pith. sign in

arxiv: 2605.23028 · v1 · pith:LLPISCM2new · submitted 2026-05-21 · 💻 cs.LG · cs.CL· cs.CV

RADAR: Relative Angular Divergence Across Representations

Xavier Cadet , Mateusz Nowak , Peter Chin This is my paper

Pith reviewed 2026-05-25 05:37 UTC · model grok-4.3

classification 💻 cs.LG cs.CLcs.CV
keywords transferability estimationfoundation modelsrepresentation trajectoriescross-domain transferangular divergencelayer-wise dynamicsnegative transfer
0
0 comments X

The pith

RADAR estimates cross-domain transferability by measuring divergence between within-domain and cross-domain representation trajectories.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes RADAR as a metric that tracks how a foundation model's internal representations change from layer to layer. It does this by calculating angular alignments and relative distance shifts along those layer-to-layer paths, then compares the statistical distribution of changes seen inside one domain against changes seen when moving from one domain to another. The central hypothesis is that smaller divergence between these two distributions signals higher likelihood of successful transfer. The authors test this on cross-lingual text classification and cross-domain image classification, finding competitive accuracy with prior metrics and stronger results when domains shift gradually or stay cleanly separated. This matters because it offers a way to choose source data that avoids performance drops from negative transfer without exhaustive trial fine-tuning.

Core claim

RADAR analyzes the layer-wise evolution of representations by measuring angular alignments and relative changes in distance along layer-to-layer displacement trajectories, and by comparing empirical distributions of within-domain and cross-domain dynamics. Domain transferability is hypothesized to relate to the divergence between these trajectory distributions. Across vision and text benchmarks the metric yields competitive predictive performance, with particularly strong results when domain transitions are smooth or cleanly separated.

What carries the argument

RADAR metric: divergence between empirical distributions of within-domain versus cross-domain layer-to-layer angular alignments and relative distance changes in representation trajectories.

If this is right

  • Practitioners can rank candidate source domains by RADAR score before fine-tuning to reduce negative transfer.
  • The metric supplies a geometric signal that can be computed from unlabeled data in both source and target domains.
  • Different modalities appear to favor different topological versions of the same trajectory comparison.
  • Representation-space geometry becomes a measurable factor in deciding whether a foundation model will adapt well.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same trajectory-divergence idea could be applied to decide which layers to freeze or adapt during transfer.
  • If the hypothesis holds, models whose internal trajectories are already similar across domains might be preferred at training time.
  • The approach invites direct comparison against other geometry-based transfer predictors that do not track layer-wise paths.

Load-bearing premise

Transfer success between domains is determined by how closely the patterns of layer-wise representation changes match inside one domain versus across domains.

What would settle it

A controlled test on paired domains where trajectory-distribution divergence is low yet fine-tuning still produces negative transfer, or divergence is high yet transfer succeeds.

Figures

Figures reproduced from arXiv: 2605.23028 by Mateusz Nowak, Peter Chin, Xavier Cadet.

