acceptodds
Under review as a conference paper at ICLR 2027

Walking Across Images: Random Walk Pairing for Self-Supervised Learning

Abstract

The ability to capture both local invariances and global semantic structures is essential for effective self-supervised learning. In this paper, we address a fundamental challenge of ensuring robust generalization to unseen data while effectively capturing global semantic structures. Motivated by a spectral embedding perspective, we identify potential sources of suboptimal generalization in SSL methods that can be viewed through a spectral lens. We propose a simple yet effective training strategy that can be easily applied to a variety of SSL methods. We introduce SAG-VICReg (Stable and Generalizable VICReg) as a detailed instantiation, demonstrating improvements in generalization and global semantic understanding. Through comprehensive experiments across diverse SSL frameworks including SimCLR and TCL, we show broad applicability with consistent improvements. The same pairing also improves the fine-tuning of today's strongest pretrained backbones, with gains that persist under linear probing, grow as labels become scarce, and transfer to unseen datasets. Our enhanced methods exhibit superior performance on metrics designed to evaluate global semantic understanding while maintaining competitive results on local evaluation metrics. We also propose a label-free evaluation metric that accounts for global data structure without requiring labels–a key advantage when labeled data is scarce or unavailable.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
SSL, Random walk, Spectral embedding, Embeddings evaluation metric, VICReg

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_k2MVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SAG-VICReg, which is a random walk-guided method for self-supervised learning. Motivated by a spectral perspective, this paper constructs a similarity graph and samples positive from neighboring images. SAG-VICReg uses similarity-weighted invariance between sampled pairs. This paper also introduces representation comparison using LCA and cophenetic correlations. Experimental results show that the proposed method outperforms existing methods.

Critical strengths
  1. The comprehensive experiments show the effectiveness of the proposed method.

  2. This paper is easy to follow with clear writing.

Critical weaknesses
  1. My main concern is the pairing mechanism using neighboring representations is well established, such as NNCLR and I-Con. I am curious whether the proposed SAG-VICReg would still hold an advantage if they were applied to VICReg.

  2. The experimental results on ImageNet-1k are conducted only with 250 epochs. The authors should provide more epochs and recent SSL methods to validate the effectiveness of SAG-VICReg.

Questions for authors

See weakness.

5 comments

We thank the reviewer for the positive assessment of our experimental coverage and the clarity of the paper. We appreciate the suggestions regarding comparisons with existing pairing methods and longer ImageNet training, and address both concerns below, including new experimental evidence.

W1 — Would SAG-VICReg still hold an advantage if NNCLR / I-Con pairing were applied to VICReg

We thank the reviewer for this suggestion. We conducted a new controlled experiment applying NNCLR-style pairing within VICReg. Both variants use the same architecture, training schedule, warm-up, and base VICReg loss coefficients. The NNCLR-style variant deterministically retrieves the nearest neighbor from a queue of 16,384 past embeddings, whereas SAG-VICReg samples probabilistically from its online mini-batch affinity graph.

Hierarchical kk-NN accuracy on CIFAR-100 (%):

MethodLevel 1Level 2Level 3Level 4
VICReg92.62 ± 0.7783.08 ± 0.5971.21 ± 0.4357.99 ± 0.48
VICReg + NNCLR pairing (new)93.18 ± 0.4183.83 ± 0.2772.21 ± 0.3958.73 ± 0.32
SAG-VICReg (Ours)93.89 ± 0.5284.89 ± 0.7073.18 ± 0.6560.01 ± 0.57

SAG-VICReg maintains its advantage at all four hierarchy levels, including a 1.28-point gain at the finest level. This comparison evaluates the two pairing strategies under matched architectures and training conditions. We have included this experiment, together with the corresponding hierarchical linear evaluation, in Appendix B.5 of the revised manuscript.

The nearest-neighbor ablation in Appendix C.2 further supports this finding under a fixed candidate pool. With uniform weighting, random walk pairing outperforms deterministic nearest-neighbor selection (59.21% vs. 57.94%). The advantage persists with affinity weighting (60.01% vs. 59.27%), suggesting that the benefit extends beyond the choice of candidate pool.

Regarding I-Con, Theorem 4 establishes a connection to a simplified VICReg objective that excludes the covariance term, rather than the full VICReg objective used in our experiments. Moreover, its experimental neighbor propagation relies on pretrained DINO features and a precomputed global neighbor graph for clustering, rather than training VICReg from scratch. In contrast, our method constructs the graph online from current representations, uses locally scaled transition probabilities, and retains the full VICReg objective, including covariance regularization. Therefore, while the methods are conceptually related, I-Con's theoretical formulation and experimental results do not directly establish the effectiveness of its neighbor-propagation mechanism within full VICReg.

