acceptodds
Under review as a conference paper at ICLR 2027

CoClusterPFN: Amortized Posterior Similarity for Tabular Bayesian Clustering

15 views

Abstract

Most clustering methods return one grouping without revealing which assignments are uncertain. Bayesian methods represent this uncertainty but typically require expensive inference. We introduce CoClusterPFN, a Prior-data Fitted Network (PFN) for Bayesian clustering of tabular data. It is trained on synthetic clustering tasks to predict pairwise posterior co-clustering probabilities in a single forward pass. The resulting posterior similarity matrix (PSM) avoids the label-switching ambiguity of cluster assignments. When the number of clusters is unspecified, the prediction marginalizes over cluster-count uncertainty. The same PSM provides an interpretable summary and can be decoded into hard partitions using decision rules targeting different metrics or selecting different numbers of clusters. Under a Gaussian-mixture prior, CoClusterPFN closely matches PSMs estimated by Markov chain Monte Carlo (MCMC) while being orders of magnitude faster. On 60 CLUBench datasets, CoClusterPFN achieves the best average rank for adjusted Rand index (ARI), normalized mutual information (NMI), and Binder loss, while ranking second for variation of information (VI), whether the cluster count is supplied or unknown. Beyond benchmark performance, an Iris case study demonstrates how CoClusterPFN can help users choose among alternative clusterings. Together, these results support CoClusterPFN as an effective approach to uncertainty-aware clustering.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Bayesian Clustering, Prior-data Fitted Networks, Amortized Clustering, Transformer, Posterior Similarity Matrix, Tabular Data

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hAyQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoClusterPFN, a Prior-data Fitted Network that, in one forward pass, predicts the posterior similarity matrix (PSM) of a tabular dataset, i.e., the probability that each pair of observations shares a cluster, with the number of clusters KK either supplied or marginalized out. Predicting pairwise probabilities avoids label switching, and the same PSM can be decoded into hard partitions with different decision rules (MaxPEAR, MinVI-LB, Binder). On 2D Gaussian mixtures, a version retrained for that prior closely matches MCMC PSMs while being orders of magnitude faster. Directly predicting the PSM gives higher decoded ARI than predicting cluster responsibilities, and on 60 CLUBench datasets it has the best average rank for ARI, NMI and Binder loss and second for VI, with or without KK. An Iris case study illustrates how the PSM exposes uncertain pairs.

Critical strengths
  1. Principled, label-invariant target with supporting evidence. The PSM is invariant to cluster relabeling, and the population minimizer of its binary cross-entropy loss is the posterior co-clustering probability, so one model supports several decision rules and both known and unknown KK.
  2. Careful and transparent evaluation. The prior and decoders were fixed without using CLUBench, and gains are tested with paired bootstrap intervals and Holm-corrected Wilcoxon tests. The paper also openly reports where the method falls short, e.g., poor cluster-count recovery, transitivity violations and quadratic cost.
Critical weaknesses
  1. Limited novelty. Pairwise same-cluster prediction with attention and BCE already exists in ABC (Coward et al., 2020), "a close architectural predecessor". The paper itself describes the method as combining "direct pairwise prediction with the PFN framework", on top of a pretrained TabICL encoder. Neither ABC nor the closest PFN, ClusterPFN, is compared. Although the method is motivated as a fast alternative to Bayesian clustering, no Bayesian clustering method is a baseline.
  2. Posterior fidelity is checked only on toy data. The MCMC agreement is for a separate model retrained on 2D Gaussian mixtures with N=48N = 48. The main tabular model is evaluated only against "sampled binary co-membership labels rather than exact posterior probabilities". On CLUBench, 18.23% of triples violate PSM transitivity with KK supplied, versus 1.29% on the training prior.
  3. Quadratic cost limits it to small tables. Attention, the pairwise head and the PSM all scale as O(N2)O(N^2), and the authors state "it remains unclear how to move beyond O(N2)O(N^2) originating from the PSM". Already at 1,001≤N≤2,0001{,}001 \le N \le 2{,}000, inference is 7.67× slower than TabClustPFN, and the model is trained only up to N=2,048N = 2{,}048, D=20D = 20 and K=10K = 10, with KK capped by design.
  4. Narrow benchmark close to the training prior. CLUBench is reduced to the 60 of 103 datasets "matching the cluster-count and dimensionality ranges of the training prior", and 19 are subsampled to 2,000 observations. Of these, 28 have K=2K = 2, and the reference "clusters" are class labels, which need not reflect cluster structure. So, no method exceeds a mean ARI of 0.39.
Questions for authors
  1. Can you compare against ABC, ClusterPFN and a Bayesian clustering baseline (e.g., a Bayesian GMM or Dirichlet-process mixture fit by MCMC or variational inference) on CLUBench, in both accuracy and runtime, and does the gain over the Responsibility-PFN hold across training seeds?
  2. Can you validate the primary tabular model against a posterior reference, e.g., MCMC on its clean Gaussian-mixture branch or on low-dimensional real datasets, and explain why transitivity violations rise to 18.23% on CLUBench with KK supplied?
  3. For scalability concerns, can you report runtime and memory for NN well beyond 2,000 (e.g., 10410^4–10510^5), and discuss possible approximations that avoid O(N2)O(N^2) cost?
  4. How does CoClusterPFN perform on the 43 excluded datasets, on full datasets without subsampling, and on benchmarks where labels are known to reflect cluster structure?
5 comments
AuthorsOpenReview ↗

Response to Reviewer hAyQ (1/2)

Regarding novelty and comparisons with ClusterPFN and ABC, please see the general comment.

