acceptodds
Under review as a conference paper at ICLR 2027

What Does Post-Training Preserve in Self-Supervised Representations?

Abstract

Self-supervised learning produces transferable visual representations, yet it remains unclear which properties are preserved during post-training. We study how full fine-tuning and LoRA reshape self-supervised vision models across network depth and within their final embedding spaces. Our evaluation covers 39 model variants and more than 64,000 images across seven widely used benchmarks: CIFAR-10, CIFAR-100, DTD, Flowers-102, Food-101, OxfordPets, and Tiny-ImageNet, yielding over two million embedding vectors. Our analysis is threefold. First, we use layerwise CKA to trace representational changes throughout the network. Second, we use embedding-level CKA and Gromov-Wasserstein distance to analyze final-embedding alignment and relational geometry. Third, we assess local neighborhood preservation through nearest-neighbor consistency. Our results reveal recurring stability patterns across the tested model configurations and adaptation targets, while alignment and geometric displacement do not always evolve together. In our rank sweep, the highest tested rank produces substantially earlier layerwise divergence than the lowest tested rank, suggesting that rank can affect the depth-wise localization of adaptation. Overall, our findings indicate that different notions of representational preservation do not always evolve together under post-training. These structured forms of specialization are shaped by model family and adaptation capacity, motivating future work on controlling where and how pretrained representations are reorganized. The code used for all our experiments is released in the supplementary material.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, representation learning, transfer learning, domain adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_M3YtOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work examines the changes in representational structure that occur when tuning pre-trained SSL models. The authors analyze various SSL models and tuning techniques, with metrics that focus on diverse kind of representational structures such as local, global or layer-wise structure. The results expose that the effects of post-training on the internal representational structure vary significantly across settings.

Critical strengths

In general, the methodology is sound, with diverse pre-trained models, datasets, metrics etc. Analyzing LoRA adaptation in vision models is also interesting, given that this tuning technique is largely absent from the visual SSL literature.

Critical weaknesses

The main limitation of this paper is the absence of key take-away. The manuscript provides a lot of technical analyses, but it remains unclear why they may be important to the SSL community. In line with this limitation, the introduction indeed clarifies that the analysis has never been conducted, but does not explain why that matters. Addressing this point likely requires to relate the present analytical results to downstream out-of-distribution/in-distribution accuracy. This could clarify how relevant a tuning technique is, why, and under what conditions.

A myriad of details make the paper hard to follow. For instance:

  • the notion of class purity is not clearly explained in the main paper.
  • the authors employ the word "target" to refer to a downstream "dataset".
  • "[..] across all families", families of what ?
  • The acronym GW is never properly introduced.
  • In the result section, the objective of each paragraph is not properly stated before describing the results.
  • Sometimes, the relevant Figure or Table is not cited prior to its description.
Questions for authors
  • Why do these technical observations matter to the SSL community ?
  • How the representational changes observed during post-training tuning relate to downstream accuracy ?
6 comments
AuthorsOpenReview ↗

Addressing reviewer M3Yt concerns

Key takeaway

We agree that the paper’s central takeaway should have been stated more clearly. Across 114 adapted models, retaining local pretrained neighborhoods outside the adaptation domain consistently predicts retained accuracy, whereas preserving global geometry does not. Our additional analysis supports three conclusions:

  • Local neighborhoods are functionally informative: neighborhood retention correlates positively with retained accuracy in all 42 backbone–dataset groups, with a median Spearman correlation of 0.90, and in 9 of 10 corruption groups.
  • The depth of divergence matters: full fine-tuning changes representations from the earliest blocks, while LoRA preserves more of the backbone and concentrates changes on the adaptation domain.
  • LoRA offers a preservation–capacity trade-off: r=16 exceeds full fine-tuning in target accuracy across all six comparable backbones while retaining more accuracy elsewhere; higher ranks provide no additional accuracy and reduce preservation.

