Pith. sign in

REVIEW 4 major objections 5 minor 23 references

The paper claims that KL-anchored penalized likelihood estimation gives test-time transduction an adaptive shrinkage mechanism between zero-shot priors and empirical statistics, and that MOON's dynamic version makes it reliable under class

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

KL-anchored penalized likelihood with class- and instance-dependent shrinkage, implemented with von Mises-Fisher mixtures, improves CLIP test-time transduction under class imbalance.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection The PL view of KL-anchored transduction is a genuinely useful contribution, and the empirical gains are large; but the BSUM convergence guarantee rests on a false PSD claim and needs fixing. the 4 major comments →

arxiv 2607.15851 v1 pith:PJG46ND7 submitted 2026-07-17 cs.CV cs.LG

Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction

Jiazhen Huang , Zhiming Liu , Changhu Wang , Wei Ju , Ziyue Qiao , Xiao Luo This is my paper
classification cs.CV cs.LG
keywords test-time transductionvision-language modelsclass imbalancevon Mises-Fisher mixturepenalized likelihood estimationKL divergence anchoradaptive shrinkageCLIP zero-shot
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper's thesis is that test-time transduction for vision-language models breaks down under realistic class imbalance because it amounts to unregularized maximum-likelihood estimation of class statistics, and that the fix is a statistical anchor. It proves that if the anchor is a KL divergence toward zero-shot class priors, the optimal parameter update is a convex combination: each class's estimate moves from its prior anchor toward the empirical estimate with weight n_k/(n_k+alpha), so abundant classes adapt fully while absent classes stay pinned to the anchor. On this basis it builds MOON, a mixture of von Mises-Fisher distributions on the unit hypersphere, and makes the shrinkage strength dynamic: uncertain instances are downweighted by entropy and classes are downweighted according to zero-shot confidence. The claimed payoff is that MOON improves zero-shot CLIP by 13.2% on ImageNet across ten imbalanced scenarios, beats the strongest baseline by 8.8%, and costs a few milliseconds per batch, all without training or task-specific hyperparameter tuning. If correct, the paper turns transduction from a brittle operation into a reliable, plug-in post-processing step for vision-language models.

Core claim

At the center is the identity from Sec. 3: for a KL-anchored penalized likelihood model with exponential-family class conditionals, the optimal class parameter satisfies a convex-combination update in the mean-parameter space, with weight beta_k = n_k/(n_k+alpha) on the empirical estimate and the complementary weight on the zero-shot anchor. This says every class update is a convex combination of the empirical class statistic and the prior anchor, with the balance set automatically by how much evidence the class has. MOON instantiates this general result using von Mises-Fisher distributions because normalized VLM embeddings live on the unit sphere, and replaces the scalar anchor weight alpha

What carries the argument

KL-anchored penalized likelihood estimation (PLE): the paper's unifying lens, whose core mechanism is the adaptive-shrinkage identity. For any exponential-family mixture, the KL anchor term turns the parameter-update step into a strictly convex problem whose unique minimizer is a convex combination of prior anchor and empirical statistic, weighted by beta_k = n_k/(n_k+alpha). MOON's concrete machinery is a mixture of von Mises-Fisher (vMF) distributions on the unit hypersphere, the natural spherical analogue of the Gaussian, with closed-form updates for mean direction and concentration; combined with dynamic instance weights gamma_i = 1 - H(z_i)/logK and class weights alpha_k = 1/lambda_k, t

Load-bearing premise

The monotone-convergence proof requires the affinity matrix W to be positive semidefinite (no negative quadratic directions), yet the implemented m-nearest-neighbor sparsified Gram matrix need not be; if W is indefinite, the first-order Taylor upper bound and the convergence guarantee in Eq. (21) do not follow.

What would settle it

Compute the smallest eigenvalue of the implemented m-NN similarity matrix on a real ImageNet batch; if it is negative, the concavity/upper-bound assumption fails. Independently, record the full objective L(z, theta) across iterations of MOON: any increase outside numerical noise directly refutes the monotone-convergence claim in Eq. (21).