Figure 1
Figure 1. Figure 1: Overview of the RADAR framework. RADAR identifies optimal source data for transfer learning by extracting layer-wise (ℓ) geometric trajectories from pre-trained models. By computing the divergence in angle (θ) and relative distance (d) densities between within- and cross-domain pairs, it efficiently ranks auxiliary candidates without requiring computationally expensive fine-tuning. We evaluate RADAR across… view at source ↗
Figure 2
Figure 2. Figure 2: Geometric triangle of displacement vectors spanning layers [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualizing ImageNet-C synthetic corruptions. Examples of various visual corruptions across increasing severities (0, 1, 3, and 5). The profound structural degradation at Severity 5 fundamentally scrambles patch-level token distributions in Vision Transformers. As shown in [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of Layer 0 centroid distances across datasets for the CLIP model. Distance matrices illustrate the spatial separation between domains right at the input embedding level. Warmer hues denote smaller centroid distances between domains, whereas cooler hues denote larger separations. At Layer 0, the domain distance distributions of DomainNet and OfficeHome closely mirror the low-severity corruptio… view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of Layer 0 centroid distances across datasets for the DINOv3 model. Distance matrices illustrate the spatial separation between domains right at the input embedding level. Warmer hues denote smaller centroid distances between domains, whereas cooler hues denote larger separations. At Layer 0, the domain distance distributions of DomainNet and OfficeHome closely mirror the ImageNet-C-1 and Ima… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of sample size (N) on RADAR metric stability. The plots illustrate the Spearman rank correlation (ρ) of RADAR computed at various pair-wise sub-sample sizes (N ∈ {16, 32, 64, . . . , 65536}) relative to a high-fidelity baseline computed with N = 65536 on the Do￾mainNet [35] dataset. We observe that for sample sizes N ≤ 128, the CLIP model maintains a Spearman correlation exceeding 0.88, which increa… view at source ↗
Figure 7
Figure 7. Figure 7: Effect of the window size (ℓ ∈ {1, 2, 3, 4, 5, 6}) on RADAR metric stability. The plots illustrate the Spearman rank correlation (ρ) of RADAR computed at various window sizes relative to a high-fidelity baseline computed with ℓ = 6 on the DomainNet [35] dataset. (a) CLIP (b) DINOv3 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of the temperature τ (τ ∈ {0.125, 0.25, 0.5, 1, 2, 4, 8}) on RADAR metric stability. The plots illustrate the Mean Correlation Improvement (MCI) of the RADAR metric computed across various weighting temperatures on the DomainNet [35] dataset. detrimental. In contrast, the DINOv3 model remains remarkably robust across the 0.125–8.0 range, with a total MCI variation of only 2.0 percentage points (pp.)… view at source ↗
read the original abstract

Machine learning methods rely on data. However, gathering suitable data can be challenging due to availability constraints, cost, or the need for domain expertise. Expanding datasets with additional sources is a common response to limited data, yet this practice does not always improve downstream performance and can sometimes lead to a loss of performance, known as negative transfer. We propose RADAR, a simple, geometrically grounded metric for estimating cross-domain transferability in foundation models. RADAR analyzes the layer-wise evolution of representations by measuring angular alignments and relative changes in distance along layer-to-layer displacement trajectories, and by comparing empirical distributions of within-domain and cross-domain dynamics. We hypothesize that domain transferability is related to the divergence between these trajectory distributions. We evaluate the metric across multiple modalities, including cross-lingual sentiment classification with text embedding models and cross-domain image classification with foundation vision models. Across several settings, RADAR provides competitive predictive performance relative to existing transferability metrics on several vision and text benchmarks, with particularly strong results when domain transitions are smooth or cleanly separated. Our ablations further suggest that the effectiveness of transferability estimation depends on the geometry of the model's internal representation space, with different modalities favoring different topological formulations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper proposes RADAR, a geometrically grounded metric for predicting cross-domain transferability of foundation models. It computes a divergence score between empirical distributions of within-domain versus cross-domain layer-wise representation trajectories, where trajectories are characterized by angular alignments and relative distance changes. The central hypothesis is that lower divergence predicts better transfer performance. The metric is evaluated on cross-lingual sentiment classification with text embeddings and cross-domain image classification with vision foundation models, where it shows competitive correlation with observed transfer performance relative to prior metrics, with stronger results on smooth or cleanly separated domain shifts. Ablations indicate that effectiveness depends on the geometry of the model's representation space and that different modalities benefit from different topological formulations of the metric.