Together, these results indicate that effective adaptation depends not only on how much representations change, but also on where and at what structural scale they change. This can guide future choices of adapter placement, updated components, and LoRA rank. We will make this takeaway explicit in the abstract, introduction, and conclusion of the camera-ready version.

How the representational changes observed during post-training tuning relate to downstream accuracy ?

On non-adaptation datasets, representational preservation predicts retained accuracy. Neighbour retention is positively correlated with linear-probe accuracy changes in 42/42 backbone–dataset groups (median ρ=0.90), k-NN accuracy in 40/42 (ρ=0.83), and robustness under corruption in 9/10. CKA is similarly informative, whereas Gromov–Wasserstein distance is not.

For LoRA, preservation comes at no target-accuracy cost up to r=16; larger ranks reduce both preservation and accuracy. This supports selecting the smallest rank within tolerance of the best target accuracy, with label-free neighbour retention serving as a practical selection signal. Full results are reported below.

BackboneTarget top-1 (%)Linear, off-target (pts)k-NN, off-target (pts)Linear under corruption (pts)
MoCo v3 ViT-B88.9 vs 87.0-2.3 vs -5.9-2.7 vs -7.6-2.8 vs -4.3
MAE ViT-L/1689.2 vs 87.4-3.8 vs -8.1+8.3 vs +5.8-2.8 vs -3.3
data2vec2 ViT-L/1688.8 vs 86.6-0.5 vs -3.3+19.0 vs +13.3-3.5 vs -5.4
I-JEPA ViT-H/1487.3 vs 84.6-4.3 vs -10.1-4.6 vs -15.5-4.6 vs -4.7
DINOv3 ViT-H+/1695.0 vs 94.4-0.3 vs -0.8-0.3 vs -1.1–
CLIP ViT-B/1688.6 vs 85.4-15.5 vs -26.1-23.1 vs -37.7-11.3 vs -11.3
Config.TrainableTarget top-1 (%)Neighbours kept (off)Linear (off, pts)k-NN (off, pts)
r = 10.2M88.5 ± 0.20.62 ± 0.00-0.5 ± 0.1+0.1 ± 0.1
r = 40.6M88.5 ± 0.10.58 ± 0.01-1.1 ± 0.1-0.8 ± 0.3
r = 162.4M89.1 ± 0.30.53 ± 0.00-2.2 ± 0.3-2.5 ± 0.4
r = 649.5M88.0 ± 0.20.43 ± 0.00-5.0 ± 0.0-7.2 ± 0.4
r = 25637.8M82.3 ± 2.20.26 ± 0.05-13.5 ± 3.1-20.0 ± 4.5
Full FT85.8M87.2 ± 0.10.37 ± 0.01-6.2 ± 0.7-8.2 ± 1.4

Clarifications

The notion of class purity is not clearly explained in the main paper. → We acknowledge that the main paper explanation of class purity is quite succinct and, upon acceptance we will use the added space to move the explanation from the appendix in the main paper.

"[..] across all families", families of what? → Here families refer to the different families of vision SSL models analyzed in this paper. We will clarify terminology to improve clarity.

In the result section, the objective of each paragraph is not properly stated before describing the results. → Along the key take-away that will be added to better frame the scope of the full paper, we will add the key intuitions of each paragraph and how they relate overall to each other.

Area_Chair_kfDWOpenReview ↗

Reminder to engage in reviewer-author discussion

Hi M3Yt,

As there are only two days left in the author-reviewer discussion period, I would like to remind you to kindly respond to the authors' rebuttal, posted last week. After November 18, reviewer responses are not be allowed and you will no longer be able to ask any follow-up questions. If you have no follow-ups, a short and polite response acknowledging you have read and understood the rebuttal and have no further questions is sufficient.

Thanks, AC

I thanks the authors for the answer.

Minor thing: why are there two numbers in the first Table.

