response.

W1-1. The contribution over prior density-based methods (DIPO, TAFSSL, CMS, and HG-VAE) needs clearer discussion.

R1-1. Compared with the prior density-based methods you mentioned, DD-FSC’s methodological advance lies in using cluster posteriors to guide density-based refinement and training representations against the refined targets, combining local density with inferred class membership to mitigate inter-class overlap. This distinction concerns how density information is used and how it affects representation optimization. DIPO iteratively refines prototypes toward higher-density regions without additional learnable parameters, but does not update feature representations during prototype refinement; overlap in the representation can therefore persist despite improved prototype locations. The mean-shift propagation (MSP) variant of TAFSSL refines prototypes within a task-adaptive feature subspace without further adjusting that subspace, so residual class overlap can still affect classification. HG-VAE uses high-density hub samples to improve feature generation, but hub selection relies on nearest-neighbor occurrence rather than inter-class separation; improved generation quality therefore does not directly guarantee class separability. CMS combines neighborhood mean shift with contrastive learning to update representations, but its mean-shift step relies on neighborhood aggregation, so cross-class neighbors may interfere with the refinement direction.

We have expanded the discussion of prior density-based methods and their differences from DD-FSC in Appendix E.5, highlighted in blue.

W1-2. Please add matched comparisons with DIPO-style refinement and CMS-style mean-shift to clarify performance gains and additional training cost.

R1-2. We conducted additional controlled comparisons with DIPO-style refinement and nonparametric mean-shift to assess the gains and training cost of learned score guidance. Under the main five-shot, known-KK protocol on ImageNet-100, two controls replace only semantic-anchor refinement, retaining score guidance elsewhere. A third replaces all score guidance with mean-shift and removes score-network training. All configurations share data splits, supports, initialization of shared modules, remaining objectives, training epochs, and the final predictor. Following Appendix C.3, timings use a single 12-GiB NVIDIA GeForce RTX 3080 Ti GPU and include score-network training where applicable and training-time DP updates, excluding feature caching and final inference.

ConfigurationNMI (%)ACC (%)ARI (%)Training time (min)
DIPO-style semantic anchors91.282.575.198.4
Mean-shift semantic anchors90.481.573.597.5
Mean-shift at all score locations89.478.971.2266.0
Full DD-FSC92.485.277.493.2

DD-FSC improves ACC over DIPO-style and mean-shift anchor refinement by 2.7 and 3.7 percentage points, respectively, with improvements in both NMI and ARI. Compared with replacing all score guidance with mean-shift, DD-FSC improves ACC by 6.3 points and reduces total training time by 65.0% (266.0 to 93.2 minutes). This timing includes score-network training, demonstrating improved clustering with lower training cost in the matched setting.

We have added the controlled comparisons with density-based refinement and the corresponding training-time results in Appendix E.5 (Table 22), highlighted in blue.

W2-1. Unlabeled imbalance needs controlled evaluation under the main few-shot protocol, especially with unknown KK; ImageNet subsets are balanced, and Appendix G.2 varies supports rather than unlabeled proportions.

R2-1. We added an unlabeled class imbalance study under the main five-shot protocol, covering both known-KK and unknown-KK settings. To isolate imbalance, paired runs keep image counts, supports, and training and prediction settings fixed, varying only unlabeled class proportions. Following the Dirichlet sampling principle in the work you recommended, Veilleux et al., we draw proportions from Dirichlet(21K)\mathrm{Dirichlet}(2\mathbf{1}_K). Balanced pools contain equal numbers per class; Dirichlet imbalanced pools follow sampled proportions while preserving the total count and retaining every class. DD-FSC is retrained in both settings; main-paper results serve only as a reference. Metrics are three-run means; estimated counts are mean ±\pm standard deviation.

