Pith. sign in

REVIEW 3 major objections 4 minor 60 references

LFM shows that foundation models can drive source-free universal domain adaptation if the label shift type is first inferred from a single variance statistic and CLIP pseudo-labels are refined by a source-initialized target model.

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 ↗

A pipeline using a vision-language model, LLM-generated unknown-class text labels, a shift-type detector, and consensus pseudo-label refinement adapts a pre-trained model to unlabeled target data across partial, open, and open-partial label shifts.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection LFM is a solid, reproducible SF-UniDA pipeline whose main generality claim rests on benchmark-calibrated hyperparameters; worth a serious referee, but the LSTD gate needs out-of-benchmark validation. the 3 major comments →

arxiv 2607.17653 v1 pith:OHSIFD4N submitted 2026-07-20 cs.CV cs.LGcs.MM

LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation

Jing Li , Pan Liu , Meng Zhao , Wanli Xue , Yanhong Yang , Xu Cheng , Fan Shi , Jianhua Zhang
show 2 more authors
Qinghua Hu Shengyong Chen
This is my paper
classification cs.CV cs.LGcs.MM
keywords source-free universal domain adaptationlabel shift type detectionfoundation modelsCLIP zero-shotLLM-generated unknown labelsGaussian mixture modelpseudo-label refinementcoefficient of variation
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

This paper tries to establish that source-free universal domain adaptation (SF-UniDA) — adapting a pretrained source model to an unlabeled target domain whose label set may be smaller, larger, or partially overlapping, with no access to source data — can be solved with a single pipeline that replaces manual thresholding and clustering with foundation-model priors. The proposed LFM framework uses a large language model to invent names for plausible unknown classes, a vision-language model (CLIP) to score each target image against known and invented labels, and the coefficient of variation of a per-sample score to decide whether unknown classes are present at all. If they are, a binary Gaussian mixture model separates them; then a consensus step lets the source-initialized target model refine CLIP's pseudo-labels. On Office-31, Office-Home, VisDA, and DomainNet, under all three label-shift types, LFM reports top or competitive H-scores/accuracy while deploying only a ResNet-based target model at inference, suggesting foundation models can serve as semantic priors rather than inference backbones.

Core claim

The paper claims that the three label-shift regimes of SF-UniDA — partial, open, and open-partial — can be unified under one adaptation mechanism, provided the pipeline first correctly infers which regime it is in. It does this by computing a per-sample score that contrasts the target image's top similarity to known-class text labels against its top similarity to LLM-envisioned unknown-class labels; the coefficient of variation of those scores across the target set is high when target-private classes exist (OSDA/OPDA) and low when they do not (PDA), with a measured gap of 0.7 vs. 1.1 on the four benchmarks. A binary GMM then separates known from unknown samples using a separate un-scaled met

What carries the argument

The central object is the per-sample score function and its coefficient of variation (cv): the score contrasts a sample's best CLIP similarity to known-class labels with its best similarity to LLM-generated unknown-class labels, and cv over the target set is the label shift type detector (threshold δ=1.0). The GMM fitted to a second metric, rate, separates unknown samples; and the PLR (pseudo-label refinement) module — top-3 CLIP candidates, target-model neighborhood soft-voting, and a confidence-and-consensus rule — fuses the two knowledge sources. The LLM's role is purely offline: it generates candidate text labels for unknown classes conditioned on the known set, and CLIP is frozen, so al

Load-bearing premise

The pipeline's first step assumes that the coefficient of variation of the sample score cleanly separates PDA (≤0.7) from OSDA/OPDA (≥1.1) on any new target domain, so a fixed threshold δ=1.0 always picks the right shift type.

What would settle it

An experiment that would settle the claim: take a target domain constructed to be pure PDA but with highly diverse known classes (e.g., a wide set of visual categories) and compute cv; if cv exceeds 1.0, the LSTD would wrongly trigger unknown separation and the accuracy would collapse versus a forced-PDA variant. Conversely, an OSDA target whose unknown classes are visually very close to known classes might produce cv below 1.0, causing unknown samples to be treated as known. Measuring cv on a held-out benchmark outside the four used would test the generality of the 0.7–1.1 gap.