Significance. If the empirical results hold under full scrutiny, RADAR supplies a simple, interpretable, and modality-aware tool for anticipating negative transfer when augmenting limited datasets. Its explicit grounding in layer-wise trajectory geometry distinguishes it from purely output-based or parameter-count-based predictors and offers a falsifiable link between internal representation dynamics and downstream transfer. The multi-modality evaluation and ablation on topological variants add value by highlighting when geometric assumptions matter.

major comments (3)
  1. [§3.2] §3.2, divergence estimator: the manuscript must specify the exact distance or divergence used to compare the two empirical trajectory distributions (e.g., Wasserstein, KL, or MMD) and whether any bandwidth or binning hyper-parameters are required; without this, it is impossible to verify that the metric is parameter-free as claimed in the abstract.
  2. [Table 2, §4.1] Table 2 and §4.1, correlation tables: the reported Pearson/Spearman values for RADAR versus baselines are competitive, but the paper should report the number of domain pairs, the exact layer sampling procedure, and whether the same layers are used for all models; small differences in correlation could be driven by inconsistent layer selection rather than the metric itself.
  3. [§4.3] §4.3, modality-specific geometry claim: the ablation concludes that vision and text favor different topological formulations, yet only two modalities and a limited set of models are tested; the central claim that “effectiveness depends on the geometry of the model’s internal representation space” therefore rests on a narrow empirical base and requires either broader model coverage or a clearer theoretical justification.
minor comments (3)
  1. [Abstract, §2] The abstract and §2 should cite the specific prior transferability metrics (LogME, H-score, etc.) against which RADAR is compared so readers can immediately locate the baselines.
  2. [§3.1] Notation for angular alignment and relative distance change should be defined once in §3.1 with consistent symbols; several equations reuse the same symbols for different quantities.
  3. [Figure 3] Figure 3 caption should state the exact number of layers sampled and whether the trajectories are normalized per model or per domain pair.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback and positive overall assessment. We address each major comment below and have revised the manuscript accordingly to improve clarity and reporting.

read point-by-point responses

  1. Referee: [§3.2] §3.2, divergence estimator: the manuscript must specify the exact distance or divergence used to compare the two empirical trajectory distributions (e.g., Wasserstein, KL, or MMD) and whether any bandwidth or binning hyper-parameters are required; without this, it is impossible to verify that the metric is parameter-free as claimed in the abstract.

    Authors: We agree this specification was missing. RADAR computes the divergence as the 2-Wasserstein distance between the two empirical distributions of trajectory features (angular alignments and relative distance changes), obtained via the exact optimal transport solution on the finite sample sets. No binning, kernels, or bandwidth parameters are used. We have updated §3.2 with this explicit description to confirm the metric is parameter-free. revision: yes

  2. Referee: [Table 2, §4.1] Table 2 and §4.1, correlation tables: the reported Pearson/Spearman values for RADAR versus baselines are competitive, but the paper should report the number of domain pairs, the exact layer sampling procedure, and whether the same layers are used for all models; small differences in correlation could be driven by inconsistent layer selection rather than the metric itself.

    Authors: We appreciate the request for additional experimental details. We have revised §4.1 and the caption of Table 2 to state that the vision experiments use 15 domain pairs and the text experiments use 12 domain pairs. All layers of each model are sampled, with layer indices normalized by total depth to ensure the same relative positions are compared across models of different architectures. revision: yes

  3. Referee: [§4.3] §4.3, modality-specific geometry claim: the ablation concludes that vision and text favor different topological formulations, yet only two modalities and a limited set of models are tested; the central claim that “effectiveness depends on the geometry of the model’s internal representation space” therefore rests on a narrow empirical base and requires either broader model coverage or a clearer theoretical justification.

