acceptodds
Under review as a conference paper at ICLR 2027

Learning to Compare Neural Representations from Parameterized Computation Graphs

Abstract

Comparing internal representations across trained models is important for identifying corresponding layers and selecting components for model reuse. Yet, direct comparisons at inference require evaluating candidate models on known reference inputs and collecting their internal activations. In this paper, we study whether internal representation similarity measured across independently trained models can instead be learned to predict in unseen models from their *parameterized computations*. We introduce a framework that applies a graph metanetwork for mapping each internal site to an embedding derived from its upstream computations while accounting for parameter symmetries. We analyze the structural properties of our estimators and derive conditional guarantees for layer retrieval and stitching candidate selection. Empirically, we find that the learned predictors support relation estimation and low-regret layer retrieval. They also select useful stitching connections while not being trained on stitching outcomes. These results show that representational relations learned across a collection of trained models can generalize to new computations, and that the predicted relations can guide candidate selection before direct measurement or adapter fitting for functional evaluations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation search, representation similarity, neural network artifacts, graph metanetworks, parameter symmetries

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_6k9kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes PARIS, a graph metanetwork that learns to predict representational similarity scores across different sites in neural networks. Their method is different from methods such as CKA and CKNNA which require computing and storing model activations on a shared set of probe images. In contrast, PARIS directly performs message massing on model weights to compare them across sites.

Critical strengths

The authors present theoretical justification of their method’s ability to account for parameter symmetries. Empirically, they show that their model beats many other methods on tasks that require accurately measuring representational similarity.

Critical weaknesses

Although the paper provides an interesting method for computing representational similarity, it should be rejected because lacks a few crucial results, and without those results its contribution is minor at best. Specifically:

  1. Results are only presented across the same kind of network with different hyperparameters. This limits the usefulness of the authors’ method, especially since previous work (such as the paper that presents the MutualNN metric used in this submission) already compares across different networks and different modalities. It is not clear if PARIS is able to train across different types of networks and/or modalities, and if so, how it performs.

  2. In figure 6, it is not clear why the target is an existing metric like a CKA-based dissimilarity. It is also not clear if there is a significant advantage in training a meta-network since the target metric can be directly computed in the situations the authors run experiments on.

  3. It seems like one of the paper’s advantages is that they do not need to run inference on samples to compute similarity scores for new networks. However, runtime results are not included for measuring how much compute the meta-network requires throughout training or during inference. Therefore, it is challenging to reason about the usefulness of their method.

Overall, the idea could be promising but the lack of results and analysis make it challenging to reason about the usefulness of their method.

Questions for authors

Is PARIS able to train across different types of networks and/or modalities? And if so, how does it compare to other methods that can?

Is there an advantage in training a meta-network to predict a metric when that metric can be directly computed in the situations the authors run experiments on?

How does inference time compare with existing methods, and if inference time is faster with PARIS, is it fast enough to justify the meta-network training cost?

6 comments
AuthorsOpenReview ↗

Response to Reviewer 6k9k (Part 1 out of 4)

We thank the reviewer for their thoughtful comments and constructive suggestions! We appreciate their recognition of PARIS's ability to compare representations directly from model weights without relying on shared probe images and the empirical results.

Following the reviewer’s feedback, we have conducted additional experiments and revised the manuscript to clarify and strengthen several aspects of PARIS.

For convenience, we group related comments in our responses below and point to the corresponding revisions, which are highlighted in blue in the revised paper.\textcolor{blue}{\text{in blue in the revised paper.}} In terms of weaknesses and questions:

W1: Results are only presented across the same kind of network with different hyperparameters. This limits the usefulness of the authors’ method, especially since previous work (such as the paper that presents the MutualNN metric used in this submission) already compares across different networks and different modalities. It is not clear if PARIS is able to train across different types of networks and/or modalities, and if so, how it performs.

Q1: Is PARIS able to train across different types of networks and/or modalities? And if so, how does it compare to other methods that can?

Architecture Coverage and Generalization

We thank the reviewer for raising this important question about the generality of PARIS!

Yes, PARIS can indeed be trained on heterogeneous architecture families -- Appendix E.5 evaluates a metanetwork trained on a population containing both ResNets and ViTs. We have expanded this experiment with Deep Meta Classifier (DMC) and Graph Metanetwork baselines, and report the results:

SettingDMCGraph MetanetworkPARIS-IndPARIS-Joint
CIFAR-10 / ResNet0.5060.5520.6080.646
CIFAR-10 / ViT0.5950.6430.7010.861
CIFAR-100 / ResNet0.5020.5450.5820.655
CIFAR-100 / ViT0.5480.6740.7130.762

We also introduce an out-of-distribution architecture evaluation, where the metanetwork is trained on architectures with smaller hidden dimensions and evaluated on held-out architectures with larger hidden dimensions. PARIS-Ind and PARIS-Joint achieve correlations of 0.715 and 0.732, respectively, compared with 0.375 for DMC and 0.654 for Graph Metanetwork. These experiments show that PARIS can learn from heterogeneous architecture populations and generalize across the tested architectural configurations.

We have expanded Appendix E.5\textcolor{blue}{\text{Appendix E.5}} to include the additional metanetwork comparisons and OOD evaluation. Thanks so much for this comment!