(a) Known-KK
DatasetSettingNMI (%)ACC (%)ARI (%)
ImageNet-50Main-paper results95.593.487.3
Balanced93.891.284.1
Dirichlet imbalanced91.489.484.4
ImageNet-100Main-paper results92.485.277.4
Balanced90.783.576.1
Dirichlet imbalanced88.380.473.2
ImageNet-200Main-paper results88.783.771.4
Balanced87.081.370.9
Dirichlet imbalanced85.578.069.2
ImageNet-1KMain-paper results86.472.451.2
Balanced85.970.448.3
Dirichlet imbalanced83.568.146.1
(b) Unknown-KK
DatasetSettingNMI (%)ACC (%)ARI (%)K^\widehat K
ImageNet-50Main-paper results82.976.066.851.7±0.951.7\pm0.9
Balanced81.273.964.752.3±3.252.3\pm3.2
Dirichlet imbalanced78.971.762.353.7±2.353.7\pm2.3
ImageNet-100Main-paper results69.756.244.7103.0±1.4103.0\pm1.4
Balanced67.754.342.5104.0±2.6104.0\pm2.6
Dirichlet imbalanced65.451.639.4108.0±3.0108.0\pm3.0
ImageNet-200Main-paper results63.646.735.3198.7±1.9198.7\pm1.9
Balanced62.145.833.2193.7±3.2193.7\pm3.2
Dirichlet imbalanced59.741.530.1186.0±2.6186.0\pm2.6
ImageNet-1KMain-paper results60.536.628.11049.0±10.21049.0\pm10.2
Balanced57.434.726.81055.3±3.81055.3\pm3.8
Dirichlet imbalanced56.731.323.31066.7±4.01066.7\pm4.0

Compared with Balanced, Dirichlet imbalanced reduces ACC by 1.8–3.3 percentage points with known KK and 2.2–4.3 points with unknown KK. Mean estimated counts remain within 8% of the true count on all four imbalanced datasets. These results show that DD-FSC remains effective beyond balanced unlabeled sets, with limited accuracy declines and near-correct category counts under the tested imbalance.

W2-2. The Herbarium19 results are useful, but use GCD supervision rather than the main few-shot protocol.

R2-2. We also added unknown-KK few-shot experiments on Herbarium19 in Appendix D.4, using its naturally unequal class sizes. This tests performance with both abundant and scarce categories. The table combines clustering metrics and estimated counts from one seed; the true count is 683.

MethodNMI (%)ACC (%)ARI (%)K^\widehat K
Few-shot methods adapted with DP-based cluster-count estimation\rlap{\textit{Few-shot methods adapted with DP-based cluster-count estimation}}
Soft kk-Means (ICLR, 2018)45.427.927.9605
cluster-FSL (CVPR, 2022)50.032.331.8605
UNEM (CVPR, 2025)48.735.331.7605
EfficientFSL (AAAI, 2026)59.041.633.9605
TIM++ (AAAI, 2026)57.240.934.1605
OSESSL (ICLR, 2026)59.142.536.9605
Clustering methods adapted with few-shot supervision\rlap{\textit{Clustering methods adapted with few-shot supervision}}
DeepDPM (CVPR, 2022)44.538.040.1775
UNSEEN+DCN (ICDMW, 2024)51.136.042.3782
SpherePair (NeurIPS, 2025)55.242.440.0635
SCMax (AAAI, 2026)54.743.340.6720
DeepDP-TGMM (CVPR, 2026)57.644.640.8715
DD-FSC (Ours)61.248.240.2711

DD-FSC achieves the highest NMI and ACC and the smallest cluster-count error on Herbarium19. Its ACC exceeds the strongest baseline by 3.6 percentage points; its estimated count is 711, an absolute error of 28. This comparison shows that its gains over baselines extend beyond balanced data to naturally unequal class sizes, unknown category counts, and sparse labels, yielding more accurate overall grouping and count estimation.

We have added the controlled unlabeled class imbalance study in Appendix G.3 (Tables 29 and 30) and the Herbarium19 unknown-KK few-shot evaluation in Appendix D.4 (Tables 16 and 17), highlighted in blue.

Reviewer_ikMzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work proposes DD-FSC that helps to perform few-shot clustering where labelled examples cover only some of the classes and the total number of classes may be unknown. The method employs frozen OpenCLIP features and a trainable adapter, and uses diffusion-based density modelling. To compute the clusters, it uses a Dirichlet process. Score-guided trajectories construct category anchors for selecting unlabelled examples and provide targets for feature refinement. The inferred cluster structure guides these trajectories and provides prototype-level contrastive supervision, while posterior component weights determine the active cluster count. Validation includes ten image-clustering benchmarks and five generalized category-discovery benchmarks under known and unknown cluster count settings.

Critical strengths

The problem is useful and well established, since labeling data is costly in many real applications. The method helps cluster the data using only a few labeled samples from a few classes, and it also handles the case where the number of clusters is unknown. The central idea of using feature density to choose reliable unlabeled samples is easy to follow. The vast set of experiments is really commendable. The hyperparameter analysis is thorough.