The authors claim that "Across 114 adapted models, retaining local pretrained neighborhoods outside the adaptation domain consistently predicts retained accuracy, whereas preserving global geometry does not. " However, they write afterwards that "CKA is similarly informative" to Neighborhood retention. It is contradictory, as CKA is a measure of global geometry.

Regarding the correlation: the analysis may also encourage the introduction of methods that specifically promote a specific level of neighborhood retention. For clarification, is the correlation computed with off-target or on-target metrics ? I also wonder if the authors considered LoRA with r=1 and r=4 in the group: their Neighbours kept is high while their accuracy is "low". Could the authors discuss these results ?

Regarding the depth of divergence: there are confounding factors so that it is unclear whether that is the depth of divergence or just LoRA that matter. Doing more experiments with frozen layers (with LoRA and finetuning) could provide more evidence for this claim.

Regarding the preservation-capacity trade-off: that could be particularly interesting if the authors could describe concrete scenarios where off-target evaluations matter, and/or provide evidence showing that post-training can improve performance on the final off-target dataset.

In general, the additional analyses may reveal interesting directions of works. However, the integration of these analyses also requires substantial rewriting of the paper, which we can not review during the rebuttal or iterate on. Therefore, I encourage the authors to complete this analysis and refocus/restructure their study.

Reviewer_f81vOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper empirically studies how full fine-tuning and LoRA reshape self-supervised vision representations across several model families and datasets. The study finds that these notions of preservation do not always evolve together, and that the patterns depend on the adaptation regime, model family, adaptation target, and LoRA rank.

Critical strengths
  • The analysis on multiple perspectives shows that different notions of representational preservation can lead to different conclusions about what adaptation preserves.
  • The study examines several SSL families and adaptation targets, with seed analyses and a LoRA rank sweep.
Critical weaknesses
  • The findings largely align with the existing understanding of LoRA and full fine-tuning, so the new insight and practical guidance offered by the study feel somewhat limited.
  • Key comparisons are confounded by variations in optimization settings, architectures, adapter coverage, and random seeds. It is difficult to isolate whether the observed differences stem from the adaptation regime or the pretraining objective.
Questions for authors
  • Compared to prior analyses of how fine-tuning changes pretrained representations [1, 2, 3], which finding would the authors highlight as a novel insight beyond existing work?
  • Could any of the proposed measures inform a practical decision, such as choosing the LoRA rank or the adaptation regime? For example, does preservation on non-target datasets relate to downstream accuracy or robustness on those datasets?

[1] LoRA Learns Less and Forgets Less (Biderman et al., 2024) [2] Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution (Kumar et al., 2022) [3] LoRA vs Full Fine-tuning: An Illusion of Equivalence (Shuttleworth et al., 2026)

4 comments
AuthorsOpenReview ↗

Addressing reviewer f81v concerns

Novelty beyond existing work

Comparison to cited works

In our introduction, we cite Biderman et al., 2024 and Shuttleworth et al., 2026 works, explaining that we differ from them respectively due to our focus on neighboors representations rather than forgetting as well as internal representations instead of weight-space structures (l39-41). Similarly, in our related work section we cite Kumar et al., 2022 work (l87) and we differ from it as our work is not a proposition for an accuracy improvement method, rather an analysis of the degree of preservation of internal representation, measured through various analysis (neighbors, geometry, embeddings, etc.). We will provide improved distinction in the camera-ready version upon acceptance.

Key takeaway

We agree that the paper’s central takeaway should have been stated more clearly. Across 114 adapted models, retaining local pretrained neighborhoods outside the adaptation domain consistently predicts retained accuracy, whereas preserving global geometry does not. Neighborhood retention correlates positively with retained accuracy in all 42 backbone–dataset groups, with a median Spearman correlation of 0.90, and in 9 of 10 corruption groups.

Could any of the proposed measures inform a practical decision, such as choosing the LoRA rank or the adaptation regime?