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Existing unanchored transductive methods should be expected to collapse under imbalance: without the anchor, beta_k is identically 1, so their parameter updates are pure MLE and overfit majority classes; the paper gives a unified explanation for that collapse.
  • Rare and absent classes are protected by construction: with zero soft count the update stays exactly at the zero-shot anchor, and with small n_k the deviation is linearly shrunk, so transduction no longer needs to know which classes are effective.
  • The anchoring mechanism is family-agnostic: because the convex-combination identity holds for any regular exponential family, the principle transfers to whichever distribution best matches the representation space, not only vMF.
  • MOON's claimed computational profile, with no gradients, no covariance matrices, and a single-pass assignment update, makes it usable as a routine post-processing step even for batch sizes up to tens of thousands.
  • In a sample-wise mode with a memory bank, the same anchored update yields immediate predictions for streaming data, extending transduction to online settings without rerunning the batch algorithm.

Where Pith is reading between the lines

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

  • The convergence proof in App. A and G.1 leans on the affinity matrix W being positive semidefinite. Since the implemented m-nearest-neighbor thresholded Gram matrix need not be PSD, a skeptical reader should check whether the objective is actually monotone on real batches; the empirical gains can survive even if this guarantee does not.
  • The same PLE-shrinkage recipe could be ported to other soft-clustering or output-calibration stages wherever a trustworthy prior anchor exists; the paper demonstrates it only for VLM image classification.
  • The dynamic weights are heuristic in form: entropy-based instance weights and inverse-confidence class weights. A testable refinement is to replace them with calibrated predictive uncertainty, which could improve the mild-imbalance regime where the paper reports a slight drop.
  • Because the vMF assumption is isotropic, MOON may miss anisotropic feature geometry; fitting an elliptical spherical distribution such as Fisher-Bingham is a natural follow-up that could recover the gap on datasets where Gaussian variants are stronger.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MOON, a training-free transductive method for vision-language models under realistic class imbalance. It casts test-time transduction as penalized likelihood estimation (PLE) with a KL-divergence anchor, and proves (Theorem D.2) that for regular exponential families the anchored M-step update is a convex combination of the empirical estimate and the prior anchor in mean-parameter space. Building on this, MOON models normalized features with a mixture of von Mises-Fisher distributions and introduces instance-level and class-level shrinkage weights derived from zero-shot prediction entropy and confidence. The method is evaluated on 11 datasets under batch and online adaptation settings, reporting large gains over zero-shot CLIP and existing transductive/TTA baselines, with very low runtime. The paper also includes a sample-wise extension, ablation studies, and experiments across multiple backbones and VLMs.

Significance. The paper makes a genuinely useful conceptual contribution: framing KL-anchored transduction as PLE and showing that the anchor induces adaptive shrinkage in the exponential family is clean and, to my knowledge, not stated in this generality before. Theorem D.2 is a solid, parameter-free theoretical result that does not depend on any label or target accuracy, which is a real strength. The empirical study is extensive (11 datasets, multiple backbones, multiple realistic imbalance scenarios, ablations) and the reported gains are large; the code is promised. The efficiency numbers are also attractive. However, the optimization-theoretic guarantees presented as part of the contribution are not supported as written, and one of the key equivalence claims used in the algorithm relies on an unvalidated approximation. These issues do not necessarily invalidate the empirical findings, but they need to be fixed or explicitly qualified before the theoretical framing can be accepted.