Critical weaknesses

My major concern is that the novelty is very limited, and the problem itself is already well established in the community. Most of the components are well known: the supervised contrastive loss, prototype contrastive learning, confidence-based pseudo-labeling, score-based denoising of features, and Dirichlet process mixtures for estimating the number of clusters. There is also no novelty in the way they are combined, as each component is simply added as a separate loss or step.

It gives an impression that true number of clusters is used in the unknown-K setting. The maximum number of clusters and the threshold for keeping a cluster are both set from the true K (Appendix C.2), even though the paper says K is withheld. So the claim of accurate cluster-count estimation is not convincing. The estimated K is also sensitive to these settings, and its improvement is not statistically significant (Appendix F).

The paper’s central claim is never directly tested. Its main promise is to spread the few labels from seen classes so that the unseen classes are also recovered, and Figure 1 even says “Novel Cluster Found.” Yet the main tables report only overall scores, with no split between labeled and unlabeled classes. The method could do well on the 30% of labeled classes and poorly on the other 70% without the reader knowing.

It is not clear where the gains come from. In the known-K setting, the final clusters come from k-means with the labeled points anchored, not from the Dirichlet process. In the unknown-K setting, accuracy drops sharply (85.2 to 56.2 on ImageNet-100) even though the estimated K is close to the true value. This suggests the DP clustering itself is weak.

The comparisons with other methods are not fair. In the unknown-K tables, DD-FSC uses labels while DeepDPM, SCMax and DeepDP-TGMM are fully unsupervised, yet DD-FSC still loses on several NMI and ARI entries. The paper does not say which backbone each baseline used or which numbers were copied from other papers rather than rerun, so baselines may differ in class subsets, splits and pretrained features. The few-shot baselines, such as TIM++, UNEM and EfficientFSL, are few-shot classification methods that expect labels for every class, and the paper never explains how they were made to cluster classes with no labels.

Theorem 1 does not cover the actual update and gives no guarantee about class correctness.

Questions for authors
  • The reliability of semantic expansion is not directly established. Candidate confidence is calculated only relative to observed-category anchors. An image from an unseen category could therefore receive a high score for an observed category simply because that category is its best available match. Such an image could then be added to the wrong semantic set and reinforce an incorrect grouping. Confidence filtering and the candidate budget may reduce this risk, but the paper should measure how many selected candidates are correct and how often unseen-category images are mistakenly included. Separate results for observed and unobserved categories under the main five-shot protocol would also help.

  • It is unclear where the gains come from, and the Dirichlet process part looks weak. In the known-K setting the final clusters seem not to be produced by the DP but by a k-means variant in which the labeled points are anchored.

  • DGCL is called a contrastive loss but is just a squared distance with no negatives. Its weight is tiny, yet removing it causes the largest accuracy drop, which needs explaining.

  • What is the accuracy on labeled (seen) vs. unlabeled (novel) classes in Tables 1–2?

  • How were all defaults and dataset-specific settings chosen without labeled validation data? Was any setting picked by looking at NMI, ACC or ARI?

  • How were TIM++, UNEM, EfficientFSL, cluster-FSL and Soft k-Means adapted to produce clusters for classes with no labels?

  • Please report the results for raw OpenCLIP + k-means and raw OpenCLIP + support-anchored k-means? This would highlight how much the proposed method adds beyond the pretrained features and the anchored k-means step

  • After the score-guided steps, how often does a labeled point move closer to another class than to its own? Since the steps move points toward denser areas, could a labeled point be pulled into a nearby class that is more crowded?

Suggestion:

The method uses OpenCLIP but only its image encoder, ignoring the text side. A simple baseline that uses CLIP text embeddings of the observed class names as anchors, plus k-means for the remaining data, can be included. It would show whether the proposed diffusion and DP components add value beyond what the vision-language model already provides. Text from a generic vocabulary could also help reject images from unseen classes during semantic expansion.

The paper does not include a limitations or future-work section.

9 comments

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following, where W, Q, and S denote weaknesses, questions, and suggestions, respectively, and R denotes our response.

Responses to Weaknesses

W1. Novelty is very limited: the problem and components are well established, and their combination simply adds separate losses or steps.

R1. Our methodological contribution addresses a central difficulty in few-shot clustering: learning representations for categories without labeled anchors while the underlying cluster structure is itself uncertain. Applying existing techniques directly leaves this difficulty unresolved: semantic anchors supervise only supported categories, density scores alone do not determine semantic cluster membership, and DP inference on fixed features cannot improve the representations on which its partition depends. DD-FSC addresses these limitations through DP-guided density trajectories (Equations (5)–(6)), where the density score captures local feature geometry and the DP posterior guides trajectories according to inferred cluster membership. The resulting sample and prototype endpoints provide learning targets for the encoder, including for unsupported categories, while updated features refine the posterior. Thus, DP inference changes the targets used by the contrastive objectives, rather than merely supplying cluster counts or contributing a separate loss.