Yes. On non-adaptation datasets, representational preservation predicts retained accuracy. Neighbour retention is positively correlated with linear-probe accuracy changes in 42/42 backbone–dataset groups (median ρ=0.90), k-NN accuracy in 40/42 (ρ=0.83), and robustness under corruption in 9/10. CKA is similarly informative, whereas Gromov–Wasserstein distance is not.

For LoRA, preservation comes at no target-accuracy cost up to r=16; larger ranks reduce both preservation and accuracy. This supports selecting the smallest rank within tolerance of the best target accuracy, with label-free neighbour retention serving as a practical selection signal. Full results are reported below.

BackboneTarget top-1 (%)Linear, off-target (pts)k-NN, off-target (pts)Linear under corruption (pts)
MoCo v3 ViT-B88.9 vs 87.0-2.3 vs -5.9-2.7 vs -7.6-2.8 vs -4.3
MAE ViT-L/1689.2 vs 87.4-3.8 vs -8.1+8.3 vs +5.8-2.8 vs -3.3
data2vec2 ViT-L/1688.8 vs 86.6-0.5 vs -3.3+19.0 vs +13.3-3.5 vs -5.4
I-JEPA ViT-H/1487.3 vs 84.6-4.3 vs -10.1-4.6 vs -15.5-4.6 vs -4.7
DINOv3 ViT-H+/1695.0 vs 94.4-0.3 vs -0.8-0.3 vs -1.1–
CLIP ViT-B/1688.6 vs 85.4-15.5 vs -26.1-23.1 vs -37.7-11.3 vs -11.3
Config.TrainableTarget top-1 (%)Neighbours kept (off)Linear (off, pts)k-NN (off, pts)
r = 10.2M88.5 ± 0.20.62 ± 0.00-0.5 ± 0.1+0.1 ± 0.1
r = 40.6M88.5 ± 0.10.58 ± 0.01-1.1 ± 0.1-0.8 ± 0.3
r = 162.4M89.1 ± 0.30.53 ± 0.00-2.2 ± 0.3-2.5 ± 0.4
r = 649.5M88.0 ± 0.20.43 ± 0.00-5.0 ± 0.0-7.2 ± 0.4
r = 25637.8M82.3 ± 2.20.26 ± 0.05-13.5 ± 3.1-20.0 ± 4.5
Full FT85.8M87.2 ± 0.10.37 ± 0.01-6.2 ± 0.7-8.2 ± 1.4

Key comparisons are confounded by variations in optimization settings, architectures, adapter coverage, and random seeds

For the key comparisons, we have:

  • (i) regime vs objective: every regime comparison is parent-child on one backbone, the regime effect has the same sign on five backbones with five objectives
  • (ii) optimization: the learning rate sweep, the depth gap holds at a matched rate
  • (iii) the Adapter coverage: Dinov3 retrained with LoRA on attention and MLP. Accuracy is unchanged (94.8 vs 95.0), trunks CKA drops from 0.971 to 0.953 and final-layer CKA from 0.905 to 0.869 . Dinov3 stays the most preserved family.
Area_Chair_kfDWOpenReview ↗

Reminder to engage in reviewer-author discussion

Hi f81v,

As there are only two days left in the author-reviewer discussion period, I would like to remind you to kindly respond to the authors' rebuttal, posted last week. After November 18, reviewer responses are not be allowed and you will no longer be able to ask any follow-up questions. If you have no follow-ups, a short and polite response acknowledging you have read and understood the rebuttal and have no further questions is sufficient.

Thanks, AC

Thank you for the detailed response and the additional experiments. The DINOv3 retraining and the matched learning rate analysis address part of my concerns. However, my main concern about novelty remains. The new key takeaway stating that off-target neighborhood retention tracks retained accuracy still seems closely aligned with prior findings. I will therefore keep my original score.