We could also revise the discussion in the main text, but we wonder whether it would be better to keep these in the appendix to maintain the flow of the paper (which is what we have done for now). From a reader’s perspective, does the reviewer have a preference for one of these two options over the other?

Extension Beyond Unimodal Models

We thank the reviewer for this helpful comment and agree that extending our framework beyond unimodal setting is a valuable direction.

Existing approaches, including the MutualNN-based analysis, can compare representations across modalities using activations computed from paired observations. Our method addresses a complementary question: whether such representation relations can be predicted from trained computations without requiring input data or activations at inference. So these two are not competing estimators -- activation-based methods provide the similarity scores that PARIS learns to predict, while PARIS aims to predict these scores for unseen models without collecting new activations.

Extending the metanetwork framework to independently trained unimodal vision and language models or more structured outputs is not trivial, since it would require constructing suitable cross-modal reference relations (i.e., paired, partial paired, unpaired data) and adapting the computation-graph representation to different model architectures (eg different positional encoding), which we view as important and promising future work. We are actively working on this extension and plan to address it in a follow-up work toward a weight-space analogue of representational-convergence tests.

AuthorsOpenReview ↗

Response to Reviewer 6k9k (Part 2 out of 4)

Extension Beyond Unimodal Models (cont.)

As a first step beyond unimodal setting, we included new results on independently pretrained CLIP models taken from Nezhurina et al. (2025) where image and text encoders were jointly trained on paired image-text data. We compare both PARIS variants against an adapted Graph Metanetwork baseline, illustrating that our PARIS has potential beyond unimodal:

Modality pairGraph MetanetworkPARIS-IndPARIS-Joint
Image-Image0.5530.7020.745
Text-Text0.5710.6880.729
Image-Text0.4820.6530.704

These results provide initial evidence that PARIS can predict representation relations involving both image and text modalities. We plan to investigate the cross-modal representational alignment (ie from independently trained unimodal vision and language models) directly from trained weights in future work (toward a weight-space Platonic test!).

We have added the pretrained CLIP experiments to Appendix E.10\textcolor{blue}{\text{Appendix E.10}} and updated the discussion in Section 5.\textcolor{blue}{\text{Section 5.}}


W3: It seems like one of the paper’s advantages is that they do not need to run inference on samples to compute similarity scores for new networks. However, runtime results are not included for measuring how much compute the meta-network requires throughout training or during inference. Therefore, it is challenging to reason about the usefulness of their method.

Q3: How does inference time compare with existing methods, and if inference time is faster with PARIS, is it fast enough to justify the meta-network training cost?

Computational Costs and Practical Motivation for Metanetwork-Based Similarity Prediction

We thank the reviewer for raising this question! We agree that it is important to understand the computational costs of our method for assessing the practical usefulness.

Training PARIS-Ind and PARIS-Joint requires approximately 5.56 and 7.64 GPU hours. We additionally benchmark both variants against direct activation-based evaluation with cached descriptors using 200 sites from distinct models:

MeasurementPARIS-IndPARIS-JointDirect evaluation
Metanetwork training~5.56 GPU h~ 7.64 GPU h—
Preparation of 200 sites~180 s~186 s~ 74.2 s
Comparison per pair~ 1.1 ms~265 ms~ 40.5 ms
Peak memory~ 0.34 GB~ 0.67 GB~ 3.82 GB

We remark that:

  1. Although direct evaluation is approximately 2.4 times faster during preparation, PARIS-Ind achieves an approximately 37-fold speedup in prepared pairwise comparisons and uses approximately 11 times less peak memory. We do not claim that PARIS is computationally preferable for one-off evaluations. Its computational advantage is most relevant when the trained metanetwork and encoded sites are reused across large numbers of comparisons.

  2. Beyond computational acceleration, PARIS enables similarity prediction for new models without requiring probe data or running additional base-model inference. This makes it applicable in settings where direct activation-based evaluation is unavailable or impractical. This also avoids introducing variability from newly sampling probe examples at query time, and it allows the learned relationships to be reused across unseen models using only their trained computations.

Our central goal is to study whether a metanetwork can learn to predict internal representation similarities from trained computations and generalize to unseen models without requiring probe data or additional base-model inference. We view this capability as the main methodological motivation for studying metanetwork-based representation similarity, which goes beyond replacing direct similarity computation in settings where it is already inexpensive.

We have expanded Appendices D.1 and D.3\textcolor{blue}{\text{Appendices D.1 and D.3}} to report training costs, inference runtime, and peak memory, and clarified the computational trade-offs and broader motivation for activation-free similarity prediction in Section 5.\textcolor{blue}{\text{Section 5.}}


AuthorsOpenReview ↗

Response to Reviewer 6k9k (Part 3 out of 4)

W2: In figure 6, it is not clear why the target is an existing metric like a CKA-based dissimilarity. It is also not clear if there is a significant advantage in training a meta-network since the target metric can be directly computed in the situations the authors run experiments on.

Q2: Is there an advantage in training a meta-network to predict a metric when that metric can be directly computed in the situations the authors run experiments on?

Why Learn to Predict Existing Representation Similarity Measures?