major comments (4)
  1. [App. G.1, Eqs. (71)–(75); App. A, Eq. (21)] The claim that the symmetrized matrix (GW+WG)/2 is PSD whenever W is PSD and γ_i ≥ 0 is false. Counterexample: N=2, K=1, W=[[1,1],[1,1]] (PSD), G=diag(1,2). Then (GW+WG)/2 = [[1,1.5],[1.5,2]], whose determinant is −0.25, so it is indefinite. Consequently, the quadratic form −Σ γ_i ω_ij z_iᵀ z_j is not guaranteed to be concave, and the first-order Taylor surrogate in Eqs. (73)–(75) is not guaranteed to majorize the objective. The monotone convergence guarantee L(z^(t+1),Θ^(t+1)) ≤ L(z^(t),Θ^(t)) in Eq. (21) is therefore not proven. This is a load-bearing theoretical claim; the paper either needs a valid proof under additional conditions (e.g., constant γ, or a different surrogate) or a clear statement that convergence is only empirical.
  2. [App. G.1, page 21; Alg. 1, line 3] The text states that since ω_ij = f_iᵀ f_j ≥ 0, the affinity matrix W is PSD. Nonnegativity of entries does not imply PSD. Moreover, the implemented matrix is the m-NN thresholded graph of Alg. 1 line 3, which is not the full Gram matrix and need not be PSD even if the full Gram is. The convergence proof should address the actual matrix used in the algorithm, or the algorithm should be changed to use a provably PSD affinity (e.g., a symmetrized normalized graph Laplacian).
  3. [Alg. 1, steps 6–14; App. A, convergence guarantee] The BSUM convergence guarantee is stated for a fixed objective. In Alg. 1, the instance-level weights γ_i are updated at every iteration (step 9) based on the current assignments z_i. Since γ_i appears as a coefficient in the objective (Eq. (8)), the objective itself changes from iteration to iteration. Thus, even if the concavity issue were resolved, the standard BSUM monotone-decrease argument would not apply to the actual algorithm. The paper should either treat γ as fixed hyperparameters in the theoretical analysis, or analyze the dynamic-weight case explicitly.
  4. [Eq. (14) and App. H] The equivalence between Eq. (13) and Eq. (14) is proved under the approximation Ad(κ'_k) ≈ 1. This approximation is not validated for the feature distributions encountered in the experiments, and it changes the effective anchor strength: without the approximation, the anchor term is α_k Ad(κ'_k), not α_k. Since Eq. (14) is the form actually implemented in Alg. 1 (steps 10–12), the paper should either justify the approximation empirically (e.g., report the range of Ad(κ'_k) across datasets) or use the exact form with α_k Ad(κ'_k).
minor comments (5)
  1. [App. A, after Eq. (21)] Typo: 'Since La is lower-bounded' should read 'Since L is lower-bounded'.
  2. [Eq. (11)] The notation '⊙' between the average and the maximum is nonstandard; the geometric mean is usually written as (avg·max)^{1/2}.
  3. [Tab. 3] The runtime for MOON is identical (0.03 s) for batch sizes 128 and 1,000; this seems suspicious and should be clarified (e.g., measurement precision or GPU overhead).
  4. [Sec. 5.3, Tab. 7] The 'linear representation hypothesis' is invoked to explain why vMF underperforms GMM in StatA; the explanation is plausible but not tested. Consider adding a simple diagnostic (e.g., covariance eigenvalue spread) to support it.
  5. [App. I.4, Tab. 19] The full-dataset (all classes) result shows MOON trailing StatA by ~2%. The text acknowledges this but frames it as a 'competitive' trade-off. Given the paper's emphasis on robustness, this limitation should appear earlier, in the main text, not only in an appendix.

Circularity Check

0 steps flagged

No significant circularity: the KL-anchored shrinkage theorem is derived from the PLE objective, and dynamic weights are computed from zero-shot predictions, not fitted to evaluation accuracy.