Our ablations isolate the mechanism that addresses this challenge. On ImageNet-100 under unknown-KK, removing DP guidance from the trajectories while retaining score diffusion, DP inference, and the remaining objectives reduces ACC from 56.2% to 47.8% and ARI from 44.7% to 36.6% (Table 4(b)). Moreover, learning features first and fitting the same DP model to frozen representations achieves only 42.2% ACC, compared with 56.2% for DD-FSC. These results demonstrate the benefit of jointly refining learning targets and inferred cluster structure, beyond retaining the individual components or arranging them in a two-stage pipeline.

W2. Unconvincing unknown-KK estimation: capacity and cluster-retention threshold use true KK despite claiming it is withheld (Appendix C.2). Estimates are sensitive to these settings, and improvements are not statistically significant (Appendix F).

R2. The ground-truth count is used only to configure the candidate-component capacity and activation threshold, but agreement between the estimated and true counts is not used as feedback to refine clustering. Accurate count estimation still requires learning from images and cannot be predetermined by these settings. We set Kmax⁡=2KrefK_{\max}=2K_{\mathrm{ref}} to scale the candidate capacity with the category space and leave additional components for learning and selection. The model learns posterior mixing weights from images and counts components exceeding the activation threshold.

Figure 9 shows that clustering performance does not depend on one specific capacity ceiling. On ImageNet-100, varying Kmax⁡∈100,150,200,300,400K_{\max}\in\\{100,150,200,300,400\\} at a fixed activation threshold yields estimated counts of 95.3–107.3, with ACC, NMI, and ARI varying by only 0.9, 0.5, and 0.8 percentage points, respectively. Thus, both smaller and larger capacities allow recovery of counts close to the true value while maintaining stable clustering performance across the tested range.

Furthermore, the ten-run analysis in Appendix F reports lower mean absolute cluster-count errors for DD-FSC than DeepDP-TGMM on both ImageNet-100 and ImageNet-200, although these improvements do not reach statistical significance after Holm correction. Further improving cluster-count estimation remains a shared challenge: recent methods including DebGCD (ICLR 2025), FAIR (AAAI 2026), and BIA (ICLR 2026) report count estimates with substantial errors on some datasets [1–3].

[1] Liu and Han. DebGCD: Debiased Learning with Distribution Guidance for Generalized Category Discovery (https://proceedings.iclr.cc/paper_files/paper/2025/file/789fc893d4b6e5b384f4eb058a35e31d-Paper-Conference.pdf). ICLR, 2025.

[2] Bian et al. Foundation-Adaptive Integrated Refinement for Generalized Category Discovery (https://ojs.aaai.org/index.php/AAAI/article/download/37231/41193). AAAI, 2026.

[3] Tang et al. Bures-Isotropy Alignment: Manifold Learning of Generalized Category Discovery (https://proceedings.iclr.cc/paper_files/paper/2026/file/5b91cefbfa52340af2f16f249572dc76-Paper-Conference.pdf). ICLR, 2026.

W3. Unseen-class recovery is not directly tested: overall scores may hide poor performance on the 70% unlabeled classes without a labeled/unlabeled breakdown.

R3. We have supplemented the main-table results with separate evaluations of observed and unobserved categories to directly assess recovery of the remaining 70% of categories. The table below reports NMI, ACC, and ARI for All, Observed, and Unobserved categories on all four ImageNet benchmarks under both known-KK and unknown-KK settings. Observed denotes the 30% of categories providing five labeled supports each; Unobserved denotes the remaining 70% providing none. All metrics exclude support images, ensuring that the breakdown measures clustering performance on unlabeled images.