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

If this is right

  • If the cv separation is reliable, practitioners can automatically choose the correct adaptation strategy for an unlabeled target domain without source data or manual shift-type labeling.
  • Pseudo-labeling becomes one to two orders of magnitude cheaper than clustering-based methods (e.g., GLC), since it avoids iterative clustering over high-dimensional features.
  • Because foundation models are used only during training, deployment remains a lightweight ResNet model, preserving architectural fairness with existing domain adaptation methods.
  • The framework's success on all three shift types suggests that a single universal algorithm can replace shift-specific methods, and the threshold δ=1.0 generalizes across benchmarks.
  • LLM-envisioned labels need not match ground-truth unknown classes exactly; visual-similarity-based invention suffices to improve unknown detection, as shown in Table XII.

Where Pith is reading between the lines

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

  • The cv-based shift detector could transfer to other modalities (e.g., text or audio) where a pretrained embedding model can score samples against known and invented class descriptors, if the same variance gap appears.
  • A possible failure mode not explored: if the LLM invents labels that are semantically close to known classes, the score contrast weakens and the cv gap may close; testing with diverse LLMs and label sets would map the reliability boundary.
  • The consensus pseudo-label rule is a testable recipe for any two classifiers (not just CLIP vs. source model); one could verify whether 'confidence + top-k overlap' generalizes to other model pairs.
  • Since β (the unknown-penalty weight) is benchmark-specific, a promising extension is a data-driven β estimator that removes the last manual hyperparameter.
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

3 major / 4 minor

Summary. The paper proposes LFM, a source-free universal domain adaptation framework. LFM first uses an LLM to generate textual labels for potential target-private classes, then uses a frozen CLIP model to compute similarities between target images and text labels for both source-known and LLM-generated unknown classes. A sample-level score is defined in Eq. (3), and the coefficient of variation of this score (Eq. (4)) is used in the LSTD module to decide whether the label shift is PDA or OSDA/OPDA. If unknown classes are detected, a binary GMM separates known and unknown samples using the alternative metric rate_i in Eq. (5). PLR then refines CLIP pseudo-labels by a confidence-and-consensus strategy with predictions from a source-initialized target model, Eq. (11). The target model is trained with cross-entropy on known samples and entropy maximization on unknowns. Experiments cover OSDA, PDA, OPDA, and CLDA on Office-31, Office-Home, VisDA, and DomainNet, with the final inference model being a ResNet backbone.

Significance. If the generality concerns can be addressed, LFM is a genuinely useful contribution to SF-UniDA. Its strengths are explicit: it sidesteps costly clustering-based pseudo-labeling, preserves a ResNet-based target model at inference, reports experiments across all three label-shift settings plus closed-set DA, includes an efficiency comparison, and releases code. The LLM-generated unknown-class labels and the consensus refinement are interesting and reasonably well motivated. However, the paper's central universality claim rests on the LSTD gate, and that gate is calibrated to the four benchmarks used for evaluation. The mismatch between the method's assumptions and the actual class splits, together with the benchmark-specific hyperparameters, currently limits the claim to the evaluated benchmark suite rather than to universal SF-UniDA.