full rationale

The central derivation (Thm. D.2 / Eq. 5) is a self-contained algebraic consequence of the KL-anchored M-step for exponential families: differentiating the objective in Eq. (27) gives ∇A(η*) = (S_k + αμ'_k)/(n_k + α), which is exactly the stated convex-combination shrinkage. The vMF updates in Eqs. (13)-(14) are derived from the same objective in App. G.2 and H, not from target accuracies. Dynamic instance-level γ_i and class-level α_k are computed from zero-shot predictions and current assignments (Eqs. 9-11), not fitted to benchmark labels or accuracies. The only notable weakness is that the BSUM convergence proof in App. A / G.1 asserts W is PSD from nonnegative entries and that per-sample γ weighting preserves concavity; both claims are questionable, and the counterexample shows the monotone-decrease guarantee may fail. However, an invalid convergence proof is a correctness risk, not circularity: the method's predictions are still produced by the stated equations and evaluated against external data. No self-citation chain or definitional equivalence is load-bearing, so the derivation is not circular.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central derivation is self-contained within exponential-family statistics, but the implemented algorithm relies on several unverified premises: the PSD property of the sparsified affinity graph, isotropic vMF geometry, zero-shot confidence as a presence proxy, and the Ad≈1 approximation. The hand-selected functional forms for γ and α are additional design degrees of freedom chosen via ablations on the same evaluation benchmark.

free parameters (5)
  • Instance-level reliability mapping γ_i = 1 − H(z_i)/log K
    Hand-chosen entropy-to-weight mapping (Eq. 9); no per-task tuning, but the functional form is selected by the authors without a formal derivation.
  • Class-level confidence mapping α_k = 1/λ_k with λ_k = sqrt(avg·max)
    Ad hoc inverse-confidence form plus geometric-mean aggregation (Eqs. 10–11); justified by ablations on the same benchmark (Tabs. 5–6) rather than by theory.
  • Number of neighbors m = 3
    Fixed hyperparameter for affinity graph construction; sensitivity shown in Fig. 4, default set by the authors.
  • Number of iterations T = 10
    Fixed iteration count; Fig. 3/4 show saturation near T≈3, but the paper uses T=10 for all experiments.
  • Hard-assignment discretization of β_k = argmax
    Shrinkage strength uses hard label counts (Alg. 1 line 10, Eq. 19); chosen for robustness based on Tab. 21 rather than derived.
axioms (6)
  • ad hoc to paper The m-NN sparsified affinity matrix W with nonnegative entries ωij = f_i^T f_j is positive semidefinite
    Invoked for concavity of the Laplacian term and the BSUM convergence guarantee (App. A, App. G.1). Nonnegativity of entries does not imply PSD for a thresholded Gram matrix; this premise is unsupported for the actual graph built in Alg. 1 line 3.
  • ad hoc to paper Class-conditional image features are isotropically distributed on the unit hypersphere and follow vMF mixtures
    This is the core distributional model (Sec. 4). The paper acknowledges in Limitation K and Tab. 7 that VLM features are anisotropic, so vMF cannot capture the full geometry.
  • domain assumption Visual and text features are ℓ2-normalized so all operations live on S^{d−1}
    Used throughout Eq. (1), Eq. (6), and App. E/F; standard for CLIP-like embeddings and stated in the problem definition.
  • domain assumption Zero-shot text prototypes are valid prior anchors for every class
    Anchors are initialized as µ'_k = t_k (Eq. 7). If text-image alignment is poor for a downstream task, the anchors mislead rather than stabilize.
  • ad hoc to paper Zero-shot prediction confidence is a valid proxy for whether a class is present in the current batch
    The class-level dynamic shrinkage uses λ_k from average/max confidence (Eq. 11) to suppress outlier classes. The paper's own results on the uniform/full-class setting (Tab. 19) and Low online scenario show this bias can hurt.
  • ad hoc to paper Ad(κ'_k) ≈ 1 in the implemented shrinkage update
    Used to equate the exact update in Eq. (13) with the simplified adaptive-shrinkage form in Eq. (14), and in Alg. 1 step 12. Described as 'mild' in App. H, but no empirical validation is provided.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction." pith.science (2026). https://pith.science/paper/PJG46ND7

@misc{pith2026260715851,
  author       = {Pith},
  title        = {Pith review of: Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJG46ND7}},
  note         = {Machine review of arXiv:2607.15851}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