Response to Q1 — Additional baselines and training-seed robustness

We thank the reviewer for suggesting these experiments.

Additional baselines

Can you compare against ABC, ClusterPFN and a Bayesian clustering baseline (e.g., a Bayesian GMM or Dirichlet-process mixture fit by MCMC or variational inference) on CLUBench, in both accuracy and runtime? The requested comparisons with ABC, ClusterPFN, and variational Bayesian mixtures are discussed in our General Comments. We have also added runtime comparisons in Appendix C.7 of the revised manuscript, covering neural inference, PSM decoding, and complete clustering pipelines where applicable. In the runtime comparisons, ABC+TabICL's matrix inference was approximately 13.4 times faster than CoClusterPFN's on an NVIDIA A40. On the nine compatible CLUBench datasets, ClusterPFN's complete CPU pipeline was approximately 71/37 times faster with supplied/unspecified KK, respectively.

Training-seed robustness

... and does the gain over the Responsibility-PFN hold across training seeds?

We extended the Responsibility-PFN comparison in Section 4.2 from one to three training seeds, matching the prior, training schedule and evaluation protocol. Direct PSM prediction achieved higher decoded ARI in every seed under all three decision rules, on both CLUBench and matched-prior synthetic tasks.

Supplied-KK CLUBench MaxPEAR ARI:

SeedDirect PSMResponsibility-PFNDifference
10.38380.3324+0.0514
20.36630.3407+0.0256
30.38200.3427+0.0393
Mean0.37740.3386+0.0388
Sample SD across seeds0.00960.00540.0129

The paired dataset-bootstrap 95% interval for the seed-averaged improvement is [0.0139, 0.0654].

All six models were evaluated in the same CPU FP32 environment. The revised Appendix C.2 documents the full results and differences from the original GPU evaluation.

Response to Q2 — Posterior reference and transitivity violations

Posterior reference

Can you validate the primary tabular model against a posterior reference, e.g., MCMC on its clean Gaussian-mixture branch or on low-dimensional real datasets?

A matched MCMC reference for the primary tabular model is difficult to obtain because its training prior combines several generative branches and latent mechanisms. Running MCMC only on the clean Gaussian branch would not provide a direct comparison with our primary model, which was trained on the full prior rather than that branch alone. Similarly, fitting a conventional Bayesian GMM to low-dimensional real data would provide a posterior for that GMM, rather than for our full tabular prior. Such comparisons would not directly validate the posterior induced by our full training prior.

This motivated our matched-prior NIW–GMM experiment (Section 4.1 and Appendix F), where we can directly compare CoClusterPFN's predicted probabilities with those estimated by MCMC. Against matched MCMC references, the model achieves PSM MAE of 0.0187/0.0302 and correlations of 0.980/0.971, with KK known/unknown.

These results validate the approach under the matched NIW–GMM prior. They do not establish posterior fidelity for the primary heterogeneous tabular model. We have explicitly acknowledged this limitation in Section 5 of the revised manuscript.

Transitivity violations

Can you explain why transitivity violations rise to 18.23% on CLUBench with KK supplied?

Our additional analysis shows that dataset characteristics explain part of the higher violation rate on CLUBench. A substantial gap remains.

We evaluated all triples on the 60 CLUBench datasets and compared the results with synthetic datasets matched in observation count, feature count, and cluster count (N,D,KN,D,K).

With KK supplied, the CLUBench violation rate had a mean of 18.22% but a median of 5.63%, indicating that datasets with particularly high violation rates raise the mean.

Matching N,D,KN,D,K increased the synthetic violation rate from 2.23% to 5.62%, suggesting that these dataset characteristics explain part of the difference. The 2.23% figure comes from a new synthetic evaluation, separate from the 1.29% reported in the original submission. Even after matching, the CLUBench rate remained 12.60 percentage points higher, with a paired 95% bootstrap interval of [7.8, 17.7] percentage points.

The synthetic and real predictions had similar average pairwise probabilities, proportions of probabilities above 0.9, and entropy. These similarities do not explain the remaining gap, whose causes are still unclear.

We have added these experiments, the per-dataset results, and further analysis to Appendix C.4 of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer hAyQ (2/2)

Response to Q3 — Scalability

For scalability concerns, can you report runtime and memory for NN well beyond 2,000 (e.g., 10410^4–10510^5), and discuss possible approximations that avoid quadratic cost?

We address this question in General Comments, Section 5, “Scalability: runtime, memory, and possible approximations,” including the measured scaling results, explicitly labeled larger-NN projections, and potential approximations.

Response to Q4 — Excluded datasets, full datasets, and label–cluster alignment

We thank the reviewer for raising concerns about the benchmark scope. We evaluated the 43 excluded CLUBench datasets, repeated the original benchmark without the row cap, and examined datasets with stronger label–cluster alignment.

On the 38 excluded datasets within our supported cluster-count range, CoClusterPFN achieves higher mean ARI than native-input TabClustPFN, although the differences are not statistically conclusive. Removing the row cap slightly reduces overall mean ARI. Finally, both models perform substantially better on datasets with stronger label–cluster alignment, where CoClusterPFN maintains the lead in both count settings.