major comments (3)
  1. [Section III-D, Section IV-B, Section IV-E (Eqs. (3)-(4), Fig. 5)] The LSTD threshold delta=1.0 and the score parameter beta are both calibrated on the four evaluation benchmarks. Eq. (3) defines score_i using beta, and Eq. (4) computes cv from that score; Section IV-B sets beta=0.25 for Office-31/DomainNet and beta=0.75 for Office-Home/VisDA. Section IV-E justifies delta=1.0 solely from the observed gap in Fig. 5 on those same benchmarks (PDA max 0.7, OSDA/OPDA min 1.1). Because LSTD decides whether the GMM and L_unk are used at all, a wrong shift classification applies the wrong adaptation strategy to every target sample. The paper provides no principled way to set beta or delta on a new target domain without labeled target data, and Table VIII only checks sensitivity on Office-Home. This is the main load-bearing generality claim; it needs either an unsupervised selection rule or validation on held-out benchmarks/splits.
  2. [Section III-B, Section IV-B, Table I, Fig. 8(c)] The number of generated unknown labels is C_unk = alpha * C_s with alpha=2 fixed in Section IV-B. This does not match the actual class splits in Table I: Office-Home OPDA has 10 common and 50 target-private classes, VisDA OPDA has 6 and 6, DomainNet OPDA has 150 and 145. The sensitivity analysis in Fig. 8(c) is claimed to show that alpha=2 is optimal, but this is not consistent with the benchmark ratios (5, 1, ~1). Moreover, Table XI lists 15 known and 30 unknown labels for Office-Home OPDA, whereas Table I lists 10 common and 50 private classes. This inconsistency matters because the score in Eq. (3) and the GMM separation in Section III-E operate on similarities to this fixed label set; if the generated unknown labels do not cover the actual unknown classes, the unknown-detection rationale changes. The authors should clarify how alpha is meant to relate to the true unknown-class count
  3. [Section IV-C, Tables II-VI] The evaluation protocol states that all compared methods must use the same ResNet-based backbone as the target model and that ViT-based extractors are excluded. However, DIFO-C-B32 is included in all main tables; by its name and the original method it uses a CLIP ViT-B/32 backbone as the feature extractor. If a ResNet-based variant of DIFO is being evaluated, that must be stated explicitly; if not, the comparison violates the paper's own fairness criterion and the claimed superiority over FM-based methods is not architecture-fair. The same concern applies to the CLIP Zero-Shot and LLM+CLIP baselines, which are not trained target models but are compared on the same inference protocol.
minor comments (4)
  1. [Table VII] The table caption does not state which metric is reported for each shift; PDA uses accuracy while OSDA/OPDA use H-score, and this should be indicated in the caption or table header.
  2. [Section IV-E, Fig. 5] The text says each bar represents a target domain, with digits showing cv values, but the figure itself is not reproduced in the manuscript text; please ensure the values discussed (0.7 and 1.1) are clearly visible in the final figure.
  3. [Eq. (11)] The notation in Eq. (11) uses a tilde superscript for the model pseudo-label, but the text uses a hat for the final refined label; the hat notation should be introduced before the equation to avoid confusion.
  4. [Section IV-F] The t-SNE and Grad-CAM visualizations are described briefly; adding quantitative metrics, such as A-distance or class-wise accuracy, would strengthen the claim that the learned features are more discriminative.

Circularity Check

1 steps flagged

LSTD gate is calibrated to the four evaluation benchmarks, making the shift-type decision fitted rather than derived; the rest of the adaptation pipeline is empirical and not circular.

specific steps
  1. fitted input called prediction [Section III-D (LSTD, Eq. (4)) with threshold set in Section IV-E (Fig. 5) and beta tuned in Section IV-B]
    "when the label shift corresponds to the PDA setting, the coefficient of variation (cv) across different target domains on the four benchmarks remains relatively low, with a maximum value of 0.7. In contrast, under the OPDA and OSDA settings which involve unknown categories, cv is significantly higher, with a minimum of 1.1. Based on these observations, we set δ=1.0 as a threshold for cv to distinguish between the presence or absence of unknown classes in the target domain. ... any choice of δ within the interval [0.7,1.1] leads to identical label-shift type determination across all benchmarks"

    LSTD is presented as inferring the shift type of the unlabeled target domain, and this decision gates whether GMM unknown separation and L_unk are used at all. But δ=1.0 is not derived; it is selected after inspecting the cv values on exactly the four benchmarks whose shift types are known from Table I. The paper states any δ in [0.7,1.1] yields identical determinations, i.e., the threshold is placed inside the observed gap, so on those benchmarks the OSDA/OPDA-vs-PDA decision is forced by the calibration data rather than independently predicted. The cv values are computed from score_i, whose β is separately tuned per benchmark (0.25/0.75, Sec. IV-B), so the 'single threshold' separation is not parameter-free. This makes the LSTD stage partially circular; the final H-scores still depend on

full rationale

Most of the pipeline is empirical and externally evaluated: Eq. (1)-(3) define CLIP-similarity features, Eq. (5) defines the GMM metric, and the reported H-scores/accuracies are measured on standard benchmarks rather than derived from the method's own equations. I found no self-definitional equality, no load-bearing self-citation, and no imported uniqueness theorem; the overlapping-author references ([2], [27], [30], [53]) appear only in related work and are not used to justify the framework. The one partial circularity is the LSTD gate: δ=1.0 is set to lie in the observed cv gap on the four evaluation benchmarks, and β is benchmark-specific, so the framework's claim to 'determine' the label-shift type on those benchmarks is fitted rather than predictive. This is a test-set calibration concern that directly limits the unified-setting claim, but it does not by construction determine the final adaptation accuracy, which depends on the GMM separation, consensus pseudo-label refinement, and training objective. Hence the circularity is real but partial, applying mainly to the shift-type inference stage rather than to the overall empirical results.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method's central claims rest on several empirical heuristics: beta, delta, alpha, t, K, and the score-selection percentage are all set based on validation/benchmark behavior rather than derived from theory. These are free parameters. The axioms are domain assumptions about CLIP separability, LLM reliability, and cv thresholding that are plausible but not proven. No new physical or conceptual entities are introduced.