Overall, the new experiment demonstrates that SAG-VICReg outperforms NNCLR-style retrieval when both are applied within the same VICReg training setup, while requiring no memory queue or externally pretrained features.

AuthorsOpenReview ↗

W2 (Part 1/2)

W2 — ImageNet-1k training duration and recent SSL methods

We agree that longer training is the common setup, and we acknowledge this limitation in the paper. Training all four methods (SimCLR, Barlow Twins, VICReg, and SAG-VICReg) for 1,000 epochs on ImageNet-1k, with multiple runs each, was beyond our computational budget. Therefore, all methods in Table 3 were trained for 250 epochs under the same protocol, ensuring a fair comparison. Importantly, our VICReg baseline already reaches 71.74% linear top-1 accuracy after 250 epochs, within 1.5 percentage points of the 73.2% reported by the original VICReg paper after 1,000 epochs. SAG-VICReg further improves this to 72.46%, demonstrating gains over an already competitive baseline. Moreover, SAG-VICReg improves VICReg across all seven ImageNet evaluation metrics in Table 3.

Regarding recent SSL methods, Table 5 already evaluates DINOv2, TIPSv2, and CAPI, showing improvements in all 20 configurations, with gains of up to +1.65 percentage points.

To further strengthen this evaluation, we conducted additional experiments with DINOv2 ViT-S/14 and TIPSv2 ViT-B/14 across six additional datasets: EuroSAT, Food-101, RESISC45, GTSRB, DTD, and Tiny-ImageNet. For each backbone, both variants use the same native objective, training data, and fine-tuning schedule; the only intervention is our random walk pairing. We evaluate both 20% and full-data fine-tuning, with five runs per configuration.

Table R1. Fine-tuning on 20% of the training data. Gains in k-NN top-1 accuracy (percentage points) over the corresponding vanilla fine-tune.

BackboneEuroSATFood-101RESISC45GTSRB
DINOv2 ViT-S/14+0.84+0.79+0.75+0.74
TIPSv2 ViT-B/14+0.68+0.51+0.90+1.64

Table R2. Fine-tuning on the full training data. Gains in k-NN top-1 accuracy (percentage points) over the corresponding vanilla fine-tune.

BackboneDTDGTSRBTiny-ImageNet
DINOv2 ViT-S/14+0.88+0.32+0.81
TIPSv2 ViT-B/14+1.83+1.22+0.67

Random walk pairing improves all 14 new configurations, with gains of up to +1.64 points on GTSRB and +1.83 points on DTD for TIPSv2. Together with Table 5, these results provide further evidence that our approach benefits recent pretrained SSL models across different architectures, datasets, and both partial-data and full-data fine-tuning. We have included these new experiments in Appendix B.8 (Tables 18 and 19) of the revised manuscript.

While longer ImageNet-1k pretraining would provide additional validation, the matched ImageNet comparisons and the broader experiments with recent SSL models support the effectiveness and general applicability of our approach.

AuthorsOpenReview ↗

W2 (Part 2/2)

Additional robustness evaluation — Distribution shift on CIFAR-100-C

To further evaluate robustness under distribution shift, we tested our existing DINOv2 and TIPSv2 checkpoints on CIFAR-100-C, without retraining. The goal is to assess whether random walk pairing produces representations that generalize more reliably when the test distribution differs from the clean training distribution.

CIFAR-100-C introduces this distribution shift by applying 19 types of corruption (noise, blur, weather, and digital distortions) at five severity levels (1 = mild, 5 = severe) to CIFAR-100 test images. The models are trained and fine-tuned exclusively on clean images, while corruptions are introduced only at test time. We evaluate the frozen representations using k-NN classification against the clean CIFAR-100 training set (k=20k=20).

The table reports mean corrupted accuracy (mCA), averaged across corruption types and evaluated severity levels, alongside the gains at the lowest and highest severities (percentage points).

Table R3. Robustness under distribution shift on CIFAR-100-C.