AuthorsOpenReview ↗

Follow-up on novelty

Thank you for the response. Could the reviewer please identify the specific prior work establishing that pretrained local-neighborhood retention predicts retained accuracy on datasets not used for adaptation? The closest works we identified concern forgetting under LoRA (Biderman et al., 2024),, feature distortion and OOD performance (Kumar et al., 2022), weight-space changes and forgetting (Shuttleworth et al., 2024), or general representational similarity (Ding et al., 2021), but do not appear to establish this relationship across self-supervised vision families or contrast neighborhood retention with alignment and relational geometry. We would be glad to add any missing references and clarify the distinction.

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

This paper investigates how full fine-tuning and Low-Rank Adaptation (LoRA) reorganize self-supervised visual representations across network depth, final-embedding geometry, and local neighborhoods. The study, conducted across 39 model variants and using over 64,000 images across seven benchmarks reveals that different notions of representational preservation do not always evolve together. LoRA maintains high layer-wise similarity much deeper into the network than full fine-tuning, while full fine-tuning causes measurable divergence starting from the earliest blocks. Post-training universally reduces global CKA alignment with pretrained models, but its effect on relational geometry varies by family, with MoCo v3 showing the strongest overall stability. Finally, structural changes concentrate significantly on the target dataset, where local neighborhoods become less stable but significantly more class-coherent

Critical strengths

These are the paper's strengths:

  • a very comprehensive empirical evaluation: 39 model variants across 7 widely used benchmarks and over 64,000 images
  • it adopts a multi-dimensional analytical framework consisting of layer-wise CKA, embedding-level CKA, Gromov-Wasserstein (GW) distance, and nearest-neighbor metrics
  • it uses a diverse selection of SSL families: I-JEPA, MoCo v3, DINOv3, data2vec2, and MAE, spanning contrastive, masked modeling, and feature prediction paradigms
  • it provides a rigorous parent–child comparison design in which the underlying pretrained backbone is held fixed when assessing adaptation variants, preventing architectural or pretraining differences from confounding the post-training analysis.
  • clearly documents training protocols, hyperparameter selection, hardware setups, and releases code in order to guarantee full reproducibility of the results
Critical weaknesses

These are the paper's weaknesses:

  • the direct comparison between LoRA and full fine-tuning is restricted to only two backbones (MoCo v3 ViT-B and MAE ViT-L/16), limiting the generalizability to other backbone architectures
  • the analysis is limited to image classification only, excluding other relevant computer vision tasks like object detection, segmentation, etc.
  • the analysis is focused on vision-centric tasks (e.g. image classification) and does not consider if these representational preservation patterns also apply to multimodal settings, like VLM (e.g., CLIP)
  • the analysis shows where and how much representations shift, but doesn't offer an explanation of why this happens
  • DINOv3 only adapts the attention parts of the model, while other models adapt both attention and multi-layer perceptron (MLP) blocks, making direct comparisons between different model families unfair or unclear
Questions for authors

These are my concerns:

  • please provide the direct comparison data for LoRA vs. full fine-tuning across all five SSL families instead of just two backbones
  • please clarify why did you use different learning rates for LoRA and full fine-tuning instead of keeping the training setup identical to isolate the effect of the low-rank updates
  • include multiple training seeds for the rank sweeps and dataset-specific nearest-neighbor analyses instead of relying on a single seed, in order to ensure statistical rigor and reproducibility
  • could you provide a more mechanistic explanation (e.g., tracking specific attention heads or pathways) for why representations diverge in early versus late layers?
  • please extend your analysis to other computer vision tasks, like object detection and segmentation
  • could you show if these representational preservation patterns also apply to multimodal settings, like VLM (e.g., CLIP)?
3 comments
AuthorsOpenReview ↗

Addressing Reviewer iHds concerns

W1/Q1: More backbones for direct comparison between LoRA and full fine-tuning

We fully fine-tuned data2vec2, I-Jepa, and DINOv3 on the three adaptation datasets with the protocol of App. D and we added CLIP under both regimes. The regime comparison now covers all five SSL families, plus CLIP.