free parameters (7)
  • beta (β) = 0.25 (Office-31, DomainNet), 0.75 (Office-Home, VisDA)
    Balances known vs unknown similarity in score_i and rate_i. Chosen per benchmark in Section IV-B and IV-E, which is post-hoc tuning on evaluation data.
  • delta (δ) = 1.0
    Threshold for coefficient of variation to decide PDA vs OSDA/OPDA. Chosen in Section IV-E from observed separation (max PDA cv 0.7, min OSDA cv 1.1) on the evaluated benchmarks.
  • alpha (α) = 2
    Number of LLM-generated unknown classes is set to twice the number of known classes. Not tied to true unknown-class counts (e.g., Office-31 OPDA has 11 unknowns vs 20 generated). Chosen from sensitivity analysis.
  • scaling factor t = 0.01
    Rescales CLIP similarities in Eq. (2) to sharpen scores. Selected by sensitivity analysis in Fig. 8(a).
  • neighborhood size K = 5
    Number of neighbors in soft voting. Chosen in Section IV-B; Fig. 8(b) shows it has minimal impact.
  • threshold w = 0.55
    Inference threshold for predicting 'unknown' from normalized entropy, taken directly from GLC [9] rather than derived or re-tuned.
  • score selection percentage = top/bottom 30%
    For cv computation, only the top and bottom 30% of scores are used. Validated via sensitivity in Table VIII but still an arbitrary choice.
axioms (5)
  • domain assumption CLIP text-image similarities provide separable signal between known and unknown classes after rescaling.
    Section III-C and III-E; the entire score/rate metrics assume that known-class samples have higher CLIP similarity to known text labels than unknown-class samples, which may fail for visually similar classes or domain gaps.
  • ad hoc to paper The coefficient of variation of the score distribution is scale-invariant and can be thresholded uniformly across datasets.
    Section III-D and IV-E. The paper asserts that cv is comparable across benchmarks and sets a single threshold δ=1.0. This is a heuristic assumption not proven from first principles.
  • domain assumption LLM-generated unknown-class labels, when conditioned on source classes with visual similarity, improve unknown detection.
    Section III-B and Table XII show this empirically for the tested prompts, but there is no guarantee the heuristic transfers to other LLMs or domains.
  • domain assumption A two-component Gaussian mixture model fits the rate_i distribution for known/unknown separation.
    Section III-E. Assumes the metric is unimodal for known and unimodal for unknown; real distributions may be multi-modal.
  • domain assumption The source model's classifier remains frozen while only the feature extractor is fine-tuned.
    Section III-A, following SHOT/GLC protocol. The validity of this choice is not justified independently from prior work.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation." pith.science (2026). https://pith.science/paper/OHSIFD4N

@misc{pith2026260717653,
  author       = {Pith},
  title        = {Pith review of: LFM: Leveraging Foundation Models for Source-Free Universal Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OHSIFD4N}},
  note         = {Machine review of arXiv:2607.17653}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Source-free universal domain adaptation (SF-UniDA) adapts a pre-trained source model to an unlabeled target domain under both covariate and label shifts, without access to source data. However, existing SF-UniDA methods rely on inefficient techniques such as threshold tuning and clustering. Foundation models (FMs), known for their generalization and zero-shot capabilities, remain underexplored in SF-UniDA. In this paper, we propose a framework that leverages foundation models (LFM) for SF-UniDA. We use a vision-language model (VLM) to compute similarities between target samples and text labels, including those for unknown classes generated by prompting a large language model. The label shift type is determined by analyzing the coefficient of variation of a similarity-based sample-level score. Unknown samples are identified using a binary Gaussian mixture model fitted to another similarity-based metric. Under a consensus strategy, the pseudo-labels generated by the VLM are refined by the target model initialized with the pre-trained source model, integrating knowledge from both the source domain and foundation models. Finally, these refined pseudo-labels are used to train the target model. Extensive experiments across all possible label shifts and multiple benchmarks demonstrate the effectiveness and superiority of our proposed LFM framework. Our code is available at https://github.com/iamjingli/LFM.