Backbone (fine-tuning data)mCA (Vanilla → Ours)mCA GainGain (Sev. 1 → 5)
DINOv2 ViT-S/14 (10%)60.33 → 62.76+2.43+2.05 → +2.60
DINOv2 ViT-S/14 (20%)62.28 → 64.49+2.22+1.60 → +2.65
DINOv2 ViT-S/14 (100%)65.82 → 66.44+0.62+0.26 → +0.83
TIPSv2 ViT-B/14 (20%)64.18 → 65.59+1.41+1.13 → +1.65
TIPSv2 ViT-L/14 (20%)72.29 → 74.15+1.85+1.23 → +2.45

For TIPSv2, mCA is averaged over severities 1, 3, and 5; DINOv2 uses all five severities.

Random walk pairing improves robustness under distribution shift across all evaluated configurations, with gains of up to +2.43 percentage points in mean corrupted accuracy. The improvements generally increase with corruption severity and extend across corruption types: accuracy improves on 18/19 corruptions for DINOv2 at 10% and 20%, and on all 19 for both TIPSv2 models.

These results demonstrate that the benefits of random walk pairing extend beyond clean-data performance to generalization under distribution shift, where models must maintain accurate representations despite changes in the input distribution. We have included this evaluation in Section 5 (Table 6) of the revised manuscript, including the experimental setup, corruption severity analysis, and results across corruption types.

We hope the new NNCLR-style comparison, together with the clarification of I-Con and the additional evaluations of recent SSL methods, addresses the reviewer's concerns. We would be happy to answer any further questions and hope the reviewer will consider updating their assessment in light of these results.

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

This paper addresses the limitations of self-supervised learning (SSL) methods that primarily rely on positive pairs generated from augmentations of the same image, potentially limiting semantic relationships across images and generalization to unseen data. Motivated by a spectral embedding perspective, the authors propose a random walk-based cross-image positive pairing strategy, with SAG-VICReg as its primary instantiation. The method is evaluated across several SSL frameworks, including SimCLR, TCL, DINOv2, TIPSv2, and CAPI, reporting consistent improvements. The paper also introduces a label-free evaluation metric based on dendrograms to assess the hierarchical structure and consistency of learned representations.

Critical strengths
  1. Simple Design and Ease of Integration
  • The proposed random walk-based positive pairing strategy can be integrated into existing SSL frameworks without major architectural modifications or additional learning modules. Its simplicity and compatibility with existing objectives make it potentially useful for improving pretrained representations.
  1. Consistent Improvements Across Multiple SSL Methods
  • The authors report improvements across VICReg, SimCLR, TCL, DINOv2, TIPSv2, and CAPI. The evaluation also includes hierarchical classification, transfer learning, and structural similarity metrics, providing multiple perspectives on the effectiveness of the proposed method.
Critical weaknesses
  1. Limited Methodological Novelty and Insufficient Justification for the VICReg-Centered Analysis
  • Although the paper emphasizes broad applicability across SSL frameworks, its theoretical motivation and detailed analysis primarily focus on VICReg. While the authors justify this choice through its simple objective and connection to spectral embedding, it remains unclear whether VICReg particularly benefits from the proposed strategy compared to other SSL objectives.
  • Furthermore, the methodological novelty relative to existing cross-image positive pairing methods, such as NNCLR, is not sufficiently established. The best-performing configuration uses a one-step random walk, which is conceptually similar to stochastic nearest-neighbor sampling. Appendix Table 16 shows only a modest improvement over similarity-weighted nearest-neighbor pairing (59.42% vs. 60.01%). Therefore, the unique contribution of random walk sampling beyond existing positive pairing strategies remains unclear.
  1. Limited Evidence for Robustness, Scalability, and Generalization
  • The proposed method constructs an affinity graph within each mini-batch, making its effectiveness sensitive to batch size and dataset characteristics. Although Appendix Table 18 investigates batch-size sensitivity, the results vary considerably across datasets. On CIFAR-100, performance decreases from 60.01% to 54.18% as batch size increases from 256 to 1024, whereas ImageNet-1k shows the opposite trend. This raises concerns about the robustness and scalability of the proposed method across different training configurations.

  • Additionally, many detailed experiments and ablations focus on CIFAR-100. While evaluations on ImageNet-1k, Caltech-256, and VOC are included, experiments on recent pretrained SSL models are largely limited to CIFAR-100 and ImageNet-100, primarily using k-NN classification. The reported improvements are also mostly below one percentage point. Consequently, the evidence remains insufficient to establish whether the proposed method consistently improves generalization across diverse data distributions and challenging downstream tasks.

Questions for authors
  1. Beyond its simple formulation and spectral interpretation, is there a specific reason for selecting VICReg as the primary instantiation? Does random walk pairing provide greater benefits for VICReg than for other SSL objectives? Additionally, what advantages does the proposed one-step random walk offer over existing stochastic nearest-neighbor pairing strategies?