(1) The 43 excluded datasets

The evaluation of all 43 excluded datasets is presented in General Comments, Section 6, “Evaluation on the 43 excluded CLUBench datasets.” Below, we address the remaining questions about full datasets and label–cluster alignment.

(2) Full datasets without subsampling

We evaluated the 19 previously capped datasets using all valid observations in their available CLUBench source files, with dataset sizes ranging from 2,001 to 10,000. The remaining 41 datasets already used all available source rows.

We also evaluated TabClustPFN on the same 19 datasets to examine whether CoClusterPFN's advantage persists without subsampling.

Mean ARI on the 19 previously capped datasets:

MethodSupplied KK: cappedSupplied KK: fullUnspecified KK: cappedUnspecified KK: full
CoClusterPFN0.37770.33030.24620.2365
TabClustPFN0.32440.31310.16660.2163

CoClusterPFN maintains higher mean ARI in both count modes when using all observations, although the gaps over TabClustPFN become smaller and their paired 95% bootstrap intervals include zero. Across all 60 datasets, CoClusterPFN's mean ARI changes from 0.3838 to 0.3688 with supplied KK and from 0.3258 to 0.3228 with unspecified KK.

The full-data results and evaluation protocols are provided in Appendix C.10.

(3) Datasets whose labels align with cluster structure

We examined cluster-label matching (CLM) using the adjusted Silhouette metric of Jeon et al. (2025). The metric compares distances between observations within the same reference class to distances between different classes. Higher scores indicate that the reference labels better reflect the cluster structure in feature space.

A threshold of CLM≥0.5\mathrm{CLM}\geq0.5 identifies 26 datasets where reference labels align more strongly with cluster structure. The same subset is used for both models, independently of their predictions.

Mean ARI by cluster-label alignment:

SettingMethodAll 60CLM ≥0.5\geq0.5 (26)CLM <0.5<0.5 (34)
Supplied KKCoClusterPFN0.38950.72350.1342
Supplied KKTabClustPFN0.34430.63100.1250
Unspecified KKCoClusterPFN0.32590.62620.0963
Unspecified KKTabClustPFN0.28020.50930.1050

Both models achieve substantially higher ARI when the reference labels align more strongly with cluster structure. On the 26 datasets with CLM≥0.5\mathrm{CLM}\geq0.5, CoClusterPFN achieves higher mean ARI in both count settings.

We keep all 60 datasets in the main benchmark and use CLM only as a supplementary analysis.

The excluded-dataset results, full-data evaluations, and CLM analysis are now included in Appendices C.9, C.10, and C.11 of the revised manuscript, respectively, together with the experimental protocols and per-dataset results.

Thanks for the great effort on the detailed experiments and the response. However, it is still unclear what the main contributions and advantages of the proposed methods are compared with existing methods, despite the reported gains in ARI and accuracy. The gain here is better might due to the prior designed by the authors for training. I’d read the result as a combination of a useful prior and a changed output design. So, I will maintain my score.

Reviewer_RMsSOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a new approach for amortized learning of the clustering "posterior similarity matrix" (PSM), the posterior probability that two sampled points are in the same cluster. The new approach achieves good results in terms of accuracy and speed.

Critical strengths
  • A new method that achieves good Bayesian clustering results, with high speed and accuracy

  • Paper was well-motivated and clearly written

Critical weaknesses

I thought the paper was fairly strong overall. There are a few points that could have been clearer:

  • The tabular aspect of the data could have been clarified. Why the focus on tabular data here? Wouldn't the same methods apply to non-tabular data?

  • The PSM is clearly an object of interest, but it doesn't get you quite what we typically want from a Bayesian clustering algorithm (ie a posterior over cluster assignments). Some more discussion about how a typical user should go from the PSM to a good cluster output (or posterior over cluster outputs) would improve the paper.

  • The NxN scaling of the PSM is obviously problematic (and is noted in the paper). Some shortcuts seem possible here (don't compute the PSM on obviously distant sample pairs, consider low-dimensional / low-rank approximations to the PSM) and could be discussed.

Questions for authors

No further questions beyond what is noted in the "weaknesses" section above

1 comment
AuthorsOpenReview ↗

Response to Reviewer RMsS

Why the focus on tabular data

The tabular aspect of the data could have been clarified. Why the focus on tabular data here? Wouldn't the same methods apply to non-tabular data?

Tabular data is widely used in science, healthcare, and industry. Transferring learned knowledge across heterogeneous tabular datasets remains challenging (van Breugel and van der Schaar, 2024). The work of TabPFN has kicked off a revolution in tabular data: they demonstrate that Bayesian priors over tables can be efficiently used for building foundation models (Müller et al., 2022; Hollmann et al., 2025). That is why we believed it was possible to formulate a similar prior for clustering tabular data.

Extending the approach to images, graphs, or sequences is not trivial: designing such priors for other types of data still poses a substantial modeling challenge. Only recently timeseries foundation models have appeared that are solely trained on synthetic data. We have included this as an interesting direction for future work.

We have clarified this motivation in Section 3 and discuss the requirements for extending the approach to other modalities in Section 5.

How to go from PSM to a good cluster output

The PSM is clearly an object of interest, but it doesn't get you quite what we typically want from a Bayesian clustering algorithm (ie a posterior over cluster assignments). Some more discussion about how a typical user should go from the PSM to a good cluster output (or posterior over cluster outputs) would improve the paper. Sections 2 and 3 of the submitted paper describe how to decode the PSM into a hard clustering. We provide further details on the decoding procedure below.