A range of methods aim to enhance the performance of vision-language models (VLMs) at test time. Among them, transduction has emerged as a promising paradigm due to its strong compatibility and efficiency. However, realistic evaluations often involve highly imbalanced class distributions, which cause performance degradation or even collapse. In this work, we systematically revisit transduction from the perspective of penalized likelihood estimation (PLE), showing that PLE with a KL-divergence anchor term naturally yields an adaptive shrinkage behavior between prior anchors and empirical estimates. From this viewpoint, the brittleness of transductive methods can be attributed to the absence of anchoring mechanism and static modeling of the shrinkage strength. Therefore, we propose Mixture of Von Mises-Fisher Models with Dynamic Shrinkage (MOON). MOON is built upon a mixture of von Mises-Fisher distributions to model feature representations on the unit hypersphere. To handle imbalance, MOON dynamically adjusts the shrinkage strength using zero-shot priors at both instance and class levels. Thus, it suppresses unreliable assignments and prevents harmful updates from outlier classes, thereby mitigating negative transfer. MOON is model-agnostic, training-free, and requires no task-specific hyperparameter tuning. Extensive experiments further validate the advantage of MOON in both performance and efficiency. Our code is available at https://github.com/walawalagoose/MOON

Figures

Figures reproduced from arXiv: 2607.15851 by Changhu Wang, Jiazhen Huang, Wei Ju, Xiao Luo, Zhiming Liu, Ziyue Qiao.

Figure 1
Figure 1. Figure 1: Performance comparison on two realistic settings. Existing transductive or online TTA methods suffer from performance degradation or even collapse, while our proposed MOON consistently enhances VLM prediction and outperforms state-of-the-art baselines. methods suffer from performance degradation or even col￾lapse. Statistically, the failures of transductive methods may arise as they implicitly treat predic… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Controlling shrinkage strength with anchor weight α of state-of-the-art method StatA. Such static modeling is suboptimal in accuracy and robustness. (b) Accuracy-Runtime Tradeoff. MOON enables effective and efficient adaptation. using a mixture of von Mises-Fisher (vMF) distributions on the unit hypersphere. To robustly handle class imbalance, MOON dynamically adjusts the shrinkage strength using zero￾… view at source ↗
Figure 3
Figure 3. Figure 3: Convergence analysis on ImageNet and DTD. We demonstrate performance curves over iterations for each method. 5. Experiments We evaluate our method in several scenarios under two real￾istic settings, as defined in Sec. 3. We report the Top-1 ac￾curacy across 11 public fine-grained classification datasets, and adopt CLIP ViT-B/16 as our default VLM backbone. Please see App. C for details on datasets, baselin… view at source ↗
Figure 4
Figure 4. Figure 4: Hyperparameter sensitivity analysis. Results are reported on batch adaptation, Medium scenario, batch size of 1,000. Bs=64 Bs=1000 [PITH_FULL_IMAGE:figures/full_fig_p033_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Detailed ablation study on components, over various batch sizes and scenarios. Each reported performance is averaged over all datasets and runs. J. Additional Analyses Hyperparameter sensitivity. We analyze the sensitivity of MOON to the existing hyperparameters, including iteration number T and the number of neighbors in Laplacian term m in [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗

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

23 extracted references · 11 linked inside Pith

  1. [1]

    The BSUM-style iterative optimization of MOON can be conceptualized as a generalized Expectation-Maximization (EM) algorithm: fixing parameters (µk, κk) to update assignmentszcorresponds to the E-step, while fixingzand updating(µ k, κk)corresponds to the M-step. Algorithm 1Overall procedure ofMOON Require: Visual feature embeddings {fi}N i=1,f i ∈ Sd−1, t...

    2025

  2. [4]

    A., Bohn ´e, J., Milgram, J., Gentric, S., and Chen, L

    Hasnat, M. A., Bohn ´e, J., Milgram, J., Gentric, S., and Chen, L. von mises-fisher mixture model-based deep learning: Application to face verification.arXiv preprint arXiv:1706.04264,

    Pith/arXiv arXiv

  3. [8]

    All the configuration is kept the same as those set in the original paper on ImageNet

    • MTA (CVPR’24): proposes a training-free strategy that leverages MeanShift on augmented views to identify dis- tribution modes, jointly optimizing a learnable inlierness score to robustly aggregate visual information. All the configuration is kept the same as those set in the original paper on ImageNet. C.3. Prompts Following (Zhang et al., 2022), we ado...