    Authors: We partially concur that the empirical base is limited to two modalities. However, the ablations already cover multiple models within each modality and demonstrate consistent differences in topological variant performance. We have revised §4.3 to add a short theoretical paragraph linking angular trajectory divergence to the intrinsic geometry of representation manifolds and to qualify the claim as an observation supported by the current experiments rather than a general assertion. Extending to additional modalities is left for future work. revision: partial

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper introduces RADAR as an explicitly defined geometric metric based on angular alignments and relative distance changes along layer trajectories, then compares within-domain vs. cross-domain distribution divergences to observed transfer performance. This is an empirical proposal with no equations reducing a claimed prediction to a fitted input by construction, no load-bearing self-citations, and no ansatz or uniqueness claim imported from prior author work. The derivation chain is self-contained as a new metric whose validity is tested externally on benchmarks rather than asserted by internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities can be identified from the provided text.

pith-pipeline@v0.9.0 · 5739 in / 940 out tokens · 41561 ms · 2026-05-25T05:37:26.434464+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

52 extracted references · 52 canonical work pages · 5 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems33, 21428–21439 (2020)

    Alvarez-Melis, D., Fusi, N.: Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems33, 21428–21439 (2020)

    work page 2020

  2. [2]

    In: 2019 IEEE international conference on image processing (ICIP)

    Bao, Y ., Li, Y ., Huang, S.L., Zhang, L., Zheng, L., Zamir, A., Guibas, L.: An information- theoretic approach to transferability in task transfer learning. In: 2019 IEEE international conference on image processing (ICIP). pp. 2309–2313. IEEE (2019)

    work page 2019

  3. [3]

    Perception Encoder: The best visual embeddings are not at the output of the network

    Bolya, D., Huang, P.Y ., Sun, P., Cho, J.H., Madotto, A., Wei, C., Ma, T., Zhi, J., Rajasegaran, J., Rasheed, H., et al.: Perception encoder: The best visual embeddings are not at the output of the network. arXiv preprint arXiv:2504.13181 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    Dong, H., Liu, M., Zhou, K., Chatzi, E., Kannala, J., Stachniss, C., Fink, O.: Advances in multimodal adaptation and generalization: From traditional approaches to foundation models. IEEE Transactions on Pattern Analysis and Machine Intelligence (2026)

    work page 2026

  5. [5]

    Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp

    Farahani, A., V oghoei, S., Rasheed, K., Arabnia, H.R.: A brief review of domain adaptation. Advances in data science and information engineering: proceedings from ICDATA 2020 and IKE 2020 pp. 877–894 (2021)

    work page 2020

  6. [6]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Feydy, J., Roussillon, P., Trouvé, A., Gori, P.: Fast and scalable optimal transport for brain tractograms. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 636–644. Springer (2019)

    work page 2019

  7. [7]

    ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik21(6), 364–379 (1941)

    Gebelein, H.: Das statistische problem der korrelation als variations-und eigenwertproblem und sein zusammenhang mit der ausgleichsrechnung. ZAMM-Journal of Applied Mathematics and Mechanics/Zeitschrift für Angewandte Mathematik und Mechanik21(6), 364–379 (1941)

    work page 1941

  8. [8]

    Nature communications12(1), 5982 (2021)

    Hénaff, O.J., Bai, Y ., Charlton, J.A., Nauhaus, I., Simoncelli, E.P., Goris, R.L.: Primary visual cortex straightens natural video trajectories. Nature communications12(1), 5982 (2021)

    work page 2021

  9. [9]

    Nature neuroscience22(6), 984–991 (2019)

    Hénaff, O.J., Goris, R.L., Simoncelli, E.P.: Perceptual straightening of natural videos. Nature neuroscience22(6), 984–991 (2019)

    work page 2019

  10. [10]

    Proceedings of the International Conference on Learning Representations (2019)

    Hendrycks, D., Dietterich, T.: Benchmarking neural network robustness to common corruptions and perturbations. Proceedings of the International Conference on Learning Representations (2019)

    work page 2019

  11. [11]

    Mathemat- ical Proceedings of the Cambridge Philosophical Society31(4), 520–524 (1935)