  2. Appendix Table 18 shows performance degradation on CIFAR-100 at larger batch sizes, whereas ImageNet-1k exhibits the opposite trend. What factors explain these differences, and under what conditions can the proposed method consistently improve performance?

  3. Could the authors provide further evidence that random walk pairing improves generalization beyond CIFAR-100 and ImageNet-100 classification benchmarks, particularly under distribution shifts or on more challenging downstream tasks?

12 comments
AuthorsOpenReview ↗

W1 - Why VICReg as the primary instantiation

We thank the reviewer for the detailed feedback. We appreciate the questions regarding methodological novelty, the choice of VICReg, batch-size sensitivity, and generalization. We address each concern below, supported by additional experiments and analyses.

W1 / Q1 — Why VICReg as the primary instantiation

We chose VICReg as our primary case study not because we expect it to benefit more than other SSL objectives, but because its formulation allows us to isolate and analyze the effect of cross-image pairing.

In particular, VICReg's invariance term can be expressed as a graph-Laplacian objective. Standard same-image pairing induces disconnected components, each containing augmentations of a single image. While the variance and covariance terms prevent collapse, they do not explicitly enforce relationships between different images. Our method addresses this limitation by introducing cross-image connections based on representation similarity.

Importantly, VICReg allows us to modify the positive-pair relationships while keeping the variance and covariance terms unchanged. This provides a particularly clean experimental setting for isolating the contribution of our pairing mechanism, without introducing changes to negative-pair interactions, as would occur in contrastive objectives such as SimCLR.

We do not claim that VICReg benefits more than other SSL objectives. In fact, on CIFAR-100, random walk pairing improves SimCLR's finest-level k-NN accuracy by 5.09 percentage points (Figure 1), compared with 2.02 points for VICReg (Table 2a). We also observe improvements with TCL across all hierarchy levels, and with DINOv2, TIPSv2, and CAPI across all 20 evaluated fine-tuning settings (Table 5).

Thus, VICReg serves as a controlled theoretical and experimental instantiation, rather than a method uniquely suited to our approach. Our results demonstrate that the benefits of random walk pairing extend across different SSL objectives, including methods with substantially different learning mechanisms.

AuthorsOpenReview ↗

W1 - Novelty relative to nearest-neighbor pairing

W1 / Q1 — Novelty relative to nearest-neighbor pairing

We thank the reviewer for raising this point. While our one-step random walk is related to nearest-neighbor sampling, there are important differences in both the pairing mechanism and how it is incorporated into SSL objectives.

First, NNCLR deterministically selects the single nearest neighbor from a memory bank, whereas our method samples from a probability distribution over the kk nearest neighbors, constructed using locally scaled Gaussian affinities. Rather than always selecting the closest neighbor, our method can select the second, third, fourth, or any other neighbor within the kk-nearest-neighbor set, with higher probabilities assigned to more similar neighbors. This allows the model to learn from a wider range of cross-image relationships during training, instead of relying exclusively on the single closest match. In SAG-VICReg, we additionally weight the invariance term by the selected pair's affinity, reducing the influence of potentially less reliable positives.

Other methods, such as SNCLR, All4One, and MUGS, also exploit multiple neighbors, but do so through mechanisms such as attention-based aggregation, additional learning modules, or auxiliary objectives. Our approach instead directly samples a positive partner from an online mini-batch affinity graph, without requiring a memory bank or additional modules.

Beyond the existing ablations, we conducted a new experiment directly comparing NNCLR-style pairing and SAG-VICReg within the same VICReg training setup, as detailed in our response to reviewer k2MV (W1). The NNCLR-style variant selects the nearest neighbor from a queue of 16,384 embeddings, while keeping the architecture, training schedule, warm-up, and base VICReg loss coefficients unchanged. On CIFAR-100, SAG-VICReg outperforms this adaptation at all four hierarchy levels, achieving 93.89% vs. 93.18% (L1), 84.89% vs. 83.83% (L2), 73.18% vs. 72.21% (L3), and 60.01% vs. 58.73% (L4). This corresponds to gains of up to +1.28 percentage points. We have included the new k-NN and linear evaluation results in Appendix B.5 of the revised manuscript.

The nearest-neighbor ablation in Appendix C.2 provides a complementary comparison with the candidate pool fixed to the batch. We acknowledge that the improvement over similarity-weighted nearest-neighbor selection is modest (60.01% vs. 59.42%), but it further supports the benefit of probabilistic pairing beyond deterministic neighbor selection.