Figures

Figures reproduced from arXiv: 2607.17653 by Fan Shi, Jianhua Zhang, Jing Li, Meng Zhao, Pan Liu, Qinghua Hu, Shengyong Chen, Wanli Xue, Xu Cheng, Yanhong Yang.

Figure 1
Figure 1. Figure 1: Method overview. Known category labels prompt the LLM to generate labels for potential unknowns, allowing CLIP to [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: LLM prompt for unknown-category text generation. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: LLM prompt to constrain the LLM response format. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Pseudo-Label Refinement. The left part shows the top-3 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rationality of cv. Each bar represents a target domain, with digits showing cv values under different label shifts. Rationality of Coefficient of Variation. As shown in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: scorei vs. ratei . Comparison of GMM-based unknown sample separation using scorei and ratei on VisDA. Real denotes the actual number of unknown samples. Benchmark (Domain) Office (A) Office-Home (Cl) VisDA (R) DomainNet (S) Time(ms) 22360 86170 291540 704510 Memory(MB) 9634 9625 9524 13712 TABLE IX: Running efficiency of LFM under the OPDA setting. Time measures the total runtime, including both one￾time p… view at source ↗
Figure 7
Figure 7. Figure 7: (a) H-score curves for 4 tasks on Office-Home under [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hyperparameter sensitivity under the OPDA setting. (a) Sensitivity to scaling factor [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE visualization for the OPDA task S→R on VisDA. F. Visualization Feature Space Visualization. For the S→R task under the OPDA setting on VisDA, t-SNE [59] is used to visualize the target domain feature space extracted by LEAD and our method. As shown in [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Confusion matrices for the PDA task Cl→Ar [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of regions vital for classification. [PITH_FULL_IMAGE:figures/full_fig_p014_11.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

60 extracted references · 5 linked inside Pith

  1. [1]

    Transfer adaptation learning: A decade survey,

    L. Zhang and X. Gao, “Transfer adaptation learning: A decade survey,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 1, pp. 23–44, 2024

    2024

  2. [2]

    Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,

    J. Li, L. Yang, and Q. Hu, “Enhancing multi-source open-set domain adaptation through nearest neighbor classification with self-supervised vision transformer,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 34, no. 4, pp. 2648–2662, 2024

    2024

  3. [3]

    Learning to transfer examples for partial domain adaptation,

    Z. Cao, K. You, M. Long, J. Wang, and Q. Yang, “Learning to transfer examples for partial domain adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2985–2994

    2019

  4. [4]

    Psdc: A prototype- based shared-dummy classifier model for open-set domain adaptation,

    Z. Liu, G. Chen, Z. Li, Y . Kang, S. Qu, and C. Jiang, “Psdc: A prototype- based shared-dummy classifier model for open-set domain adaptation,” IEEE Transactions on Cybernetics, vol. 53, no. 11, pp. 7353–7366, 2022

    2022

  5. [5]

    Universal domain adaptation,

    K. You, M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Universal domain adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 2720–2729

    2019

  6. [6]

    Universal domain adaptation through self supervision,

    K. Saito, D. Kim, S. Sclaroff, and K. Saenko, “Universal domain adaptation through self supervision,”Advances in neural information processing systems, vol. 33, pp. 16 282–16 292, 2020

    2020

  7. [7]

    Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation,

    L. Chen, Y . Lou, J. He, T. Bai, and M. Deng, “Geometric anchor correspondence mining with uncertainty modeling for universal domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 16 134–16 143

    2022

  8. [8]

    The eu general data protection regu- lation (gdpr),

    P. V oigt and A. V on dem Bussche, “The eu general data protection regu- lation (gdpr),”A Practical Guide, 1st Ed., Cham: Springer International Publishing, vol. 10, no. 3152676, pp. 10–5555, 2017

    2017

  9. [9]

    Upcycling models under domain and category shift,

    S. Qu, T. Zou, F. Röhrbein, C. Lu, G. Chen, D. Tao, and C. Jiang, “Upcycling models under domain and category shift,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 20 019–20 028

    2023

  10. [10]

    Universal domain adaptation via compressive attention matching,

    D. Zhu, Y . Li, J. Yuan, Z. Li, K. Kuang, and C. Wu, “Universal domain adaptation via compressive attention matching,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 6974–6985

    2023

  11. [11]