    Hirschfeld, H.O.: A connection between correlation and contingency. Mathemat- ical Proceedings of the Cambridge Philosophical Society31(4), 520–524 (1935). https://doi.org/10.1017/S0305004100013517

    work page doi:10.1017/s0305004100013517 1935

  12. [12]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Huang, L., Cao, X., Lu, H., Meng, Y ., Yang, F., Liu, X.: Mind the gap: Preserving and compensating for the modality gap in clip-based continual learning. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3777–3786 (2025)

    work page 2025

  13. [13]

    In: International conference on machine learning

    Huang, L.K., Huang, J., Rong, Y ., Yang, Q., Wei, Y .: Frustratingly easy transferability estimation. In: International conference on machine learning. pp. 9201–9225. PMLR (2022) 10

    work page 2022

  14. [14]

    https://huggingface.co/google/ embeddinggemma-300m(2026), accessed: 2026

    Hugging Face: Embeddinggemma model card. https://huggingface.co/google/ embeddinggemma-300m(2026), accessed: 2026

    work page 2026

  15. [15]

    https://huggingface.co/openai/ clip-vit-base-patch32(2026), accessed: 2026

    Hugging Face: Model card: Clip. https://huggingface.co/openai/ clip-vit-base-patch32(2026), accessed: 2026

    work page 2026

  16. [16]

    https://huggingface.co/facebook/ dinov3-vits16-pretrain-lvd1689m(2026), accessed: 2026

    Hugging Face: Model card for dinov3. https://huggingface.co/facebook/ dinov3-vits16-pretrain-lvd1689m(2026), accessed: 2026

    work page 2026

  17. [17]

    https://huggingface.co/Qwen/ Qwen3-Embedding-0.6B(2026), accessed: 2026

    Hugging Face: Qwen3-embedding-0.6b. https://huggingface.co/Qwen/ Qwen3-Embedding-0.6B(2026), accessed: 2026

    work page 2026

  18. [18]

    In: Forty-first International Conference on Machine Learning (2024)

    Huh, M., Cheung, B., Wang, T., Isola, P.: Position: The platonic representation hypothesis. In: Forty-first International Conference on Machine Learning (2024)

    work page 2024

  19. [19]

    In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases

    Ibrahim, S., Ponomareva, N., Mazumder, R.: Newer is not always better: Rethinking transfer- ability metrics, their peculiarities, stability and performance. In: Joint European Conference on Machine Learning and Knowledge Discovery in Databases. pp. 693–709. Springer (2022)

    work page 2022

  20. [20]

    In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum?id=LsmUgStXby

    Internò, C., Geirhos, R., Olhofer, M., Liu, S., Hammer, B., Klindt, D.: AI-generated video detec- tion via perceptual straightening. In: The Thirty-ninth Annual Conference on Neural Information Processing Systems (2025),https://openreview.net/forum?id=LsmUgStXby

    work page 2025

  21. [21]

    arXiv preprint arXiv:2201.05867 (2022)

    Jiang, J., Shu, Y ., Wang, J., Long, M.: Transferability in deep learning: A survey. arXiv preprint arXiv:2201.05867 (2022)

    work page arXiv 2022

  22. [22]

    https://github

    Junguang, J., Baixu, C., Bo, F., Mingsheng, L.: Transfer-learning-library. https://github. com/thuml/Transfer-Learning-Library(2020)

    work page 2020

  23. [23]

    Kempf, E., Schrodi, S., Argus, M., Brox, T.: When and how does clip enable domain and compositional generalization? arXiv preprint arXiv:2502.09507 (2025)

    work page arXiv 2025

  24. [24]

    arXiv preprint arXiv:2602.10099 (2026)

    Kumar, A., Patel, V .M.: Learning on the manifold: Unlocking standard diffusion transformers with representation encoders. arXiv preprint arXiv:2602.10099 (2026)

    work page arXiv 2026

  25. [25]

    Li, D., Yang, Y ., Song, Y .Z., Hospedales, T.M.: Deeper, broader and artier domain generalization (2017),https://arxiv.org/abs/1710.03077

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  26. [26]

    Advances in Neural Information Processing Systems35, 17612–17625 (2022)