    2022

  4. [10]

    R., and Shah, M

    Soomro, K., Zamir, A. R., and Shah, M. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402,

    Pith/arXiv arXiv

  5. [12]

    Tent: Fully test-time adaptation by entropy minimiza- tion.arXiv preprint arXiv:2006.10726,

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., and Darrell, T. Tent: Fully test-time adaptation by entropy minimiza- tion.arXiv preprint arXiv:2006.10726,

    Pith/arXiv arXiv 2006

  6. [13]

    A hard-to-beat baseline for training-free clip-based adaptation.arXiv preprint arXiv:2402.04087,

    Wang, Z., Liang, J., Sheng, L., He, R., Wang, Z., and Tan, T. A hard-to-beat baseline for training-free clip-based adaptation.arXiv preprint arXiv:2402.04087,

    Pith/arXiv arXiv

  7. [14]

    Boostadapter: Improving test-time adaptation via regional bootstrapping.arXiv preprint arXiv:2410.15430, 2024a

    Zhang, T., Wang, J., Guo, H., Dai, T., Chen, B., and Xia, S.-T. Boostadapter: Improving test-time adaptation via regional bootstrapping.arXiv preprint arXiv:2410.15430, 2024a. Zhang, Y ., Zhu, W., Tang, H., Ma, Z., Zhou, K., and Zhang, L. Dual memory networks: A versatile adaptation ap- proach for vision-language models. InProceedings of the IEEE/CVF conf...

    Pith/arXiv arXiv

  8. [15]

    These experiments cover the full range of realistic batch and online adaptation settings. MOON demonstrates remarkable universality, achieving the highest accuracy in nearly all evaluated scenarios (winning in 33 out of 36 cases) with only negligible margins in the few exceptions. This consistent superiority across diverse architectures and model capaciti...

    2023

  9. [17]

    Experimental Details C.1

    C. Experimental Details C.1. Datasets We evaluate our proposed MOON and other baselines on 11 widely-used public datasets for fine-grained visual classification. These datasets cover a diverse range of domains, including generic objects, scenes, textures, satellite imagery, and specific fine-grained categories. Specifically, the benchmark includes: ImageN...

    2009

  10. [18]

    C.2. Baselines We compare ourMOON against a comprehensive set of baselines, which are categorized into: (1)Transductive methods: EM- Dirichlet (Dirichlet) (Martin et al., 2024), ZLaP (Kalantidis et al., 2024), GDA-CLIP (Wang et al., 2024), TransCLIP (Zanella et al., 2024), ADAPT (Zhang et al., 2025), and StatA (Zanella et al.,

    2024

  11. [19]

    We also incorporate another TTA method MTA (Zanella & Ben Ayed,

    6; (2)Online TTA methods: TENT (Wang et al., 2020), TDA (Karmanov et al., 2024), DMN (Zhang et al., 2024b), and OGA (Fuchs et al., 2025). We also incorporate another TTA method MTA (Zanella & Ben Ayed,

    2020

  12. [21]

    a photo of a [ ]

    Table 10.Dataset information and prompt templates. Name Other name #K#NDescription Prompt template SUN397 SUN397 397 19,850 Scenes classification"a photo of a [ ]." Aircraft FGVCAircraft 100 3,333 Aircraft classification"a photo of a [ ], a type of aircraft." EuroSAT EuroSAT 10 8,100 Satellite images classification"a centered satellite photo of [ ]." Stan...