We thank the reviewer for raising this question! We agree that existing similarity measures can be computed directly when suitable input data (e.g., paired) and activations are available. However, our goal is to study a different question: whether previous activation-based measurements can be used to learn a predictor that subsequently estimates relations between sites in newly encountered models from their trained computations. Beyond this scientific motivation, such an approach also offers practical advantages, as the upfront training cost can be amortized across many comparisons, potentially leading to substantial savings in computation and memory when comparing large numbers of representations.

We compute activation-based reference scores on held-out models to evaluate prediction quality. These activations are not inputs to PARIS. We use established measures as prediction targets because they capture different properties of internal representations. By training PARIS on similarity scores computed from activations, we investigate whether a metanetwork can learn these relationships from model parameters (i.e., those we called neural artifacts) and generalize to unseen models. Once trained, PARIS can predict these scores without requiring probe data or collecting new activations.

Prior metanetwork work has demonstrated that the parameters and computational structure of trained neural networks contain useful information about their properties. PARIS builds on this insight but addresses a different prediction target. Instead of predicting model-level properties (eg performance scores or characteristics of the training data), we aim to predict how similar two specific internal representations are under a given activation-based similarity measure. These similarities provide a more fine-grained view of how representations relate to one another, capturing aspects of their underlying geometry that go beyond individual model properties.

We also investigate whether this approach offers practical benefits:

  1. PARIS-Ind produces compact site embeddings that can be reused across different comparisons and similarity measures. In our additional runtime experiments, prepared PARIS-Ind comparisons are approximately 37 times faster per pair of sites than direct activation-based comparison with cached descriptors, while using approximately 11 times less peak memory in the evaluated workload. Although training the metanetwork and encoding the graphs introduce upfront costs, these costs can be amortized across subsequent comparisons. Specifically, when performing many comparisons with large probe sets, the substantial gains in computational efficiency and memory usage can readily outweigh the initial investment, making PARIS-Ind advantageous at scale (Appendix D.3).

  2. We evaluated whether predicted similarity scores can support downstream tasks (Sections 4.3 and 4.4), including model retrieval and stitching-candidate selection. We have further extended our evaluation to pretrained CLIP models, including representation comparisons across different modalities (updated in Appendix E.10\textcolor{blue}{\text{Appendix E.10}}), and investigate transfer with adaptation to a new population of pretrained vision encoders on Flowers-102 (updated Appendix E.12\textcolor{blue}{\text{Appendix E.12}}). We observe that both PARIS variants outperform the adapted Graph Metanetwork baseline in cross-modal image-text similarity prediction (0.653/0.704 vs. 0.482) and Flowers-102 transfer (0.64/0.73 vs. 0.58). These experiments provide additional evidence for the utilities and broader applicability of learning representation relations from trained computations.

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

PARIS addresses whether the similarity between intermediate representations of independently trained neural networks can be predicted from their architectures, trained parameters, and inference state, without collecting new activations from the compared models at inference time. Each representation site is described by a graph of its entire upstream computation, rather than only the parameters of the layer that immediately produces it. The graph construction is designed to account for hidden-coordinate and channel permutations, as well as certain changes of basis within Transformer attention heads, by representing the composed query-key and value-output operators. The authors introduce two predictors: PARIS-Ind independently encodes each site into a reusable embedding and predicts relation-specific distances, whereas PARIS-Joint uses bidirectional cross-attention to compare two site graphs jointly. Both models are trained using activation-derived similarity targets, combining regression and ranking losses, on populations of 20,000 ResNets and 13,500 Vision Transformers trained for 300 epochs on CIFAR-10 and CIFAR-100. The experiments evaluate five representation relations and show that PARIS-Ind is generally competitive with other independent weight-space encoders, while PARIS-Joint achieves stronger aggregate rank correlations and lower layer-retrieval regret; the appendix reports aggregate Spearman correlations ranging from 0.641 to 0.764 for PARIS-Ind and 0.729 to 0.890 for PARIS-Joint across the four dataset–architecture settings. Finally, the authors show that the predictors can help select model-stitching candidates without being trained on stitching outcomes, although the evaluation is limited to small-scale vision benchmarks and a restricted candidate pool.

Critical strengths

The paper identifies a well-motivated problem, a genuine bottleneck in representation analysis and model reuse.

The method is reasonably designed. PARIS represents each layer using its upstream computation graph, including network structure and trained weights. This provides more information than using only local layer weights or layer positions. The method accounts for some common changes in parameter values that do not change the network's computation. This makes comparisons between networks more reliable.

The paper tests five similarity measures and compares PARIS with several existing methods. It also evaluates layer retrieval and model stitching, rather than only measuring prediction accuracy. PARIS performs well on layer retrieval and helps select useful stitching candidates, even though it is not trained directly on stitching results.

Critical weaknesses

(W1) The novelty over existing methods is not fully established. The paper builds on graph-based weight encoders and pairwise graph comparison methods, and it discusses earlier work on activation-free network similarity. However, some of the closest parameter-based similarity methods are not evaluated directly. Comparisons on suitable shared settings would help show how much PARIS improves over existing approaches.