    Liang, V .W., Zhang, Y ., Kwon, Y ., Yeung, S., Zou, J.Y .: Mind the gap: Understanding the modal- ity gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems35, 17612–17625 (2022)

    work page 2022

  27. [27]

    Neural Networks181, 106796 (2025)

    Liu, X., Bai, Y ., Lu, Y ., Soltoggio, A., Kolouri, S.: Wasserstein task embedding for measuring task similarities. Neural Networks181, 106796 (2025)

    work page 2025

  28. [28]

    In: International conference on machine learning

    Long, M., Cao, Y ., Wang, J., Jordan, M.: Learning transferable features with deep adaptation networks. In: International conference on machine learning. pp. 97–105. PMLR (2015)

    work page 2015

  29. [29]

    IEEE transactions on pattern analysis and machine intelligence 29(9), 1546–1562 (2007)

    Ma, Y ., Derksen, H., Hong, W., Wright, J.: Segmentation of multivariate mixed data via lossy data coding and compression. IEEE transactions on pattern analysis and machine intelligence 29(9), 1546–1562 (2007)

    work page 2007

  30. [30]

    In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/forum?id= VVVfuIcmKR

    Mistretta, M., Baldrati, A., Agnolucci, L., Bertini, M., Bagdanov, A.D.: Cross the gap: Exposing the intra-modal misalignment in CLIP via modality inversion. In: The Thirteenth International Conference on Learning Representations (2025), https://openreview.net/forum?id= VVVfuIcmKR

    work page 2025

  31. [31]

    In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics

    Muennighoff, N., Tazi, N., Magne, L., Reimers, N.: Mteb: Massive text embedding bench- mark. In: Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics. pp. 2014–2037 (2023)

    work page 2014

  32. [32]

    In: International conference on machine learning

    Nguyen, C., Hassner, T., Seeger, M., Archambeau, C.: Leep: A new measure to evaluate transferability of learned representations. In: International conference on machine learning. pp. 7294–7305. PMLR (2020) 11

    work page 2020

  33. [33]

    arXiv preprint arXiv:2501.18901 (2025)

    Nguyen, K., Nguyen, H., Pham, T., Ho, N.: Lightspeed geometric dataset distance via sliced optimal transport. arXiv preprint arXiv:2501.18901 (2025)

    work page arXiv 2025

  34. [34]

    In: Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies (NoDaLiDa/Baltic-HLT 2025)

    Nielsen, D.S., Enevoldsen, K., Schneider-Kamp, P.: Encoder vs decoder: Comparative analysis of encoder and decoder language models on multilingual nlu tasks. In: Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies (NoDaLiDa/Baltic-HLT 2025). pp. 561–572 (2025)

    work page 2025

  35. [35]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Peng, X., Bai, Q., Xia, X., Huang, Z., Saenko, K., Wang, B.: Moment matching for multi-source domain adaptation. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1406–1415 (2019)

    work page 2019

  36. [36]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021)

    work page 2021

  37. [37]

    Acta Mathematica Academiae Scientiarum Hungarica 10(3), 441–451 (1959)

    Rényi, A.: On measures of dependence. Acta Mathematica Academiae Scientiarum Hungarica 10(3), 441–451 (1959)

    work page 1959

  38. [38]

    DINOv3

    Simoni, O., et al.: Dinov3: Self-supervised learning for vision at unprecedented scale. arXiv preprint arXiv:2508.10104 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Tan, Y ., Li, Y ., Huang, S.L.: Otce: A transferability metric for cross-domain cross-task rep- resentations. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 15779–15788 (2021)

    work page 2021

  40. [40]

    IEEE Transactions on Neural Networks and Learning Systems 36(2), 2423–2436 (2024)