For users seeking a single clustering, MaxPEAR provides a natural starting point. MaxPEAR approximates the posterior expected ARI. Given the PSM, and a candidate clustering, MaxPEAR estimates the expected ARI of the candidate clustering solution. In our implementation, we use 1−S^1-\hat S as the dissimilarity matrix to construct average- and complete-linkage hierarchies. When KK is unknown, we consider cuts with K∈{2,…,10}K\in\{2,\ldots,10\} from both hierarchies. When KK is supplied, we consider the cuts with that number of clusters. In either case, we return the candidate partition with the highest MaxPEAR score.

Note that CoClusterPFN also supports other decision criteria. MinVI-LB minimizes a lower bound on posterior expected VI, an information-theoretic clustering loss. MinBinder minimizes posterior expected Binder loss over the same candidate partitions.

The PSM does not determine the full posterior over partitions: different posterior distributions can produce the same pairwise probabilities. Therefore, while the PSM supports different hard-clustering decisions, it does not contain enough information to recover samples from the full posterior. Doing so would require modeling additional dependencies between cluster assignments. CoClusterPFN does not currently return posterior samples; learning to generate them is a future direction discussed in Section 5.

Potential scaling improvements

The NxN scaling of the PSM is obviously problematic (and is noted in the paper). Some shortcuts seem possible here (don't compute the PSM on obviously distant sample pairs, consider low-dimensional / low-rank approximations to the PSM) and could be discussed.

We discuss selective pairwise prediction, low-rank representations, and alternatives to dense PSM decoding in General Comments, Section 5, “Scalability: runtime, memory, and possible approximations.” This shared response also includes the new runtime and memory measurements.

Reviewer_2gDmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors apply the Prior-data Fitted Network (PFN) paradigm to Bayesian clustering by directly predicting the posterior similarity matrix (PSM). The resulting CoClusterPFN models uncertainty in cluster assignments without the need for per-dataset training. The authors validate CoClusterPFN on synthetic tasks and 60 CLUBench tabular datasets, with an additional case study on the Iris dataset.

Critical strengths
  • S1 The paper is clearly written and well formatted, and the authors did a good job presenting the problem. This makes the central idea easy to follow.
  • S2 The authors provide the inference code. I ran the code, and the main CLUBench results are reproducible, which makes the results trustworthy.
Critical weaknesses
  • W1 Novelty against [1,2,3] is limited. The Cluster-PFN method already extends PFN models to unsupervised Bayesian clustering, while TabClustPFN [2] already applies PFNs to tabular clustering with a pretrained TabICL encoder, and ABC [3] already learns a pairwise similarity matrix. Moreover, Cluster-PFN [1] is not compared against, although it is a natural choice for the 2D Gaussian-mixture setting, and it could also be applied to the low-dimensional CLUBench datasets (e.g., Iris).

  • W2 One of the main design decisions of PFN family models is the synthetic data used for training (the prior). While the prior is documented in the paper, I miss ablations on its multiple design choices. Previous methods such as ZEUS [4] and TabClustPFN [2] use Gaussian mixtures with invertible ResNet transformations, while CoClusterPFN replaces it with several different transformations (e.g. sinusoids, independent nuisance features, duplicate features, noise introduced by cluster label reassignments). The ablations for these components are missing, while they are central for the design of the PFN models. The relative contributions of PSM prediction and the training prior to the improvement over TabClustPFN remain unclear. Moreover, the prior is restricted to at most 20 features, whereas TabClustPFN [2] is trained with up to 64 features and evaluated beyond its training range (e.g., with PCA for higher dimensions). The CLUBench comparison is therefore limited to datasets with at most 20 features, and CoClusterPFN's behavior beyond its training range is not evaluated.

  • W3 The abstract claims that CoClusterPFN "can help users choose among alternative clusterings", and the conclusion states that it "supports exploration of alternative clusterings", but CoClusterPFN outputs a single posterior similarity matrix (PSM), while alternative clustering is a separate line of work [5, 6] that is not cited in the paper and aims to recover a set of distinct, individually valid partitions. The case study is presented on Iris, but Iris is not an alternative clustering benchmark and provides only a single partitioning. Alternative clustering algorithms (e.g., MISC [7]) are designed for such scenarios, and some of them can also estimate the number of clusters and partitions. This undermines the positioning of this contribution and the interpretation of results.

Minor suggestions for the presentation:

  • Consider presenting the prior as an algorithm. It will make it easier to interpret Table 3.
  • Related Work (Sec. 2) and CoClusterPFN (Sec. 3) could be streamlined, which would allow presenting additional results in the main paper.
  • The main text should also present the training objective formally.

References:

[1] Bhaskaran, Prajit, and Tom Viering. "Transformers can do Bayesian Clustering." arXiv preprint arXiv:2510.24318 (2025).

[2] Zhao, Tianqi, et al. "TabClustPFN: A prior-fitted network for tabular data clustering." arXiv preprint arXiv:2601.21656 (2026).

[3] Coward, Samuel, Erik Visse-Martindale, and Chithrupa Ramesh. "Attention-Based Clustering: Learning a Kernel from Context." arXiv preprint arXiv:2010.01040 (2020).