(W2) The experiments use only CIFAR-10 and CIFAR-100 with ResNets and Vision Transformers. Although the test models are unseen during meta-training, they belong to the same datasets and architecture families. The paper therefore does not establish whether PARIS works well on large pretrained models, other datasets, or substantially different architectures. The stitching evaluation is also small and narrow. The method selects five candidates from a pool of 20 and is tested only on same-family ResNet and ViT stitching using CIFAR datasets. These results are promising, but they do not show whether PARIS remains effective with larger candidate pools or heterogeneous pretrained models.

(W3) The practical speed advantage is unclear. PARIS requires constructing and encoding upstream computation graphs, which can be costly. The paper discusses computational complexity but does not provide enough wall-clock or memory measurements to show when PARIS becomes cheaper than direct activation-based comparisons, including methods that cache activations.

(W4) The symmetry guarantees are limited. PARIS handles graph relabelings and certain attention basis changes, but it does not cover all function-preserving parameter changes. Since symmetry handling is an important part of the method, the authors should state more clearly which transformations are supported and how the remaining limitations affect comparisons.

(W5) The paper mainly reports Spearman correlation, which measures whether predicted scores rank pairs correctly, rather than how close the predicted scores are to the reference values. Reporting MAE or RMSE would help clarify whether PARIS can estimate similarity scores accurately or is mainly useful for ranking candidates.

Questions for authors

(Q1) Does the main score-prediction evaluation use only pairs of sites from different base models? The model-disjoint split is appropriate, but the description of pair construction does not make clear whether Section 4.2 includes pairs of sites from the same model. If it does, could the authors report the score-prediction results separately for within-model and cross-model pairs, using the existing test population? The cross-model-only result is particularly important because matching sites across independently trained networks is the main use case.

(Q2) How far does the method generalize beyond the CIFAR model populations? Could the authors add one modest test on a held-out dataset or a genuinely unseen architecture family, ideally using existing pretrained models rather than training a new population from scratch? If this is not feasible, a more systematic analysis of the failure modes in the current diverse-architecture experiment would still help.

(Q3) Please report wall-clock time and peak memory for graph construction, site encoding, and comparison.

(Q4) How does PARIS compare with existing activation-free similarity approaches where their assumptions overlap? The paper discusses Inner Product-based Neural Network Similarity and tensor similarity, but neither appears in the empirical baseline set. Could the authors either provide a comparison on an appropriate subset or explain in more detail why these methods cannot be compared fairly?

(Q5) Could the authors clearly distinguish the transformations that leave the graph and predictions invariant by construction from function-preserving reparameterizations that are not covered? A small check using a known equivalent reparameterization outside the supported set e.g. an appropriate positive rescaling of ReLU coordinates, would be useful if the transformation is applicable to the selected site and relation. If the current encoding is intentionally invariant only to a restricted family of transformations, I would prefer that scope to be stated prominently rather than relying on the broad phrase “accounting for parameter symmetries.”

8 comments
AuthorsOpenReview ↗

Response to Reviewer HAKZ (Part 1 out of 6)

We thank the reviewer for their careful assessment and constructive feedback! We are delighted to read that the reviewer recognizes the motivation behind PARIS, our use of upstream computation graphs, and the potential of PARIS for applications beyond score prediction.

Following the reviewer’s feedback, we have conducted additional experiments and revised the manuscript to clarify and strengthen several aspects of PARIS. We also found a typo in the model population sizes in our initial submission: the model zoo actually contains 2,000 ResNet models and 1,350 ViT models. We have fixed them in the revised paper (all experiments and results in the original submission used these populations -- this correction changes the reported model counts but does not change the experimental results).

To make our responses easier to follow, we address related weaknesses and questions together below. We point to the relevant manuscript revisions throughout, with the updated text highlighted in blue in the revised paper.\textcolor{blue}{\text{in blue in the revised paper.}} In terms of weaknesses and questions:

(W1) The novelty over existing methods is not fully established. The paper builds on graph-based weight encoders and pairwise graph comparison methods, and it discusses earlier work on activation-free network similarity. However, some of the closest parameter-based similarity methods are not evaluated directly. Comparisons on suitable shared settings would help show how much PARIS improves over existing approaches.

(Q4) How does PARIS compare with existing activation-free similarity approaches where their assumptions overlap? The paper discusses Inner Product-based Neural Network Similarity and tensor similarity, but neither appears in the empirical baseline set. Could the authors either provide a comparison on an appropriate subset or explain in more detail why these methods cannot be compared fairly?

The novelty and contribution over metanetwork literature

We thank the reviewer for encouraging us to better clarify our novelty and strengthen the comparison with prior metanetworks!

Our central contribution is to formulate and investigate a new problem for metanetworks: Can we predict the similarity between two internal representations using only the trained computations that produce them, without running the models on new data? 

Prior metanetwork work has demonstrated that the parameters and computational structure of trained neural networks contain useful information about their properties. PARIS builds on this insight but addresses a different prediction target. Instead of predicting model-level properties (eg performance scores or characteristics of the training data), we aim to predict how similar two specific internal representations are under a given activation-based similarity measure. These similarities provide a more fine-grained view of how representations relate to one another, capturing aspects of their underlying geometry that go beyond individual model properties. To learn this relationship, PARIS is trained using similarity scores computed from activations. At inference, it predicts these scores for internal sites in held-out models using only their upstream computations without requiring input data or activations.