DatasetEvaluation scopeKnown-KKUnknown-KK
NMI (%)ACC (%)ARI (%)NMI (%)ACC (%)ARI (%)
ImageNet-50All95.593.487.382.976.066.8
Observed94.789.890.281.087.971.9
Unobserved96.994.992.282.570.970.8
ImageNet-100All92.485.277.469.756.244.7
Observed93.389.986.273.979.367.0
Unobserved89.983.278.069.246.346.6
ImageNet-200All88.783.771.463.646.735.3
Observed89.890.079.371.161.676.7
Unobserved88.281.074.362.540.334.3
ImageNet-1KAll86.472.451.260.536.628.1
Observed87.568.162.366.945.440.0
Unobserved86.874.256.259.832.826.1

The breakdown shows that DD-FSC's overall performance is not confined to categories with labeled supports: unobserved categories are also recovered. Under known-KK, unobserved-category ACC ranges from 74.2% to 94.9% across the four benchmarks and exceeds observed-category ACC on two benchmarks; unobserved-category NMI and ARI range from 86.8% to 96.9% and from 56.2% to 92.2%, respectively. Under the more challenging unknown-KK setting, unobserved-category ACC reaches 70.9%, 46.3%, 40.3%, and 32.8% on ImageNet-50, ImageNet-100, ImageNet-200, and ImageNet-1K, respectively. These results directly support our central claim: few labels from observed categories can guide the recovery of both observed and unobserved categories.

We have added the separate evaluations of observed and unobserved categories under known-KK and unknown-KK in Appendix D.2 (Table 13), highlighted in blue.

W4-1. The source of gains is unclear: known-KK clusters come from support-anchored kk-means, not DP.

R4-1. Under known-KK, the gains arise from Dirichlet process (DP) posterior guidance in density-guided sample alignment and prototype learning, while final clustering is indeed performed by support-anchored kk-means. DP posteriors construct prototypes and guide sample and prototype trajectories, whose endpoints provide encoder learning targets. Keeping the final anchored kk-means unchanged, removing DGCL or PCL reduces ImageNet-100 ACC from 85.2% to 76.9% or 80.3%, respectively (Table 4(a)). Thus, these known-KK gains reflect representation learning guided by DP, rather than simply the use of anchored kk-means.

W4-2. Unknown-KK ACC drops sharply (85.2% to 56.2% on ImageNet-100) despite near-correct KK, suggesting weak DP clustering.

R4-2. On ImageNet-100, the ACC decrease from 85.2% with known KK to 56.2% with unknown KK reflects the difficulty of jointly inferring cluster counts and sample memberships under sparse supervision: a near-correct count does not guarantee correct assignments. Only 30 of the 100 categories have labeled supports, with five images each, while the remaining 70 categories are entirely unlabeled. Erroneous splits and merges can offset numerically, producing a near-correct count despite incorrect sample memberships. Moreover, agreement between the estimated and true counts is not used as feedback to refine clustering. Controlled comparisons within the same unknown-KK setting show that the assignment quality of the same Dirichlet process (DP) model depends on how representations are learned. Learning features first and then fitting this DP model yields only 42.2% ACC, versus 56.2% for joint DD-FSC. Retaining DP inference and the remaining objectives while removing trajectory-level DP guidance also lowers ACC to 47.8% (Table 4(b)). These results show that a near-correct count does not remove the difficulty of inferring sample memberships, while joint learning substantially improves DP's assignment quality.

W5-1. Unfair comparisons: DD-FSC uses labels while DeepDPM, SCMax, and DeepDP-TGMM are fully unsupervised, yet DD-FSC still loses on several unknown-KK NMI and ARI entries.

R5-1. Regarding supervision fairness, DeepDPM, SCMax, and DeepDP-TGMM in Tables 2–3 are adapted to receive exactly the same few-shot supervision as DD-FSC. Thus, these tables compare methods under matched supervision rather than comparing our few-shot method against fully unsupervised baselines. Under this protocol, DD-FSC ranks first in nine of twelve NMI/ACC/ARI comparisons in Table 2 and first or second in cluster-count estimation accuracy on every dataset in Table 3, achieving the strongest overall performance despite not leading every entry. We also report a separate fully unsupervised comparison in Appendix D.3, where DD-FSC-U ranks first in 23 of 24 metric–dataset comparisons in Table 14.

W5-2. Baseline backbones and copied versus rerun results are unspecified; class subsets, splits, and pretrained features may differ. How do TIM++, UNEM, and EfficientFSL, which expect labels for every class, handle classes without labels?

R5-2. Regarding backbone, result provenance, and baseline adaptations, Appendix C.4 now provides the corresponding details. All baseline results in Tables 2–3 were rerun using the same frozen OpenCLIP ViT-B/16 features, category subsets, data splits, support samples, and evaluation images as DD-FSC. For TIM++, UNEM, and EfficientFSL, support-anchored components represent supported classes, while additional components initialized from unlabeled features accommodate classes without labeled examples; their unknown-KK versions use a DP mixture to infer the component structure. These shared protocols align the representation and evaluation conditions across methods. We have also clarified result provenance in Tables 1–3, with all additions and revisions highlighted in blue.