[4] Marszałek, Patryk, et al. "Zeus: Zero-shot embeddings for unsupervised separation of tabular data." Advances in Neural Information Processing Systems 38 (2026): 86464-86501.

[5] Bailey, James. "Alternative clustering analysis: A review." Data clustering (2018): 535-550.

[6] Zou, Bangyu, et al. "Deep Multiple Clustering: From Foundations to Context-Aware Approaches." Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2026.

[7] Wang, Xing, et al. "Multiple independent subspace clusterings." Proceedings of the AAAI conference on artificial intelligence. Vol. 33. No. 01. 2019.

Questions for authors

Please see weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2gDm

Response to W1 — Limited novelty

Our response to W1 is provided in the General Comments above.

Response to W2 — Training prior and benchmark scope

We thank the reviewer for highlighting the importance of the synthetic training prior.

Performance advantage is not only due to the prior

Our experimental comparison with the Responsibility-PFN (Section 4.2 and Appendix C.2) does provide some insight already. In this experiment we find that Direct PSM prediction consistently improves decoded ARI over Responsibility-PFN under all three evaluated decision rules on both benchmarks. These results demonstrate that the advantages of the direct-PSM formulation are not solely attributable to differences in the training prior. The new comparisons with ABC and ClusterPFN, trained on our tabular and NIW–GMM priors respectively, further support this conclusion.

Prior ablations

We have completed three controlled prior ablations. Removing the optional nonlinear transformations produced the largest reduction in supplied-KK ARI; the other interventions had smaller or less conclusive effects on CLUBench.

All variants use the same architecture, training budget, validation protocol, and fixed decoders, with one training seed and a shared CPU FP32 evaluation.

Training priorSupplied KK ARIUnspecified KK ARI
Full prior0.38380.3258
Clean Gaussian only0.36400.3139
Without nuisance/duplicate features0.38260.3311
Without optional nonlinear transformations0.35140.3136

Removing the nonlinear block reduced supplied-KK ARI by 0.0324, with an unadjusted paired 95% interval of [−0.0696, −0.0025]. The other five ARI-change intervals include zero. These are single-seed sensitivity results on CLUBench; a small effect here does not establish that a prior component is unhelpful across other tabular tasks.

The results and experimental details are included in Appendix C.3, Table 11, of the revised manuscript.

Feature-range extrapolation

General Comments, Section 6, “Evaluation on the 43 excluded CLUBench datasets,” reports the additional evaluation, including the 38 datasets with more than 20 features and comparisons with TabClustPFN using native and PCA inputs. Full results appear in Appendix C.9.

Response to W3 — “Alternative clusterings”

We agree that our use of the term “alternative clusterings” was ambiguous. In the literature, alternative or multiple clustering methods such as MISC (Wang et al., 2019) seek several distinct, individually meaningful partitions. These often correspond to different feature subspaces or views.

CoClusterPFN does not optimize for diversity between partitions, and we did not intend to claim that it solves this task. CoClusterPFN predicts one PSM. Users can decide on different hard partitions from a single PSM (using e.g. MaxPEAR, MinBinder, etc.). These decisions summarize the same pairwise uncertainty in different ways. The resulting clusterings are not guaranteed to represent diverse or independent clustering views, posterior modes or posterior samples.

The Iris case study illustrates different resolutions of one hierarchy, rather than serving as a benchmark for recovering distinct clustering views.

In response to this concern, we have added a discussion of alternative-clustering methods and the relevant citation in Section 2 of the revised manuscript. We have also revised the terminology in the abstract, Figure 1 caption, and conclusion:

  • Abstract: “An Iris case study shows how the predicted PSM reveals uncertain relationships and supports comparison of PSM-derived partitions at different resolutions.”
  • Conclusion: “The Iris case study illustrates how the PSM reveals uncertain relationships and supports exploration of different clustering decisions and resolutions through a dendrogram.”

We also consider Bayesian alternative clustering an interesting direction for future work. Specialized priors over distinct clustering views could support the development of such methods, as we now discuss in Section 5 of the revised manuscript.

Response to minor presentation suggestions

We thank the reviewer for these suggestions. We have added an algorithmic overview of the tabular-prior sampler in Appendix A, alongside the existing numerical settings in Table 3, and reference it from Section 3.

We have also streamlined Sections 2 and 3 by shortening the intermediate mean-field formulation and row-encoder description while retaining the key methodological distinctions.

Finally, we now state the training objective explicitly in Section 3, including both supplied- and unspecified-KK prediction. Appendix B.4 retains the detailed pair-sampling estimator and the derivation of its posterior-probability targets.

I thank the authors for their substantial effort in the rebuttal and extensive amount of new evidence. I keep my score and provide the justification for my decision below.

On W1: While I disagree with some of the statements on prior work, in particular that ClusterPFN and TabClustPFN suffer from label switching, since both propose their own solutions to it, I agree with the authors that the uncertainty angle is the main contribution of this work. However, the positioning of the paper is unclear on this, and the empirical evidence for it is limited. While the authors provide many results on clustering quality (by measuring ari / nmi / etc.), the uncertainty itself receives far less attention. As uncertainty is central to the paper, I would expect the experiments to focus more on that and show the failure / limited capabilities of existing methods, which is not visible in the current version of the paper.