The problem we studied introduces two challenges that are not addressed by existing graph-based weight encoders or pairwise graph-comparison methods:

  1. The representation at a given site depends not only on its own layer but also on the computations before it (i.e., we need to identify the exact sites being compared). A whole-model representation (as is commonly done in the literature) does not directly identify a particular site, while the weights of a single layer leave out earlier computations that also affect its output. This motivates our use of site-specific graphs that capture the upstream computation that lead to it. These graphs also account for neuron/head permutations and attention-basis transformations, so that equivalent parameterizations can be treated consistently.
  2. The goal is not simply to determine whether two computation graphs are structurally or parametrically similar, but to predict how similar their resulting representations are. Different measures compare different properties, so two sites may be similar under one measure but not another. A general-purpose weight encoder or graph-comparison method can not establish how its outputs should correspond to these activation-based quantities. PARIS addresses this challenge by learning the relationship between pairs of trained computations and their target similarity scores, allowing it to approximate different measures through measure-specific supervision.
AuthorsOpenReview ↗

Response to Reviewer HAKZ (Part 2 out of 6)

The novelty and contribution over metanetwork literature (cont.)

These considerations motivate the two comparison designs we study: PARIS-Ind independently encodes the upstream computation graph of each site and then compares the resulting embeddings, whereas PARIS-Joint allows information to flow between the two site graphs before pooling, enabling pair-dependent interactions during representation learning. We build on existing graph metanetwork encoders and graph-comparison tools, and do not claim these underlying primitives as novel, but: our methodological contribution lies in adapting and integrating them into a framework for site-level representation similarity prediction, through site-specific upstream computation graphs and measure-specific supervision, while respecting the relevant permutation and attention-basis symmetries.

Beyond the formulation and design, we provide empirical and theoretical evidence for this approach. We evaluate how accurately PARIS predicts multiple representation similarity measures on held-out models and whether these predictions are useful for downstream tasks, including model retrieval and stitching-candidate selection. We also characterize structural limitations of the two designs and analyze how prediction errors affect retrieval and candidate selection under stated conditions. These results establish both what can be learned about internal representation similarity from trained computations and when such predictions can support downstream decisions.

In summary, our contribution is not a new graph encoder or pairwise comparison primitive in isolation, but: the formulation and methodological design of learning to predict internal representation similarity from trained computations. This distinguishes PARIS from prior metanetwork work in its prediction target and the site-specific information required to address it. We have revised Sections 1-2\textcolor{blue}{\text{Sections 1-2}} to make these distinctions and our relationship to prior work more explicit.

Comparison with additional parameter-based metanetworks

Following the reviewer’s comment, we added Neural Graphs (NG) from Kofinas et al. (2024) as an additional baseline for PARIS-Ind, alongside the weight-space encoders already included in our experiments. We updated Section 4 and Figures 6-8\textcolor{blue}{\text{Section 4 and Figures 6-8}} to include this comparison. The results show that PARIS-Ind still achieves better performance than NG. We also updated Appendix D.4\textcolor{blue}{\text{Appendix D.4}} with details of the new baseline.

Comparison with activation-free similarity approaches: Inner Product-based Neural Network Similarity and tensor similarity

We thank the reviewer for this thoughtful suggestion! We agree that these activation-free methods are closely related to our work and that empirical comparisons would help clarify how PARIS differs from them. Following the reviewer's comment, we have expanded Appendix E.7\textcolor{blue}{\text{Appendix E.7}} to explain their assumptions and added three complementary experiments to compare them with PARIS where their settings overlap.

Let us first clarify why these methods cannot be directly applied to our main benchmarks. Filter-subspace similarity (Chen et al., 2023) assumes that models use a decomposition with shared coefficients. However, our ResNet and ViT checkpoints were trained without this constraint. Tensor similarity (Gonzalez et al., 2026) supports polynomial or multilinear computations, but our benchmarks also contain non-polynomial operations, such as ReLU and softmax attention. Applying these methods directly would then require us to change their assumptions or the computations being compared.

To provide empirical comparisons, we consider three additional settings:

  1. Filter-subspace similarity on standard ResNets

    We relax the subspace constraints for compatible tensor shapes, allowing us to compare models using shared coefficients. In this setting, the atom-cosine score reduces to kernel cosine similarity without reconstruction error. We evaluate this approach on 50 ResNet models trained on CIFAR-10. The filter-subspace score achieves correlations of 0.27/0.28/0.24/0.17/0.19 with linear CKA, CKA, GULP, MutualNN, and readout disagreement, respectively. PARIS-Ind achieves 0.64/0.54/0.63/0.79/0.52, while PARIS-Joint achieves 0.61/0.75/0.67/0.93/0.80.

AuthorsOpenReview ↗

Response to Reviewer HAKZ (Part 3 out of 6)

Comparison with additional parameter-based metanetworks (cont.)

  1. Filter-subspace similarity on models trained with shared coefficients

    We also construct a population of 60 ResNet models trained on CIFAR-10, where the atoms are trained with fixed and shared coefficients (as the setup in Chen et al., (2023)). We then evaluate PARIS using the exact effective kernels and inference state of these same models. In this setting, the filter-subspace score achieves correlations of 0.93/0.95/0.89/0.74/0.81 with the same five measures. PARIS-Ind achieves 0.96/0.99/0.90/0.78/0.85, while PARIS-Joint achieves 0.95/0.99/0.92/0.76/0.88.