BackboneRegime= 0.99= 0.95= 0.90Trunk CKABlocks > 0.95Neighbours kept (adapt. / other)Target top-1 (%)
MoCo v3 ViT-BLoRA0.150.70 (1)0.82 (4)0.94769%0.49 / 0.6788.9
Full FT0.000.000.000.8290%0.46 / 0.5687.0
MAE ViT-L/16LoRA0.400.810.86 (2)0.94677%0.22 / 0.3389.2
Full FT0.000.240.48 (1)0.89740%0.21 / 0.2887.4
data2vec2 ViT-L/16LoRA0.050.270.540.84926%0.12 / 0.1688.8
Full FT0.000.000.000.7753%0.11 / 0.1586.6
I-JEPA ViT-H/14LoRA0.090.500.740.90048%0.31 / 0.4587.3
Full FT0.000.000.010.8060%0.26 / 0.3484.6
DINOv3 ViT-H+/16LoRA (attn)0.290.71 (4)0.87 (6)0.97190%0.62 / 0.8095.0
LoRA (attn + MLP)0.130.16 (2)0.34 (4)0.95373%0.59 / 0.7794.8
Full FT0.000.200.14 (2)0.90651%0.57 / 0.7494.4
CLIP ViT-B/16LoRA0.010.320.490.79430%0.36 / 0.3188.6
Full FT0.000.000.000.6450%0.32 / 0.1885.4

W2-3/Q5-6: Analysis limited to image classification only

We acknowledge that extending our evaluation framework is key to further our analysis. However, as stated in your strengths, our current evaluation framework is already rather wide and we believe that extending it further would actually result in reduced readability. For these reasons, other tasks than image classification are out of the scope of this work and left for future research as a natural extension.

W4/Q4: Explanation of why representations shift

An intuition would be that full fine-tuning propagates error throughout the whole network thus even modifying the earliest layers while the "low-rank" aspect of LoRA might filter out components that are less significant thus focusing the changes in the more fine-grained features of the latest layers.

Q2: Different learning rates for LoRA and full fine-tuning

Each regime is trained at its own optimal learning rate, selected by a sweep on MoCov3, ViT-B (App. D.2). A shared rate would not isolate the low-rank update: the two optimal differ by an order of magnitude, so at any shared rate one regime is far from its optimum and the comparison mostly reflects that mistunning. The sweep is referred to App. D2 but was missing. We provide it below:

RegimeLearning rateTarget top-1 (%)Trunk CKA (on-target)First crossing = 0.95
Full FT10⁻⁵45.00.9760.00 (6)
Full FT10⁻⁴ (used)87.00.7980.00
Full FT10⁻³74.40.5690.00
LoRA10⁻⁴85.80.9010.40
LoRA10⁻³ (used)88.90.9110.70 (1)
LoRA10⁻²4.2 (diverged)––

The depth-wise difference does not come from the learning rate. The only full fine-tuning run that preserves as much as LoRA (10e-5, trunk CKA 0.976) learns almost nothing (45%). The paper compares the regimes at their respective optima, as App. D2 states.

Q3: Multiple training seeds for the rank sweeps and dataset-specific nearest-neighbor analyses

Every configuration of the MoCov3 rank sweep (Sec. 6) now has three training seeds: r = 16 and full fine-tuning already has seeds 43 and 44 (Tab.8) and we trained them for r=1, 4, 64 and 256. Table below reports means and std over the three seeds :

Config.TrainableTarget top-1 (%)Trunk CKAFinal CKANeighbours kept (adapt.)Neighbours kept (other)= 0.95 crossing
r = 10.2M88.5 ± 0.20.942 ± 0.0040.820 ± 0.0100.574 ± 0.0050.743 ± 0.0070.73
r = 40.6M88.5 ± 0.10.923 ± 0.0030.769 ± 0.0080.532 ± 0.0090.711 ± 0.0100.68
r = 162.4M89.1 ± 0.30.912 ± 0.0010.723 ± 0.0080.487 ± 0.0020.672 ± 0.0010.71
r = 649.5M88.0 ± 0.20.875 ± 0.0020.643 ± 0.0070.447 ± 0.0000.597 ± 0.0030.42
r = 25637.8M82.3 ± 2.20.822 ± 0.0170.545 ± 0.0190.416 ± 0.0200.401 ± 0.0670.14
Full FT85.8M87.2 ± 0.10.798 ± 0.0040.691 ± 0.0050.456 ± 0.0040.559 ± 0.0080.00