    2025

  13. [22]

    Hence, ∥fi −µ ′ k∥2 2 =∥f i∥2 2 +∥µ ′ k∥2 2 −2(f ⊤ i µ′ k) = 2 1−cosθ i ,(65) wherecosθ i ≜f ⊤ i µ′ k. Taking the weighted average overifor classkgives MSEk = 2 1−E k[f ⊤µ′ k] ,E k[f ⊤µ′ k]≜ P i zi,k f ⊤ i µ′ kP i zi,k .(66) Therefore, Ek[f ⊤µ′ k] = 1− MSEk 2 .(67) For a vMF distribution onS d−1 with densityp(f) =C d(κ) exp(κµ⊤f), it is well-known that9 E...

    2024

  14. [2012]

    Eva- clip: Improved training techniques for clip at scale.arXiv preprint arXiv:2303.15389,

    Sun, Q., Fang, Y ., Wu, L., Wang, X., and Cao, Y . Eva- clip: Improved training techniques for clip at scale.arXiv preprint arXiv:2303.15389,

    Pith/arXiv arXiv

  15. [2013]

    Correct- ing visual blur induced by attention distraction to reduce hallucinations: Algorithm and theory

    Li, Q., Liu, Z., Luo, W., Luo, T., and Hou, C. Correct- ing visual blur induced by attention distraction to reduce hallucinations: Algorithm and theory. InForty-third In- ternational Conference on Machine Learning, 2026a. Li, Q., Liu, Z., Xu, T., Luo, T., and Hou, C. Adap- tive disentangled representation learning for incomplete multi-view multi-label cla...

    arXiv

  16. [2014]

    Dino as a von mises-fisher mixture model.arXiv preprint arXiv:2405.10939,

    Govindarajan, H., Sid´en, P., Roll, J., and Lindsten, F. Dino as a von mises-fisher mixture model.arXiv preprint arXiv:2405.10939,

    Pith/arXiv arXiv

  17. [2019]

    What drives test-time adaptation for clip? a controlled em- pirical study from an update perspective.arXiv preprint arXiv:2606.14299,

    Huang, J., Chen, X., Liu, Z., Sun, Y ., Jiang, J., and Wang, Z. What drives test-time adaptation for clip? a controlled em- pirical study from an update perspective.arXiv preprint arXiv:2606.14299,

    arXiv

  18. [2020]

    Do all individual layers help? an empirical study of task-interfering layers in vision-language models

    Liu, Z., Wei, Y ., Feng, L., Su, X., Xia, X., Guan, W., Xie, Z., and Yang, S. Do all individual layers help? an empirical study of task-interfering layers in vision-language models. arXiv preprint arXiv:2602.01167,

    Pith/arXiv arXiv

  19. [2021]

    Learning generative visual models from few training examples: An incremen- tal bayesian approach tested on 101 object categories

    Fei-Fei, L., Fergus, R., and Perona, P. Learning generative visual models from few training examples: An incremen- tal bayesian approach tested on 101 object categories. In 2004 conference on computer vision and pattern recogni- tion workshop, pp. 178–178. IEEE,

    2004

  20. [2022]

    Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151,

    Maji, S., Rahtu, E., Kannala, J., Blaschko, M., and Vedaldi, A. Fine-grained visual classification of aircraft.arXiv preprint arXiv:1306.5151,

    Pith/arXiv arXiv

  21. [2023]

    J., and Veitch, V

    11 Von Mises-Fisher Mixture Model with Dynamic Shrinkage for Realistic Test-Time Transduction Park, K., Choe, Y . J., and Veitch, V . The linear represen- tation hypothesis and the geometry of large language models.arXiv preprint arXiv:2311.03658,

    Pith/arXiv arXiv

  22. [2024]

    Z., and Zhang, C

    Han, Z., Yang, J., Wang, G., Li, J., Xu, Q., Shou, M. Z., and Zhang, C. Dota: Distributional test-time adaptation of vision-language models.arXiv preprint arXiv:2409.19375,

    arXiv

  23. [2025]

    On pitfalls of test-time adaptation.arXiv preprint arXiv:2306.03536,

    Zhao, H., Liu, Y ., Alahi, A., and Lin, T. On pitfalls of test-time adaptation.arXiv preprint arXiv:2306.03536,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.