    PARIS remains competitive with or outperforms filter-subspace similarity even when the latter's assumptions are satisfied, while offering broader applicability to conventionally trained models.

  2. Tensor similarity on linear models

    To compare with tensor similarity under supported computations, we use 300 independently trained three-layer linear models on different regression problems. Tensor similarity achieves correlations of 0.94/0.64/0.67/0.72/0.52 with linear CKA, CKA, MutualNN, GULP, and readout disagreement, respectively. PARIS-Ind achieves 0.95/0.77/0.82/0.84/0.63, while PARIS-Joint achieves 0.97/0.80/0.84/0.89/0.67.

Across these settings, both PARIS variants achieve stronger performance. These experiments provide comparisons where the methods' assumptions overlap, while the expanded discussion explains why a direct comparison on our original ResNet and ViT benchmarks would require further changes.

We have added both the experimental results and these clarifications to  Appendix E.7\textcolor{blue}{\text{ Appendix E.7}} in the revised paper.

We could also briefly discuss these comparisons in the main text to better highlight how PARIS differs from existing activation-free methods, but we wonder whether it would be better to keep the detailed analysis in the appendix to maintain the flow of the paper (which is what we have done for now). From a reader’s perspective, does the reviewer have a preference for one of these two options over the other?


(W2) The experiments use only CIFAR-10 and CIFAR-100 with ResNets and Vision Transformers. Although the test models are unseen during meta-training, they belong to the same datasets and architecture families. The paper therefore does not establish whether PARIS works well on large pretrained models, other datasets, or substantially different architectures. The stitching evaluation is also small and narrow. The method selects five candidates from a pool of 20 and is tested only on same-family ResNet and ViT stitching using CIFAR datasets. These results are promising, but they do not show whether PARIS remains effective with larger candidate pools or heterogeneous pretrained models.

(Q2) How far does the method generalize beyond the CIFAR model populations? Could the authors add one modest test on a held-out dataset or a genuinely unseen architecture family, ideally using existing pretrained models rather than training a new population from scratch? If this is not feasible, a more systematic analysis of the failure modes in the current diverse-architecture experiment would still help.

Generalization Beyond CIFAR and Stitching-Candidate Selection from Lager Pool

We thank the reviewer for this valuable point! Following the reviewer's suggestion, we have run several additional experiments to investigate these points:

  1. Generalization to pretrained models, different modalities, and new datasets

    To explore this, we have extended our evaluation to two new populations of pretrained CLIP models:

    (a) We consider independently pretrained CLIP checkpoints from Nezhurina et al. (2025), where the image and text encoders within each model were jointly trained on paired image-text data. We investigated image-image, text-text, and image-text, and compare both PARIS variants against an adapted Graph Metanetwork baseline, illustrating that our method has potential beyond unimodal. We observe that both PARIS variants outperform the adapted Graph Metanetwork baseline in cross-modal image-text similarity prediction (0.653/0.704 vs. 0.482). These results have updated in Appendix E.10\textcolor{blue}{\text{Appendix E.10}}, which provides initial evidence that PARIS can learn representation relations involving both image and text modalities.

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

This paper proposes a framework called Parameter-based Analysis for Relational Inference of Sites (PARIS), which builds on graph metanetworks (Lim et al., 2024a; Kofinas et al., 2024). PARIS takes an internal site of a neural network, together with its entire upstream computation (architecture, trained parameters, and inference state), as input, and can be used to predict representation (dis)similarity between internal sites directly from their weights. PARIS supports multiple (dis)similarity measures (linear CKA, RBF CKA, GULP, MutualNN, and linear-readout disagreement). PARIS has two variants: PARIS-Ind (which encodes each site into a reusable embedding) and PARIS-Joint (a cross-encoder that takes two inputs and outputs a similarity prediction). The paper also includes some theoretical analysis and applications to model stitching candidate selection.

Critical strengths
  • From my understanding, this is the first work to apply upstream computations to graph metanetworks.
  • Comprehensive experiments and theoretical results (though I didn't check any proofs in the appendix).
Critical weaknesses

I still have some concerns on its practical value. I think graph metanetworks are already difficult to apply to large models, and PARIS seems to be even more expensive to train.

Questions for authors
  1. Could the authors report (estimated) GPU hours for (i) training the model zoo, and (ii) training each metanetwork?

  2. Let's consider an alternative: train a metanetwork that takes the entire neural network as input and have it output an embedding for each site (like PARIS-Ind), or pick a specific site and output its (dis)similarity (like PARIS-Joint). This seems to have a cost advantage. Could the authors explain why train and use each upstream computation separately?

  3. Fig. 5 shows a pattern that could also be produced by a simple relative-depth rule. My understanding is that PARIS is trying to be a more general framework, but these tasks (similarity, stitching) happen to be strongly correlated with relative depth. So:

  • To what extent can the similarity be modeled by relative depth? Can we compare against a relative-depth baseline?
  • PARIS is not given relative depth explicitly, but to what extent can it indirectly learn relative depth (for example, inferring from the shape of the input), and what does it learn beyond depth?
6 comments
AuthorsOpenReview ↗

Response to Reviewer moXh (Part 1 out of 4)

We thank the reviewer for their thorough reading and insightful comments! We are glad to know that the reviewer recognized the contribution of using upstream computations for graph metanetworks and appreciated our comprehensive experimental and theoretical analysis.