    Umad: Universal model adaptation under domain and category shift,

    J. Liang, D. Hu, J. Feng, and R. He, “Umad: Universal model adaptation under domain and category shift,”arXiv preprint arXiv:2112.08553, 2021

    Pith/arXiv arXiv 2021

  12. [12]

    Universal source-free domain adaptation,

    J. N. Kundu, N. Venkat, R. V . Babuet al., “Universal source-free domain adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 4544–4553

    2020

  13. [13]

    Learning to detect open classes for universal domain adaptation,

    B. Fu, Z. Cao, M. Long, and J. Wang, “Learning to detect open classes for universal domain adaptation,” inComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XV 16. Springer, 2020, pp. 567–583

    2020

  14. [14]

    Unveiling the unknown: Unleashing the power of unknown to known in open-set source-free domain adaptation,

    F. Wan, H. Zhao, X. Yang, and C. Deng, “Unveiling the unknown: Unleashing the power of unknown to known in open-set source-free domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 24 015–24 024

    2024

  15. [15]

    Lead: Learning decomposition for source-free universal domain adapta- tion,

    S. Qu, T. Zou, L. He, F. Röhrbein, A. Knoll, G. Chen, and C. Jiang, “Lead: Learning decomposition for source-free universal domain adapta- tion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 334–23 343

    2024

  16. [16]

    Domain invariant and class discriminative feature learning for visual domain adaptation,

    S. Li, S. Song, G. Huang, Z. Ding, and C. Wu, “Domain invariant and class discriminative feature learning for visual domain adaptation,”IEEE transactions on image processing, vol. 27, no. 9, pp. 4260–4273, 2018

    2018

  17. [17]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskillet al., “On the opportunities and risks of foundation models,”arXiv preprint arXiv:2108.07258, 2021

    Pith/arXiv arXiv 2021

  18. [18]

    Cross-domain open-world discovery,

    S. Wen and M. Brbic, “Cross-domain open-world discovery,” inInter- national Conference on Machine Learning. PMLR, 2024, pp. 52 744– 52 761

    2024

  19. [19]

    Source-free domain adaptation guided by vision and vision-language pre-training,

    W. Zhang, L. Shen, and C.-S. Foo, “Source-free domain adaptation guided by vision and vision-language pre-training,”International Jour- nal of Computer Vision, vol. 133, no. 2, pp. 844–866, 2025

    2025

  20. [20]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  21. [21]

    Source-free domain adaptation with frozen multimodal foundation model,

    S. Tang, W. Su, M. Ye, and X. Zhu, “Source-free domain adaptation with frozen multimodal foundation model,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 23 711–23 720. 15

    2024

  22. [22]

    Learning transfer- able features with deep adaptation networks,

    M. Long, Y . Cao, J. Wang, and M. I. Jordan, “Learning transfer- able features with deep adaptation networks,” inProceedings of the 32nd International Conference on International Conference on Machine Learning - Volume 37, ser. ICML’15. JMLR.org, 2015, p. 97–105

    2015

  23. [23]

    Deep transfer learning with joint adaptation networks,

    M. Long, H. Zhu, J. Wang, and M. I. Jordan, “Deep transfer learning with joint adaptation networks,” inInternational conference on machine learning. PMLR, 2017, pp. 2208–2217

    2017

  24. [24]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavi- olette, M. March, and V . Lempitsky, “Domain-adversarial training of neural networks,”Journal of machine learning research, vol. 17, no. 59, pp. 1–35, 2016

    2016

  25. [25]

    Domain prompt tuning via meta relabeling for unsupervised adversarial adaptation,

    X. Jin, C. Lan, W. Zeng, and Z. Chen, “Domain prompt tuning via meta relabeling for unsupervised adversarial adaptation,”IEEE Transactions on Multimedia, vol. 26, pp. 8333–8347, 2024

    2024

  26. [26]

    Cross- domain contrastive learning for unsupervised domain adaptation,

    R. Wang, Z. Wu, Z. Weng, J. Chen, G.-J. Qi, and Y .-G. Jiang, “Cross- domain contrastive learning for unsupervised domain adaptation,”IEEE Transactions on Multimedia, vol. 25, pp. 1665–1673, 2023

    2023

  27. [27]

    Coarse helps fine: A multi- granularity discriminative adversarial network for fine-grained open- set domain adaptation,