More importantly, our contribution is not limited to improving nearest-neighbor selection. Our pairing strategy is lightweight and applicable across different SSL objectives. We demonstrate improvements across VICReg, SimCLR, TCL, DINOv2, TIPSv2, and CAPI, without introducing new objectives, auxiliary losses, memory banks, or architectural modifications. We believe this combination of spectral motivation, probabilistic cross-image pairing, simplicity, and broad applicability is the main contribution of our approach.

Regarding walk length, as discussed in our response to reviewer MtUq (W1/Q1), the optimal number of steps depends on the dataset's semantic structure. For datasets with many classes relative to the batch size, such as CIFAR-100, Caltech-256, and ImageNet-1k, relatively few same-class images appear in each batch. Longer walks are therefore more likely to reach unrelated images and introduce noisy positives, making one-step sampling preferable. In contrast, datasets with fewer classes contain larger same-class neighborhoods, allowing longer walks to capture more diverse positive relationships. Indeed, our new CIFAR-10 experiment shows that three-step walks outperform one-step walks by 0.43 percentage points in k-NN accuracy. We therefore adopt one-step sampling as our default for the datasets studied in the paper, while recognizing that longer walks can be beneficial when the batch contains denser semantic neighborhoods. This experiment and the corresponding training statistics have been added to Appendix C.2 of the revised manuscript.

AuthorsOpenReview ↗

W2 - Batch-size sensitivity

W2 / Q2 — Batch size sensitivity

We thank the reviewer for this question. The degradation on CIFAR-100 at larger batch sizes is also observed in vanilla VICReg, and is even more pronounced. We repeated the batch-size ablation using the same training setup (fine-grained kk-NN accuracy, %):

Batch size2565121024
VICReg57.9955.98 ± 0.2350.28 ± 0.12

From batch 256 to 1024, VICReg loses 7.71 points, compared with 5.83 points for SAG-VICReg (Table 23), which remains better at every batch size.

The main factor appears to be the number of optimization steps. At a fixed 500-epoch budget, increasing the batch size from 256 to 1024 reduces the number of updates from approximately 98k to 24k. To verify this, we trained VICReg at batch 512 for 1,000 epochs, matching the number of updates at batch 256. Accuracy recovered from 55.98% to 57.65%, approaching the 57.99% baseline. On ImageNet-1k, the much larger dataset provides substantially more updates per epoch, making this limitation less pronounced.

Meanwhile, the quality of our cross-image pairing does not deteriorate with larger batches: the fraction of selected positives sharing the anchor's fine-grained class increases from 15% at batch 128 to 22% at batches 512 and 1024.

Thus, our method improves over VICReg at every tested batch size, while absolute performance remains sensitive to the base method's optimization budget. Larger batches can provide more meaningful neighbors, provided sufficient optimization steps are maintained. We have added the new VICReg experiments and this analysis to Appendix C.2.2 (Table 23) of the revised manuscript.

Comments on the paper

6 comments

We thank the reviewer for recognizing the simplicity, practical value, and broad experimental coverage of our approach. We appreciate the concerns regarding the role of random walk sampling and the consistency of improvements, and address both below with additional experiments and clarifications.

W1 / Q1 — Random walk vs. nearest-neighbor pairing

We agree that our default configuration uses a one-step random walk, which is closely related to stochastic nearest-neighbor sampling. Importantly, it remains a random walk on the constructed affinity graph, where positives are sampled according to transition probabilities rather than deterministically selecting the closest neighbor. Our method can select any of the kk nearest neighbors, with probabilities determined by locally scaled Gaussian affinities. Table 16 shows that this improves over deterministic nearest-neighbor pairing (60.01% vs. 59.42%).

Regarding walk length, our goal is to introduce meaningful cross-image connections, rather than necessarily explore distant regions of the graph. We find that the optimal number of steps depends on the structure of the dataset. In datasets with many classes, such as CIFAR-100, Caltech-256, and ImageNet-1k (Table 15), relatively few same-class images appear in each batch. For example, with a batch size of 256, CIFAR-100 has approximately 2.6 images per class, while Caltech-256 and ImageNet-1k have around one or fewer. Longer walks are therefore more likely to reach unrelated images, introducing noisy positives.