Following the reviewer’s feedback, we have conducted additional experiments and revised the manuscript to clarify and strengthen several aspects of PARIS. To make our responses easier to follow, we point to the relevant manuscript revisions throughout, with the updated text highlighted in blue in the revised paper.\textcolor{blue}{\text{in blue in the revised paper.}} In terms of weakness and questions:

W: I still have some concerns on its practical value. I think graph metanetworks are already difficult to apply to large models, and PARIS seems to be even more expensive to train.

Practical Value

We thank the reviewer for this point! We agree that using metanetworks to learn neural networks can be expensive, and that it is important to consider both the benefits and computational costs of PARIS.

Our main goal is to study whether metanetworks can learn how internal representations relate to each other from trained neural networks. We train PARIS using activation-based similarity scores, but once trained, it can predict these scores for unseen models using only their model parameters (i.e., those we called neural artifacts) and generalize to unseen models: PARIS can predict these scores without requiring probe data or collecting new activations.

Prior metanetwork work has demonstrated that the parameters and computational structure of trained neural networks contain useful information about their properties. PARIS builds on this insight but addresses a different prediction target. Instead of predicting model-level properties (eg performance scores or characteristics of the training data), we aim to predict how similar two specific internal representations are under a given activation-based similarity measure. These similarities provide a more fine-grained view of how representations relate to one another, capturing aspects of their underlying geometry that go beyond individual model properties.

We also investigate whether this approach offers practical benefits:

  1. PARIS-Ind produces compact site embeddings that can be reused across different comparisons and similarity measures. In our additional runtime experiments, prepared PARIS-Ind comparisons are approximately 37 times faster per pair of sites than direct activation-based comparison with cached descriptors, while using approximately 11 times less peak memory in the evaluated workload. Although training the metanetwork and encoding the graphs introduce upfront costs, these costs can be amortized across subsequent comparisons. Specifically, when performing many comparisons with large probe sets, the substantial gains in computational efficiency and memory usage can readily outweigh the initial investment, making PARIS-Ind advantageous at scale (Appendix D.3).

  2. We evaluated whether predicted similarity scores can support downstream tasks (Sections 4.3 and 4.4), including model retrieval and stitching-candidate selection. We have further extended our evaluation to pretrained CLIP models, including representation comparisons across different modalities (updated in Appendix E.10\textcolor{blue}{\text{Appendix E.10}}), and investigate transfer with adaptation to a new population of pretrained vision encoders on Flowers-102 (updated in Appendix E.12\textcolor{blue}{\text{Appendix E.12}}). We observe that both PARIS variants outperform the adapted Graph Metanetwork baseline in cross-modal image-text similarity prediction (0.653/0.704 vs. 0.482) and Flowers-102 transfer (0.64/0.73 vs. 0.58). These experiments provide additional evidence for the utilities and broader applicability of learning representation relations from trained computations.

Our goal is not solely to optimize computational efficiency, but: we investigate how effectively a metanetwork can learn from upstream computations and transfer this knowledge to unseen models. Beyond its practical advantages in settings where suitable probe data are unavailable or impractical, or where many comparisons are required, this approach offers a more fundamental perspective on trained neural networks as a new data modality: it allows us to investigate what information about a model's internal representations can be extracted directly from its weights, without requiring access to input data or explicit activation-based comparisons. We believe that understanding what can be learned directly from neural network parameters is an important and promising direction toward a deeper understanding of how neural networks organize and relate their internal representations.

We have updated Sections 1 and 5\textcolor{blue}{\text{Sections 1 and 5}} to better explain this motivation and the practical benefits. Thanks again for the comment!


AuthorsOpenReview ↗

Response to Reviewer moXh (Part 2 out of 4)

Q1: Could the authors report (estimated) GPU hours for (i) training the model zoo, and (ii) training each metanetwork?

GPU-Hour Costs for Model Zoo and Metanetwork Training

We thank the reviewer for this question! We have added the estimated GPU hours for the model zoo and both PARIS variants in Appendices D.1 and D.3.

We also found a typo in the model population sizes in our initial submission: the model zoo actually contains 2,000 ResNet models and 1,350 ViT models. We have fixed them in the revised paper (all experiments and results in the original submission used these populations -- this correction changes the reported model counts but does not change the experimental results). The training costs are:

Training componentGPU hours
Model zoo~658
PARIS-Ind~5.56
PARIS-Joint~7.64

Training the model zoo accounts for most of the computational cost (but we also showed that our method can be applied to existing model zoos -- see updated Appendices E.10 and E.12\textcolor{blue}{\text{Appendices E.10 and E.12}}). However, this cost is incurred once when building the training population. After training PARIS, we can use it to predict representation similarities in held-out models without retraining the metanetworks or collecting new activations. Although preparing the site graphs and embeddings still takes time, PARIS-Ind can reuse these embeddings across different comparisons. PARIS-Ind is about 37 times faster per pair than direct activation-based comparison with cached descriptors, once preparation is complete.