W6. Theorem 1 does not cover the actual update and gives no guarantee about class correctness.

R6. We have extended Theorem 1 to cover the complete stochastic DP–diffusion update and added a conditional class-preservation guarantee. The revised theorem bounds the expected change in the Gaussian-smoothed log-density, explicitly accounting for score-estimation error, normalized DP guidance, and Gaussian exploration noise. It also provides a sufficient condition for expected density ascent, with the derivation given in Appendix B.1.2.

Proposition 1 addresses class correctness under explicit class-separation and score-error conditions. It derives a high-probability condition for a finite annealed trajectory to preserve its starting class by tracking the class margin through the actual updates. The analysis includes initialization noise, subsequent stochastic perturbations, and DP guidance, and proves that endpoint normalization preserves the resulting positive class margins. A further consequence establishes class preservation for category anchors formed by averaging the support endpoints and normalizing their mean, as detailed in Appendix B.1.3.

We have extended Theorem 1 and added theoretical analyses to Appendix B.1.2–B.1.3. All these additions and revisions are highlighted in blue.

Reviewer_SouTOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DD-FSC, a nonparametric few-shot clustering framework that jointly models feature-space density and latent cluster structure. DD-FSC characterizes the feature distribution with a diffusion model and employs score-guided Langevin dynamics to propagate limited supervision toward high-density semantic regions. It identifies high-confidence same-category samples and constructs density-aware positive pairs. Experiments cover known- and unknown-cluster-number settings, together with generalized category discovery.

Critical strengths

The paper addresses an interesting clustering problem where labels cover only some categories and the total category count is unknown. The experiments are extensive, including component ablations and sensitivity analyses, and show promising results.

Critical weaknesses
  1. The performance of clustering with an unknown number of clusters drops substantially. For ImageNet-1K, accuracy decreases from 72.4% to 36.6%, making the results less convincing as evidence of meaningful clustering. More importantly, this occurs despite a relatively accurate estimated category count of 1049 versus 1000. This discrepancy raises concerns about the results and requires further explanation.

  2. The unknown-K setting needs stronger validation. I could not find how the maximum number of clusters is determined in the main text. In Appendix C.2, I found that sets Kmax to twice Kref, the ground-truth category count used for evaluation. Thus, the configuration still uses information about the true count. This is problematic because the true category count would be unavailable in the intended real-world setting.

  3. The methodological novelty appears limited. Few-shot clustering is an established problem, and the framework combines existing techniques, including Dirichlet process mixtures, semantic anchoring, diffusion-based density modeling, and contrastive learning. The potentially novel contribution lies in their coupling, but it remains unclear what substantive methodological advance this provides beyond integrating these components.

  4. The confidence-based selection rule may admit samples from unknown categories into observed ones. Equation (2) normalizes assignment scores only over labeled categories, so a high score does not necessarily indicate that a sample belongs to any labeled category. The authors should report the proportion of incorrectly admitted unknown category samples and explain how this affects subsequent representation learning.

  5. The literature review appears incomplete. Although the paper respectively discusses deep clustering and clustering with an unknown number of clusters, some important work about deep clustering with an unknown number of clusters appears to be missing. Since the method falls within semi-supervised clustering, the discussion of related semi-supervised and few-shot clustering methods also needs expansion. The authors should review this literature more comprehensively, including recent developments.

  6. The practical motivation could be clearer. Although the paper considers limited labels and an unknown category count, it provides few concrete application scenarios explaining why these conditions arise together. The evaluation mainly focuses on standard image benchmarks, leaving the practical benefits insufficiently demonstrated.

Questions for authors

Please refer to the questions raised in the Weaknesses section.

5 comments

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following, where W denotes a weakness raised by the reviewer and R denotes our response.

W1. Questionable unknown-KK clustering: ImageNet-1K ACC drops from 72.4% to 36.6% despite near-correct KK (1049 versus 1000).

R1. ImageNet-1K is challenging under unknown-KK clustering because the model must jointly infer the cluster count and sample memberships with sparse supervision: only 30% of its 1,000 categories have labeled supports, with five images each, while the remaining 70% are entirely unlabeled. Erroneous class splits and merges can offset numerically, producing a near-correct count despite inaccurate assignments. The consistent performance declines across state-of-the-art methods further underscore this challenge: the six few-shot baselines in Tables 1–2 lose 21.5–31.8 percentage points in NMI, 35.0–43.9 in ACC, and 25.2–33.7 in ARI when moving from their known-KK configurations to their DP-adapted unknown-KK versions. DD-FSC nevertheless achieves the strongest overall results in Table 2, ranking first in nine of twelve metric–dataset comparisons.