On W2: I thank the authors for the new ablations. As mentioned in my initial concern, the authors replace multiple elements of the priors used by previous methods such as ZEUS [4] and TabClustPFN [2], however, the benefit of these replacements remains unclear. I could not find an argument for why one would replace them, or evidence that the proposed transformations are better than the existing priors. A direct comparison, ceteris paribus, would have been needed to support this design decision. My concern about prior limitation also remains. The authors limit the proposed prior to at most 20 features during training, whereas TabClustPFN [2] was trained with up to 64 features, and I couldn't find a justification for this reduction. Naturally, one would expect to extend the amount of features the method is trained on and not to narrow it.

On W3: I thank the authors for revising the terminology. With this clarification, however, the Iris experiment shows different cuts of a single dendrogram, which standard hierarchical clustering also provides. The introduction still argues that a dataset may support multiple plausible partitions and that most algorithms return a single clustering, while the work actually addresses uncertainty estimation. I believe a more precise scope of the work, focused on uncertainty estimation and supported by corresponding experiments, would make it a better paper.


I appreciate the effort the authors have put into this rebuttal. However, in my opinion the claims of the paper and the supporting results still need substantial revision, and I therefore cannot raise my score.

AuthorsOpenReview ↗

Response to Reviewer 2gDm — Follow-up (1/2)

Thank you for considering the additional evidence and explaining your remaining concerns.

W1. Label switching

Our original response acknowledged that both ClusterPFN and TabClustPFN address label switching. Immediately after introducing the problem, we described how each method handles it. Our phrase "suffer from the label switching problem" may have obscured that acknowledgment.

Each method's solution has consequences for training. ClusterPFN introduces an arbitrary, coordinate-dependent ordering convention into its assignment targets. Those targets can change when the ordering changes, even though the underlying partition stays the same. TabClustPFN uses the permutation-invariant SoftARI objective, which optimizes partition agreement rather than posterior probabilities. Its outputs therefore have no guaranteed Bayesian posterior interpretation.

CoClusterPFN's pairwise co-membership targets are invariant to cluster relabeling. We train directly against these indicators using binary cross-entropy, whose population optimum is the true posterior co-clustering probability under the training prior.

W1. Evidence for uncertainty estimation

We disagree with the assessment that the uncertainty-estimation contribution has little empirical support. Our evaluation includes comparisons with posterior references and probability scores, alongside downstream clustering metrics.

Section 3 and Appendix B.4 establish that the population optimum of our training objective is the true posterior co-clustering probability under the specified prior, marginalizing over KK when unspecified.

Section 4.1 directly evaluates posterior approximation accuracy against MCMC references under the NIW–GMM prior.

Section 4.2 and Appendix C.2, Table 9 provide complementary evidence through a controlled comparison with Responsibility-PFN. Direct PSM prediction achieves lower BCE and Brier scores on held-out tasks from the primary tabular prior, demonstrating improvements in probabilistic co-membership prediction beyond downstream clustering quality.

Section 4.4 and Figure 3 further investigate uncertainty estimation through the Iris case study, illustrating how CoClusterPFN reveals uncertain pairwise relationships that are largely absent from the PSMs derived from TabClustPFN and GMM+BIC responsibilities.

Regarding the correctness of uncertainty estimates from our primary tabular model, obtaining a reliable posterior reference under its full heterogeneous training prior is computationally intractable in practice, as explained in our response to reviewer hAyQ (Q2). This motivated our matched NIW–GMM experiment, where posterior accuracy can be evaluated directly. We explicitly acknowledge this limitation in Section 5.

AuthorsOpenReview ↗

Response to Reviewer 2gDm — Follow-up (2/2)

W2. Training prior and feature range

Prior design

We respectfully disagree that a direct comparison with the ZEUS or TabClustPFN prior is necessary to justify the use of our prior. Our generator was designed around the modeling assumptions described in Section 3 and Appendix A, including potentially multimodal classes and irrelevant or redundant features. These mechanisms explicitly represent sources of ambiguity that we aim to include in the training distribution and reflect established modeling considerations in PFN design and Bayesian clustering.

The reviewer asks for evidence that our transformations are better than those used in previous methods. A controlled comparison would be needed to establish that comparative claim, but we do not claim that our prior is superior to theirs or attribute our performance advantage over those methods specifically to replacing their prior mechanisms. The use of a particular generator in previous work does not make it a required starting point for subsequent methods.

Feature range

The reviewer writes:

"The authors limit the proposed prior to at most 20 features during training, whereas TabClustPFN [2] was trained with up to 64 features."

We would like to emphasize that this difference in training ranges is not an architectural limitation of CoClusterPFN, nor does it establish a performance disadvantage. In fact, our new experiments show higher mean ARI for CoClusterPFN than TabClustPFN on datasets beyond CoClusterPFN's training feature range, with both known and unknown KK, despite CoClusterPFN having been trained on at most 20 features and TabClustPFN on up to 64 (General Comments, Section 6; Appendix C.9, Table 25).

The reviewer further states:

"Naturally, one would expect to extend the amount of features the method is trained on and not to narrow it."

We would also like to clarify that TabClustPFN is contemporaneous work, and our prior was developed independently of it. Framing our 20-feature training range as a "reduction" relative to TabClustPFN is therefore misleading, as we did not start from its prior and narrow the dimensionality. Furthermore, we respectfully disagree that increasing the maximum training dimensionality is inherently preferable. Doing so increases computational requirements and changes the distribution of training tasks. Under a finite computational budget, a narrower training range may therefore represent a reasonable trade-off, instead of an inherent limitation.