    Tan, Y ., Zhang, E., Li, Y ., Huang, S.L., Zhang, X.P.: Transferability-guided cross-domain cross-task transfer learning. IEEE Transactions on Neural Networks and Learning Systems 36(2), 2423–2436 (2024)

    work page 2024

  41. [41]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Tran, A.T., Nguyen, C.V ., Hassner, T.: Transferability and hardness of supervised classification tasks. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 1395–1405 (2019)

    work page 2019

  42. [42]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Venkateswara, H., Eusebio, J., Chakraborty, S., Panchanathan, S.: Deep hashing network for unsupervised domain adaptation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 5018–5027 (2017)

    work page 2017

  43. [43]

    EmbeddingGemma: Powerful and Lightweight Text Representations

    Vera, H.S., Dua, S., Zhang, B., Salz, D., Mullins, R., Panyam, S.R., Smoot, S., Naim, I., Zou, J., Chen, F., et al.: Embeddinggemma: Powerful and lightweight text representations. arXiv preprint arXiv:2509.20354 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    arXiv preprint arXiv:2507.03175 (2025)

    Wang, H., Wang, J., Zhao, Z., Tan, Y ., Wu, Y ., Liu, H., Yang, J., Zhang, E., Chen, X., Rong, Z., et al.: Understanding knowledge transferability for transfer learning: A survey. arXiv preprint arXiv:2507.03175 (2025)

    work page arXiv 2025

  45. [45]

    In: International conference on machine learning

    You, K., Liu, Y ., Wang, J., Long, M.: Logme: Practical assessment of pre-trained models for transfer learning. In: International conference on machine learning. pp. 12133–12143. PMLR (2021)

    work page 2021

  46. [46]

    Advances in neural information processing systems32 (2019)

    Yurochkin, M., Claici, S., Chien, E., Mirzazadeh, F., Solomon, J.M.: Hierarchical optimal transport for document representation. Advances in neural information processing systems32 (2019)

    work page 2019

  47. [47]

    IEEE/CAA Journal of Automatica Sinica10(2), 305–329 (2022)

    Zhang, W., Deng, L., Zhang, L., Wu, D.: A survey on negative transfer. IEEE/CAA Journal of Automatica Sinica10(2), 305–329 (2022)

    work page 2022

  48. [48]

    Zhang, Y ., Li, M., Long, D., Zhang, X., Lin, H., Yang, B., Xie, P., Yang, A., Liu, D., Lin, J., et al.: Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176 (2025) 12 A RADAR robustness on ImageNet-C dataset In this section, we examine the robustness of RADAR using the ImageNet-C dataset, whic...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  49. [49]

    no replacement

    or extreme (Severity 5), but exhibits vulnerabilities to intermediate, unstructured perturbations (Severity 3). 13 Figure 3:Visualizing ImageNet-C synthetic corruptions.Examples of various visual corruptions across increasing severities (0, 1, 3, and 5). The profound structural degradation at Severity 5 fundamentally scrambles patch-level token distributi...

    work page

  50. [50]

    Specifically, we restrict the size of DomainNet and OfficeHome by adopting sampling protocols standard in the domain adaptation literature [47]

    to ensure tractable evaluation. Specifically, we restrict the size of DomainNet and OfficeHome by adopting sampling protocols standard in the domain adaptation literature [47]. Furthermore, we uniformly downsample the Amazon Reviews dataset from its original 256,000 samples per language to create a balanced, computationally efficient subset. Finally, to e...

    work page arXiv

  51. [51]

    Single-layer baselines are biased.Metrics that compare only final-layer feature distribu- tions (such as l2 centroid distance or s-OTDD) inherit a one-sided bias: they may report small divergence even when the underlying domains are meaningfully distinct at earlier layers

    work page

  52. [52]

    Multi-layer extraction is justified.Multi-layer extraction captures the full divergence profile across depth, recovering divergence signals that are monotonically suppressed as depth increases and would be invisible to any single-layer metric. The trajectory description leveraged by RADAR utilizes this insight by integrating geometric in- formation across...

    work page