DD-FSC improves category recovery beyond count estimation. In Table 20(b), full DD-FSC and its variant without DGCL estimate similar cluster counts on ImageNet-1K (1,049.0 vs. 1,058.7), yet DGCL improves ACC by 5.9 and ARI by 4.9 percentage points. This demonstrates that density-guided feature alignment helps recover more accurate sample memberships even when both count estimates are close to the ground truth.

W2. Insufficient unknown-KK validation: the main text omits the capacity setting, while Appendix C.2 uses Kmax⁡=2KrefK_{\max}=2K_{\mathrm{ref}}, requiring unavailable ground-truth counts in practice.

R2. The ground-truth count is used only to configure the candidate-component capacity and activation threshold, but agreement between the estimated and true counts is not used as feedback to refine clustering. Accurate count estimation still requires learning from images and cannot be predetermined by these settings. We set Kmax⁡=2KrefK_{\max}=2K_{\mathrm{ref}} to scale the candidate capacity with the category space and leave additional components for learning and selection. The model learns posterior mixing weights from images and counts components exceeding the activation threshold.

Figure 9 shows that clustering performance does not depend on one specific capacity ceiling. On ImageNet-100, varying Kmax⁡∈100,150,200,300,400K_{\max}\in\\{100,150,200,300,400\\} at a fixed activation threshold yields estimated counts of 95.3–107.3, with ACC, NMI, and ARI varying by only 0.9, 0.5, and 0.8 percentage points, respectively. Thus, both smaller and larger capacities allow recovery of counts close to the true value while maintaining stable clustering performance across the tested range. Furthermore, the ten-run analysis in Appendix F reports lower mean absolute cluster-count errors for DD-FSC than DeepDP-TGMM on both ImageNet-100 and ImageNet-200.

For real-world applications where the true category count is unavailable, Appendix B.3 develops a theoretical extension that configures both capacity and activation threshold without this information. It selects capacity by balancing evidence from unlabeled images with a capacity prior, and sets the activation threshold from the selected capacity.

We have added the theoretical extension to Appendix B.3, highlighted in blue.

W3. Limited methodological novelty: the substantive advance from coupling existing components for few-shot clustering remains unclear.

R3. Our methodological contribution addresses a central difficulty in few-shot clustering: learning representations for categories without labeled anchors while the underlying cluster structure is itself uncertain. Applying existing techniques directly leaves this difficulty unresolved: semantic anchors supervise only supported categories, density scores alone do not determine semantic cluster membership, and DP inference on fixed features cannot improve the representations on which its partition depends. DD-FSC addresses these limitations through DP-guided density trajectories (Equations (5)–(6)), where the density score captures local feature geometry and the DP posterior guides trajectories according to inferred cluster membership. The resulting sample and prototype endpoints provide learning targets for the encoder, including for unsupported categories, while updated features refine the posterior. This mechanism resolves the disconnect between latent cluster discovery and representation learning by allowing each to inform the other during optimization.

Our ablations isolate the mechanism that addresses this challenge. On ImageNet-100 under unknown-KK, removing DP guidance from the trajectories while retaining score diffusion, DP inference, and the remaining objectives reduces ACC from 56.2% to 47.8% and ARI from 44.7% to 36.6% (Table 4(b)). Moreover, learning features first and fitting the same DP model to frozen representations achieves only 42.2% ACC, compared with 56.2% for DD-FSC. These results demonstrate the benefit of jointly refining learning targets and inferred cluster structure, beyond retaining the individual components or arranging them in a two-stage pipeline.

W4-1. Unquantified unknown-category admissions: confidence normalized only over observed categories (Equation (2)) does not establish observed-class membership.

R4-1. We directly measured candidate reliability in a single-seed ImageNet-100 experiment under the main five-shot, known-KK protocol. The table below reports candidate precision and the proportion of accepted candidates originating from unobserved categories.

EpochSelection ruleAcceptedPrecision (%)Unobserved intrusion (%)
45Threshold + top-PP16091.885.00
45Top-PP only24081.6712.50
225Threshold + top-PP20096.002.50
225Top-PP only24090.836.25
450Threshold + top-PP22097.731.36
450Top-PP only24094.583.33

At epoch 450, thresholding reduces unobserved-category intrusion from 3.33% to 1.36%, while retaining 220 candidates with 97.73% precision. Thus, filtering reduces erroneous admissions while retaining most of the candidate budget.