W3. Scope of the contribution and Iris case study

We respectfully disagree with the characterization of the Iris experiment as merely showing different dendrogram cuts. We do not claim that hierarchical clustering is novel. Our contribution is estimating posterior co-clustering probabilities, which can then be inspected and used for downstream clustering decisions.

Importantly, Section 4.4 and Figure 3 explicitly compare CoClusterPFN's predicted PSM with PSMs derived from TabClustPFN and GMM+BIC responsibilities. Approximately 44% of CoClusterPFN's off-diagonal pairwise probabilities lie between 0.1 and 0.9, compared with only 0.7% for TabClustPFN and 0.0% for GMM+BIC. The baseline-derived matrices largely reproduce their respective hard partitions, whereas CoClusterPFN exposes substantial intermediate probabilities that allow users to inspect uncertain co-clustering relationships. These differences are already present in the predicted matrices before any hierarchical clustering.

We also disagree that the introduction's motivation is inconsistent with uncertainty estimation. The possibility that a dataset supports multiple plausible partitions is precisely why estimating uncertainty is valuable. A single hard partition does not communicate uncertainty about pairwise cluster membership. CoClusterPFN addresses this by estimating posterior co-clustering probabilities, providing a pairwise summary of that uncertainty. We do not claim to recover distinct clustering views or the full posterior over partitions, as already clarified in our original response and the revised manuscript.

Closing remarks

Finally, we would like to emphasize that the revised manuscript and rebuttal have substantively addressed several concerns raised in the original review. We have added comparisons with ClusterPFN and ABC, conducted controlled prior ablations, evaluated CoClusterPFN beyond its training feature range, clarified the distinction from alternative clustering, and implemented the requested presentation improvements, including an explicit training objective and an algorithmic description of the prior.

We recognize that the reviewer may still have reservations about the contribution, but we believe these substantial additions directly address the main experimental and presentation concerns raised in the original review and deserve consideration in the final assessment.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Authors' General Comments (1/5)

We thank the reviewers for their careful assessment.

Substantive additions and revisions are highlighted in blue in the updated manuscript.

Both reviewers 2gDm and hAyQ raise novelty concerns relative to ClusterPFN, TabClustPFN, and ABC. We respond in three ways: (1) discuss the scope of the requested comparisons under the ICLR guidelines; (2) clarify the methodological distinctions between these methods and our work; and (3) present new comparisons with ClusterPFN and ABC, alongside an additional Bayesian baseline.

Overview of the General Comments

  1. Scope of the requested comparisons: ICLR guidelines on contemporaneous work and arXiv preprints.
  2. Methodological distinctions: differences from ClusterPFN, TabClustPFN, and ABC.
  3. ClusterPFN comparison: new results on matched NIW–GMM tasks and CLUBench.
  4. ABC and Bayesian baselines: new comparisons on CLUBench.
  5. Scalability: runtime, memory, and potential improvements.
  6. Excluded CLUBench datasets: evaluation beyond the training feature and cluster-count ranges.

1. Scope of the requested comparisons

The ICLR 2027 reviewer guidelines state that comparisons with contemporaneous work or papers available only as arXiv preprints are not required, and that their absence cannot be a basis for rejection. We hope this policy will be taken into account when assessing the absence of these comparisons in the original submission. Nevertheless, we agree that comparisons with ABC and ClusterPFN are informative and have now added them to the revised manuscript.

2. Our contribution in detail compared to Cluster-PFN, TabClustPFN and ABC

As explained in Section 2 of the original submission, CoClusterPFN combines direct pairwise prediction with the PFN framework. It predicts a label-invariant PSM without factorizing co-clustering probabilities through cluster-specific responsibilities, and the same output supports multiple clustering decision rules. Section 3 and Appendix B.4 define its conditional and count-marginalized posterior targets.

ClusterPFN and TabClustPFN predict cluster-specific assignment probabilities. These methods therefore suffer from the label switching problem, which we resolve with the label invariant PSM. Cluster-PFN solves this with a canonical ordering (the cluster labels are determined by distance to the origin), and TabClust-PFN resolves it by using a differentiable version of the ARI, which is label-invariant. However, TabClustPFN's SoftARI objective optimizes partition agreement rather than directly targeting posterior probabilities.

CoClusterPFN explicitly uses the predicted PSM for uncertainty inspection, hierarchical visualization, and clustering decisions under different loss functions. The PSM can be decoded by different algorithms and loss functions to obtain different hard cluster labels, following a Bayesian decision theoretic treatment.

ABC also learns pairwise similarities, but does not explicitly formulate its predictions as a Bayesian PSM under a specified clustering prior or investigate their use for Bayesian uncertainty analysis and loss-dependent partition estimation. ABC also does not aim to build a foundation model for tabular clustering; its experiments focus on Omniglot and synthetic circles.

Our contribution therefore lies in combining prior-fitted Bayesian inference with direct PSM prediction, enabling uncertainty-aware interpretation and flexible downstream clustering decisions from a single predicted matrix. We additionally demonstrate this approach on a substantially broader collection of real-world tabular clustering datasets and provide new empirical comparisons against the related methods.

In the revised Section 2, we have expanded the comparison with ABC to clarify the distinction between its learned-kernel formulation and our prior-fitted Bayesian PSM formulation and decision workflow. We have also made the contrast in cluster-count conditioning explicit.