    J. Li, L. Yang, Q. Wang, and Q. Hu, “Coarse helps fine: A multi- granularity discriminative adversarial network for fine-grained open- set domain adaptation,” in2023 IEEE International Conference on Multimedia and Expo (ICME), 2023, pp. 2675–2680

    2023

  28. [28]

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,

    J. Liang, D. Hu, and J. Feng, “Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation,” inInternational conference on machine learning. PMLR, 2020, pp. 6028–6039

    2020

  29. [29]

    Self-mining the confident prototypes for source-free unsupervised domain adaptation in image segmentation,

    Y . Tian, J. Li, H. Fu, L. Zhu, L. Yu, and L. Wan, “Self-mining the confident prototypes for source-free unsupervised domain adaptation in image segmentation,”IEEE Transactions on Multimedia, vol. 26, pp. 7709–7720, 2024

    2024

  30. [30]

    Domain-division based progressive learning for source-free domain adaptation,

    P. Liu, J. Li, M. Zhao, W. Xue, Q. Hu, and S. Chen, “Domain-division based progressive learning for source-free domain adaptation,”IEEE Transactions on Multimedia, vol. 27, pp. 7081–7092, 2025

    2025

  31. [31]

    Hierarchical unsupervised relation distillation for source-free domain adaptation,

    B. Xing, X. Ying, R. Wang, R. Guo, J. Shi, and W. Yue, “Hierarchical unsupervised relation distillation for source-free domain adaptation,” in European Conference on Computer Vision. Springer, 2024, pp. 393– 409

    2024

  32. [32]

    De- confusing pseudo-labels in source-free domain adaptation,

    I. Diamant, A. Rosenfeld, I. Achituve, J. Goldberger, and A. Netzer, “De- confusing pseudo-labels in source-free domain adaptation,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 108–125

    2024

  33. [33]

    Robust nearest neighbors for source-free domain adaptation under class distribution shift,

    A. Tejero-de Pablos, R. Togashi, M. Otani, and S. Satoh, “Robust nearest neighbors for source-free domain adaptation under class distribution shift,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 1–17

    2024

  34. [34]

    Multi- granularity class prototype topology distillation for class-incremental source-free unsupervised domain adaptation,

    P. Deng, J. Zhang, X. Sheng, C. Yan, Y . Sun, Y . Fu, and L. Li, “Multi- granularity class prototype topology distillation for class-incremental source-free unsupervised domain adaptation,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30 566– 30 576

    2025

  35. [35]

    Simplifying source-free domain adaptation for object detection: Effective self-training strategies and per- formance insights,

    Y . Hao, F. Forest, and O. Fink, “Simplifying source-free domain adaptation for object detection: Effective self-training strategies and per- formance insights,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 196–213

    2024

  36. [36]

    Source-free unsu- pervised domain adaptation: A survey,

    Y . Fang, P.-T. Yap, W. Lin, H. Zhu, and M. Liu, “Source-free unsu- pervised domain adaptation: A survey,”Neural Networks, p. 106230, 2024

    2024

  37. [37]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkatet al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  38. [38]

    Universal domain adaptation from foundation models: A baseline study,

    B. Deng and K. Jia, “Universal domain adaptation from foundation models: A baseline study,”arXiv preprint arXiv:2305.11092, 2023

    Pith/arXiv arXiv 2023

  39. [39]

    Open-set domain adaptation with visual- language foundation models,

    Q. Yu, G. Irie, and K. Aizawa, “Open-set domain adaptation with visual- language foundation models,”Computer Vision and Image Understand- ing, vol. 250, p. 104230, 2025

    2025

  40. [40]

    Decoupling domain invariance and variance with tailored prompts for open-set domain adaptation,

    S. Zeng, X. Liu, and Y . Zhou, “Decoupling domain invariance and variance with tailored prompts for open-set domain adaptation,” in2024 IEEE International Conference on Image Processing (ICIP), 2024, pp. 645–651

    2024

  41. [41]

    Cosmo: Clip talks on open-set multi-target domain adaptation,

    M. Monga, S. K. Giroh, A. Jha, M. Singha, B. Banerjee, and J. Chanussot, “Cosmo: Clip talks on open-set multi-target domain adaptation,” in35th British Machine Vision Conference 2024, BMVC 2024, Glasgow, UK, November 25-28, 2024. BMV A, 2024. [Online]. Available: https://papers.bmvc2024.org/0031.pdf

    2024

  42. [42]