As discussed above, we believe that trained neural networks themselves offer a useful source of data (i.e., a new data modality). As more models become available, studying their architectures and learned parameters may help us better understand how they form internal representations.

We have corrected the model population sizes in Appendix D.1\textcolor{blue}{\text{Appendix D.1}} and added the training costs and runtime analysis in Appendix D.3\textcolor{blue}{\text{Appendix D.3}}.


AuthorsOpenReview ↗

Response to Reviewer moXh (Part 3 out of 4)

Q2: Let's consider an alternative: train a metanetwork that takes the entire neural network as input and have it output an embedding for each site (like PARIS-Ind), or pick a specific site and output its (dis)similarity (like PARIS-Joint). This seems to have a cost advantage. Could the authors explain why train and use each upstream computation separately?

Why Encode Site-Specific Upstream Computations Instead of Entire Networks?

We thank the reviewer for this thoughtful suggestion! The question is a good one: indeed, the metanetwork and graph metanetwork methods really did inspire this work and are aligned since a whole-network metanetwork with site-conditioned outputs can amortize computation when many sites from the same network -- so much so that when we explored this idea originally, we did it in the way suggested by the reviewer! However, we found it harder to obtain stable generalization with limited model populations. This motivated us to consider which computations are actually relevant to each representation and how we can have the same learned metanetwork parameters that are shared across all site-specific graphs within each meta-training setting.

Our main motivation is to use only the computations that produce the representation we want to compare. In a fixed model, changing layers after a selected site does not affect its representation, eg, replacing the final classifier leaves all earlier representations unchanged. However, a whole-network encoder also sees these downstream parameters and may learn to rely on them when predicting similarity. This can lead to overfitting, especially when only a small number of models are available for meta-training. By using only the upstream computations, PARIS avoids dependence on these irrelevant parameters by construction.

To expand on this point, we added the comparison between PARIS with whole-network encoding using different numbers of models in Appendix D.3.7. We test both independent and joint designs on ResNet and ViT populations trained on CIFAR-10 and CIFAR-100. Across these settings, PARIS achieves higher prediction. The difference is large when fewer models are available, suggesting that focusing on upstream computations helps PARIS generalize.

This design also allows us to reuse the same site description when a model changes, eg, replacing the classifier, adding layers after the selected site, or reusing the same upstream computation in another model does not change its site graph. We believe this property is useful for comparing representations across models and studying how changes to model parameters affect their internal representations. Also, importantly, PARIS does not train a separate metanetwork for each site and the same learned parameters are shared across all site-specific graphs.

We have updated the method discussion to clarify this design motivation in Section 3.2\textcolor{blue}{\text{Section 3.2}} and added the whole-network comparison to Appendix D.3.7\textcolor{blue}{\text{Appendix D.3.7}}.


Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions and additions

Dear Reviewers and Area Chair,

We thank all reviewers for their thoughtful and constructive feedback!

As the discussion period comes to a close, we would like to summarize the additions made during it and where each can be found:

  • We expanded our baseline evaluations to include Neural Graphs (Section 4, Figures 6-8), as well as activation-free similarity methods based on filter-subspace and tensor similarity (Appendix E.7).
  • We showed that PARIS can generalize across different architectures by jointly training on ResNet and ViT models and, separately, by evaluating on held-out larger widths (Appendix E.5).
  • We further extended our experiments to CLIP image-text representations (Appendix E.10) and transfer to Flowers-102 (Appendix E.12), both using existing pretrained checkpoints without training a new model collection, and to stitching with larger candidate pools (Appendix E.11).
  • We compared whole-network encoding with site-specific encoding (Appendix D.3.7), added relative-depth baselines (Appendix E.9), and evaluated the effect of positive weight rescaling (Appendix E.8).
  • We provided details on GPU-hours, runtime, and memory usage (Appendices D.1 and D.3), and reported prediction errors as MAE and RMSE (Appendix E.4). For scale, a collection of 50 models with 21 sites each already contains 540,225 distinct cross-model site pairs, and at the measured ~1.1 ms per pair once sites are encoded (Appendix D.3), PARIS-Ind can compare these in about ten minutes.
  • We presented a new weight-perturbation selection experiment (on November 18) in our response to Reviewer 6k9k, currently available only in the discussion. PARIS selects three of 24 candidate weight modifications per model and achieves a mean regret of 0.013, compared with 0.030 for the strongest baseline.

These additions provide further evidence for the central claim of our paper: relationships between internal representations can be predicted from architecture and trained weights alone, even for previously unseen models, and these predictions can guide downstream candidate selection, as in retrieval (Section 4.3) and stitching-candidate selection (Section 4.4).

Looking ahead, we believe this direction provides a starting point for studying representational convergence from model artifacts. Inspired by the Platonic Representation Hypothesis (Huh et al., 2024), a further question is whether the shared structure observed across representation spaces can be inferred and identified from the trained computations that produce them. We regard such weight-space tests of the Platonic Representation Hypothesis as a longer-term research direction, and we see PARIS as an initial step toward investigating these questions directly from neural artifacts.

We thank the Reviewers and the Area Chair again for the reviewing process.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.