AuthorsOpenReview ↗

Authors' General Comments (2/5)

3. Empirical comparison with ClusterPFN

CoClusterPFN achieves higher mean ARI than ClusterPFN in both count modes on the shared 2D NIW–GMM tasks and on the nine CLUBench datasets with two to five active features, which are compatible with ClusterPFN's published five-feature checkpoint.

For a controlled comparison, we retrained ClusterPFN on the shared 2D NIW–GMM prior using its published architecture and objective, matching CoClusterPFN's training budget of 1.92 million tasks. We also evaluated the published ClusterPFN checkpoint, which was trained on a different distribution with substantially more examples.

EvaluationMethodSupplied KKUnspecified KK
2D NIW–GMMCoClusterPFN0.18750.1886
ClusterPFN, matched-prior retraining0.03910.0428
ClusterPFN, published checkpoint0.07380.0374
CLUBench (9 datasets)CoClusterPFN0.39670.3361
ClusterPFN, published checkpoint0.16580.1887

We have added these comparisons to Sections 4.1 and 4.3, with the corresponding experimental protocols in Appendices G and C.1.

4. Empirical comparison with ABC and Bayesian clustering

On all 60 CLUBench datasets, CoClusterPFN achieves higher mean ARI than our ABC+TabICL adaptation and the variational Bayesian mixture baseline in both count modes.

We adapted ABC to heterogeneous tabular inputs using a TabICL row encoder, while keeping its attention-based affinity architecture. We trained the adaptation using the same synthetic prior and training budget as CoClusterPFN. This extends ABC beyond its original evaluation on synthetic circles and Omniglot.

MethodSupplied KKUnspecified KK
CoClusterPFN0.38950.3259
ABC+TabICL, native spectral0.35060.2768
Variational Bayesian mixture0.32470.2415
TabClustPFN (existing baseline)0.34430.2802

Here, Variational Bayesian mixture is the added Bayesian baseline. We used a finite variational Bayesian GMM when KK was supplied. When KK was unspecified, we used a variational Dirichlet-process GMM truncated at 10 components. We used a mature implementation available in scikit-learn for this baseline.

We have added ABC+TabICL and the variational Bayesian mixture baseline to Section 4.3, with their experimental protocols in Appendix C.1 and complete results in Appendix C.5.

AuthorsOpenReview ↗

Authors' General Comments (3/5)

5. Scalability: runtime, memory, and possible approximations

We have added runtime and memory measurements beyond the training observation range to Appendix C.7 of the revised manuscript. We have also expanded the discussion of potential scaling improvements in Section 5 and Appendix C.7.

We profiled CoClusterPFN and TabClustPFN at N∈{1,000,2,000,4,000,8,000}N\in\{1{,}000,2{,}000,4{,}000,8{,}000\} on an NVIDIA A40, using cached inputs with 20 features, FP32 inference, and five repetitions. At N=8,000N=8{,}000, median runtimes were:

MethodSupplied KK (s)Unspecified KK (s)
CoClusterPFN (GPU inference)7.237.22
TabClustPFN (GPU inference)0.1860.391
CoClusterPFN (CPU MaxPEAR decoding)4.5217.27

MaxPEAR was measured separately on saved PSMs using four CPU threads. CoClusterPFN's peak PyTorch-allocated GPU memory was 2.55 GiB (3.89 GiB reserved), and peak process RSS was approximately 3.00 GiB, including decoding.

Extrapolation to N=104N=10^4–10510^5. Doubling NN from 4,000 to 8,000 increased inference time by 3.9×3.9\times, consistent with quadratic scaling. Assuming quadratic scaling for inference and decoding, we obtain the following estimates for the slower, unspecified-KK setting:

NNGPU inferenceCPU decodingTotalFP32 PSM memory
8,000 (measured)7.22 s17.27 s24.49 s256 MB
10410^4 (estimated)11.3 s27.0 s38.3 s400 MB
10510^5 (estimated)18.8 min45.0 min63.8 min40 GB

Hierarchical clustering can be implemented in O(N2)O(N^2) (Müllner, 2011), consistent with our extrapolation. The PSM can be constructed in chunks and stored in CPU memory or on disk, avoiding the need to keep it entirely on the GPU.

These runtime projections do not establish that the current implementation can run at N=105N=10^5: the dense PSM alone would require 40 GB of storage, with additional memory needed for the encoder and decoding. Given these resource requirements, we did not attempt a full evaluation at this scale on our A40 setup.

Scaling inference and memory. CoClusterPFN's full self-attention and dense PSM prediction both require O(N2)O(N^2) computation. Inducing-point attention (Lee et al., 2019) could reduce attention complexity to O(Nm)O(Nm) for mm inducing points, while predicting PSM entries only for selected pairs, such as neighboring observations or pairs involving landmarks, could reduce pairwise computation. Recently TabICLv2 (Qu et al., 2026) demonstrates million-row inference in 450 s on an NVIDIA H100, using under 50 GB GPU memory, 24 GB CPU RAM, and approximately 250 GB disk storage. Although TabICLv2 retains O(N2+ND2)O(N^2+ND^2) computation, it demonstrates that quadratic complexity need not prevent practical scalability through efficient implementation and memory offloading. However, CoClusterPFN additionally requires dense pairwise prediction and PSM storage, which would remain a bottleneck at such scales. Addressing this would require reducing pairwise costs alongside improvements in attention efficiency.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.