Seed variation is small relative to the effect of rank: up to r = 64, the std. is at most 0.33 points for top-1 and 0.004 for trunk CKA. Every trend of Sec. 6 holds in each seed separately: trunk CKA, final-layer CKA and neighbours kept fall monotonically with rank in all three seeds, target accuracy peaks at r = 16 in all three, and r = 256 is the least accurate configuration in every seed.

Area_Chair_kfDWOpenReview ↗

Reminder to engage in reviewer-author discussion

Hi iHds,

As there are only two days left in the author-reviewer discussion period, I would like to remind you to kindly respond to the authors' rebuttal, posted a week ago. After November 18, reviewer responses are not be allowed and you will no longer be able to ask any follow-up questions. If you have no follow-ups, a short and polite response acknowledging you have read and understood the rebuttal and have no further questions is sufficient.

Thanks, AC

Reviewer_iHdsOpenReview ↗

Response to authors

I want to thank the authors for the comprehensive rebuttal and for addressing the questions regarding multiple backbones comparisons, different learning rates, and multi-seed evaluations. However, my other concerns have not been addressed: the evaluation is still restricted to image classification only, and the explanation behind the layer-wise representation shifts (divergence) remains intuitive. Therefore, I have decided to maintain my initial rating.

Comments on the paper

4 comments

Dear Reviewers,

As the discussion period is nearing its end, we would appreciate your feedback on our responses.

We have completed the requested experiments and provided additional evidence across backbones, learning rates, and seeds, added analyses addressing potential confounds, clarified our key takeaways, and pointed to the specific manuscript lines that distinguish our work from the prior works cited in the reviews.

Given the scope of the additional experiments and analyses provided, we would particularly appreciate confirmation of whether our responses address your concerns, or guidance on any remaining questions.

I would like to remind the authors that they can post a rebuttal version of their paper using "Edit > Rebuttal version" on this page. I encourage them to incorporate all the changes they have made during the rebuttal period into the paper, and to highlight any pertinent changes they would like reviewers to note using coloured text, margin notes, or the changebar package. This will make it easier for reviewers to see how the paper has ended up as a result of their changes, and for the AC to assess both the final version of the paper and how it was changed during the rebuttal period.

Rebuttal versions can be posted multiple times (you do not need to only post one version) and you can not post a rebuttal version after the author-reviewer discussion period closes at the end of 18 Nov.

Thanks, AC

AuthorsOpenReview ↗

Thanks for the reminder

Dear Area Chair,

We have uploaded a revised version incorporating the requested minor changes and the clarifications arising from the discussion. These include clearer key takeaway (abstract, introduction and conclusion), expanded LoRA versus full fine-tuning (Tab.1) and rank results (Tab. 4), the learning-rate control (Tab. 7), a clearer definition of class purity (Sec. 3.2(4)), and improved formatting in the Results (Sec 5) with an introduction paragraph at each sub-section. Relevant changes are highlighted.

We waited for the remaining rebuttals to further evolve so that we could incorporate the relevant feedback in one consolidated version, rather than ask the reviewers and AC to check several successive revisions. Our apologies for the inconvenience.

Thank you.

No inconvenience to me, I just wanted to ensure you remembered to update it before the window closed. I would generally recommend uploading a revised version with the rebuttal so reviewers can check it when reading the rebuttal, and then you can revise the paper further if there are additional changes to be made. Reviewers are not notified about each update to the rebuttal version, so you don't have to explain why you made an upload to fix each minor change; only explain significant updates which you actually want reviewers to be aware of (as you have done above).

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.