W4-2. Unclear effects of erroneous admissions on subsequent representation learning.

R4-2. To examine their effect on subsequent learning, we compared the original candidate schedule with targeted removal of unobserved-category candidates from semantic sets and random removal matched in number per category and epoch. All three configurations replay the same recorded candidate schedule. Targeted removal uses ground-truth labels for diagnosis; all images remain in the training pool. Besides ACC, we report the same-class proportion among each unobserved sample's ten nearest neighbors in the learned feature space (10-NN).

ConfigurationAll ACC (%)Observed ACC (%)Unobserved ACC (%)Unobserved 10-NN (%)
Original replay85.3691.2082.8786.01
Targeted removal (diagnostic)86.0791.6783.6986.61
Matched random removal84.9490.8682.4185.70

Targeted removal improves unobserved-category ACC and 10-NN over the original replay by 0.82 and 0.60 percentage points, respectively, and outperforms matched random removal. These results indicate that the remaining erroneous admissions have a limited effect on unobserved-category representations and clustering in this run.

We have added the candidate-reliability measurements and matched removal experiments in Appendix G.4 (Tables 31 and 32), highlighted in blue.

W5. Incomplete coverage of unknown-KK deep clustering, semi-supervised and few-shot clustering, and recent developments.

R5. We have expanded the Related Work to cover unknown-KK deep clustering and semi-supervised and few-shot clustering, including recent developments. For unknown-KK clustering, we added DipDECK and DeepSynC (2026), and clarified that these methods' original unsupervised formulations do not exploit sparse category labels to resolve semantic ambiguity. For limited supervision, we added DC-GMM and SimGCD and expanded the discussion of SpherePair, cluster-FSL, OSESSL (2026), GCD, and FSNCD (2025). The revised discussion explicitly identifies limitations concerning prescribed cluster counts, labeled category coverage, and the separation of representation learning from clustering.

We have expanded and revised the Related Work, highlighted in blue.

Comments on the paper

1 comment

Dear Area Chair,

Thank you for overseeing the discussion. Reviewers ikMz and KEGU have both expressed positive assessments of our work. We have not yet received a follow-up from Reviewer SouT. As the discussion closes, we highlight the key evidence addressing the reviewers’ concerns. All references below refer to the revised manuscript.

Main concerns shared across the reviews

  • Methodological contribution and category recovery [SouT W1, W3; ikMz W1, W3–W5, Q2, Q4, Q6; KEGU W1-2]. DP-guided trajectories couple representation learning with cluster inference. On ImageNet-100 under unknown-KK, DD-FSC achieves 56.2% ACC, versus 47.8% without trajectory-level DP guidance and 42.2% for a two-stage pipeline. Under matched supervision, features, and protocols, it ranks first in 9/12 unknown-KK metric–dataset comparisons. We also report observed/unobserved-category results, document baseline adaptations and result provenance, and isolate representation-learning gains with the final predictor held fixed. On ImageNet-1K, density-guided consistency improves ACC/ARI by 5.9/4.9 percentage points despite similar estimated counts, demonstrating better assignments beyond count recovery. Known-KK comparisons further show 2.7/3.7-point ACC gains over DIPO-style/mean-shift anchor refinement, with lower training time (Tables 2 and 4; Appendices C.4, D.2–D.3, E.3, and E.5).

  • Reliability and practical applicability [SouT W4, W6; ikMz Q1; KEGU W2]. Semantic-expansion filtering reduces unknown-category intrusion from 3.33% to 1.36%, with 97.73% candidate precision, at the final checkpoint of a single-seed ImageNet-100 known-KK audit; matched removal controls assess effects on both observed and unobserved categories. To support the ecological-monitoring motivation, we extended Herbarium19 evaluation from GCD to partial-category five-shot learning and added IP102 and iWildCam experiments to assess practical benefits. We also conducted controlled imbalance tests under known-KK and unknown-KK (Appendices D.4, G.3, and G.4).

Additional concerns raised by Reviewer SouT

  • Capacity selection [W2]. We clarified benchmark settings, added sensitivity analyses, and developed a theoretical extension for selecting capacity and threshold without the true category count (Appendix B.3; Figure 9).

  • Literature coverage [W5]. We expanded related work on unknown-KK deep clustering, semi-supervised clustering, and few-shot clustering, including recent developments (Section 2).

Thank you for your time and consideration. We would be happy to clarify any remaining questions.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.