In contrast, datasets with fewer classes contain larger same-class neighborhoods, allowing longer walks to reach more diverse positives while remaining within the same class. To test this, we conducted a new experiment on CIFAR-10, using the same architecture and training protocol as CIFAR-100. The table reports kk-NN accuracy gains relative to one-step sampling (percentage points):

1-step2-step3-step
+0.00+0.15+0.43

Here, 3-step walks outperform one-step walks, demonstrating that multi-step propagation can indeed be beneficial.

Training statistics further support this explanation. On CIFAR-10, the fraction of training pairs connecting different images of the same class increases from 35% (1 step) to 45% (2 steps) and 49% (3 steps), while class agreement among cross-image positives remains relatively stable (68%, 65%, and 63%). On CIFAR-100, by contrast, only about 20% of cross-image positives share the anchor's class, and a third step introduces no additional same-class positives. We have added the CIFAR-10 experiment and these training statistics to Appendix C.2 of the revised manuscript.

These findings suggest that walk length should reflect the dataset's semantic structure: longer walks can benefit datasets with larger same-class neighborhoods, while one-step walks are more effective on larger, more diverse datasets with many fine-grained classes, such as CIFAR-100, Caltech-256, and ImageNet-1k. This is why we adopted one-step walks as our default configuration, while retaining the flexibility to use longer walks when beneficial.

W2 / Q2 — Consistency of improvements

We thank the reviewer for pointing this out. We clarify that our improvements are primarily measured against each method's corresponding vanilla baseline, rather than necessarily achieving the best result among all SSL methods.

In fact, across Tables 2–5, our approach improves the corresponding baseline in 38 out of 39 comparisons, covering different datasets, evaluation protocols, and SSL objectives. Regarding the two examples:

  • Table 3: Although Barlow Twins achieves 65.81% compared with our 65.62%, the corresponding VICReg baseline achieves 65.18%, meaning our method improves it by 0.44 percentage points.
  • Table 2b: We acknowledge the small decrease from 55.95% to 55.82% at Level 4. However, this difference of just 0.13 points is smaller than the reported standard deviations (0.33 and 0.23), while the other three hierarchy levels all improve.

Importantly, these improvements are not limited to VICReg. Our random walk pairing consistently improves the corresponding vanilla versions of SimCLR, TCL, DINOv2, TIPSv2, and CAPI as well, often with substantial gains. This demonstrates that the benefit extends across different SSL objectives, rather than being specific to a single method.

In addition, we conducted new experiments evaluating random walk pairing across additional downstream datasets and under distribution shift on CIFAR-100-C. These results, presented in our response to reviewer juMX (W3/Q3), provide further evidence of improvements across different datasets, SSL methods, and challenging evaluation conditions.

We hope the new multi-step experiment, the clarification of the matched-baseline comparisons, and the additional generalization results reported in our response to reviewer juMX (W3/Q3) address the reviewer's concerns. We would be happy to answer any further questions and hope the reviewer will consider revisiting their score in light of this evidence.

AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

We thank all three reviewers for their careful evaluations and constructive feedback. We have addressed the concerns raised in our individual responses and uploaded a revised manuscript incorporating the new experiments and clarifications. All new additions are highlighted in red to make them easy to identify.

The main updates, in order of appearance in the revised manuscript, are:

  1. Robustness to image corruptions under distribution shift (Section 5, Table 6): We evaluated existing DINOv2 and TIPSv2 checkpoints on CIFAR-100-C, which applies 19 types of image corruption at different severity levels to clean test images. Random walk pairing improves corrupted-image classification accuracy across all five configurations, with gains of up to 2.43 percentage points in mean corrupted accuracy.

  2. Controlled comparison with NNCLR-style pairing (Appendix B.5): We applied NNCLR-style retrieval within VICReg under matched training conditions. SAG-VICReg outperforms this adaptation across all four hierarchy levels under both k-NN and linear evaluation.

  3. Broader fine-tuning evaluation (Appendix B.8, Tables 18–19): We added 14 configurations across six additional datasets using DINOv2 and TIPSv2, with improvements in all 14 settings.

  4. Walk length and batch-size analysis (Appendix C.2): New CIFAR-10 experiments show that multi-step walks can outperform one-step sampling when same-class neighborhoods are denser. Additional VICReg experiments clarify that the degradation at larger batch sizes on CIFAR-100 is largely related to the optimization budget rather than our pairing mechanism.

We believe these additions substantially strengthen the experimental evidence and directly address the main concerns regarding novelty, robustness, scalability, and generalization.

We appreciate the reviewers' time and consideration, and would be happy to clarify any remaining questions. We hope the reviewers and Area Chair will consider the revised manuscript and additional evidence in their final assessments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.