    Adversarial experts model for black-box domain adaptation,

    S. Xiao, M. Ye, Q. He, S. Li, S. Tang, and X. Zhu, “Adversarial experts model for black-box domain adaptation,” inProceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 8982–8991

    2024

  43. [43]

    Envisioning outlier exposure by large language models for out-of-distribution detec- tion,

    C. Cao, Z. Zhong, Z. Zhou, Y . Liu, T. Liu, and B. Han, “Envisioning outlier exposure by large language models for out-of-distribution detec- tion,” inProceedings of the 41st International Conference on Machine Learning, vol. 235. PMLR, 21–27 Jul 2024, pp. 5629–5659

    2024

  44. [44]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778

    2016

  45. [45]

    Monitoring the coefficient of variation: A literature review,

    Z. Jalilibal, A. Amiri, P. Castagliola, and M. B. Khoo, “Monitoring the coefficient of variation: A literature review,”Computers & Industrial Engineering, vol. 161, p. 107600, 2021. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0360835221005040

    2021

  46. [46]

    Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation,

    M. Litrico, A. Del Bue, and P. Morerio, “Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 7640–7650

    2023

  47. [47]

    Adapting visual cate- gory models to new domains,

    K. Saenko, B. Kulis, M. Fritz, and T. Darrell, “Adapting visual cate- gory models to new domains,” inComputer Vision–ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11. Springer, 2010, pp. 213–226

    2010

  48. [48]

    Deep hashing network for unsupervised domain adaptation,

    H. Venkateswara, J. Eusebio, S. Chakraborty, and S. Panchanathan, “Deep hashing network for unsupervised domain adaptation,” inPro- ceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 5018–5027

    2017

  49. [49]

    Visda: The visual domain adaptation challenge,

    X. Peng, B. Usman, N. Kaushik, J. Hoffman, D. Wang, and K. Saenko, “Visda: The visual domain adaptation challenge,”arXiv preprint arXiv:1710.06924, 2017

    Pith/arXiv arXiv 2017

  50. [50]

    Moment matching for multi-source domain adaptation,

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

    2019

  51. [51]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems, I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, Eds., vol. 30. Curran Associates, Inc., 2017, pp. 1–11

    2017

  52. [52]

    Billion-scale similarity search with GPUs,

    J. Johnson, M. Douze, and H. Jégou, “Billion-scale similarity search with GPUs,”IEEE Transactions on Big Data, vol. 7, no. 3, pp. 535– 547, 2019

    2019

  53. [53]

    Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set do- main adaptation,

    J. Li, L. Yang, Q. Wang, and Q. Hu, “Wdan: A weighted discriminative adversarial network with dual classifiers for fine-grained open-set do- main adaptation,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 33, no. 9, pp. 5133–5147, 2023

    2023

  54. [54]

    Domain consensus clustering for universal domain adaptation,

    G. Li, G. Kang, Y . Zhu, Y . Wei, and Y . Yang, “Domain consensus clustering for universal domain adaptation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 9757–9766

    2021

  55. [55]

    Ovanet: One-vs-all network for universal do- main adaptation,

    K. Saito and K. Saenko, “Ovanet: One-vs-all network for universal do- main adaptation,” inProceedings of the ieee/cvf international conference on computer vision, 2021, pp. 9000–9009

    2021

  56. [56]

    Unified optimal transport framework for universal domain adaptation,

    W. Chang, Y . Shi, H. Tuan, and J. Wang, “Unified optimal transport framework for universal domain adaptation,”Advances in Neural Infor- mation Processing Systems, vol. 35, pp. 29 512–29 524, 2022

    2022

  57. [57]

    Conditional adversarial domain adaptation,

    M. Long, Z. Cao, J. Wang, and M. I. Jordan, “Conditional adversarial domain adaptation,” inProceedings of the 32nd International Confer- ence on Neural Information Processing Systems, 2018, pp. 1647–1657

    2018

  58. [58]

    Bridging theory and algo- rithm for domain adaptation,

    Y . Zhang, T. Liu, M. Long, and M. Jordan, “Bridging theory and algo- rithm for domain adaptation,” inInternational conference on machine learning. PMLR, 2019, pp. 7404–7413

    2019

  59. [59]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.”Journal of machine learning research, vol. 9, no. 11, 2008

    2008

  60. [60]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inProceedings of the IEEE international conference on computer vision, 2017, pp. 618–626

    2017

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