Pith. sign in

REVIEW 3 major objections 5 minor 60 references

COSMO: Consensus-Driven Shift Modulation for Source-Free Domain Adaptation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read COSMO adapts source-free models by learning from an anchored shared consensus instead of letting one expert teach the other.

desk verdict A well-engineered anchored-consensus method for VLM-guided SFDA with sound math and honest ablations, but the SOTA claim is not yet established because the baselines are taken from their own papers without matched source checkpoints or repeated runs. read the letter →

arxiv 2608.04604 v1 pith:TJXTLOSA submitted 2026-08-05 cs.CV

Bo Li , Junjie Peng , Xiaohua Xie , Jianhuang Lai This is my paper · ORCID
classification cs.CV
keywords source-freedomainadaptationvision-languagemodelsconsensuslearningproductofexpertsentropy-conditionedweightingpromptshiftmodulationreliabilityallocation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

COSMO is a method for source-free domain adaptation that combines a frozen source-trained model with a pretrained vision-language model (VLM) to adapt to an unlabeled target domain. The paper argues that existing VLM-guided approaches let one expert directly supervise the other, which can overwrite valid source-derived evidence when the experts disagree. COSMO instead builds a sample-specific consensus prediction that both branches learn from, and it regulates how far that consensus moves from its initial state. A sympathetic reader would care because the method reports state-of-the-art results on four benchmarks under matched VLM backbones while retaining more correct source-only decisions than direct distillation.

What carries the argument

The central object is the entropy-conditioned reverse-KL barycenter, defined as the minimizer of a weighted sum of reverse KL divergences to the expert predictions, whose closed form is a weighted product of experts in probability space and a weighted mean in centered-logit space. The weights are proportional to $\max\{\log K - H(p_m), \epsilon\}$, measuring each expert's predictive concentration relative to uniformity. This consensus is computed twice: once with the frozen source expert and initial VLM to form the cached anchor $c_i^{(0)}$, and once each step with the current target and VLM branches to form the dynamic consensus $c_i^{(\ell)}$. Consensus Shift Modulation then rescales the displacement $\Delta c_i^{(\ell)} = c_i^{(\ell)} - c_i^{(0)}$ by a factor $\gamma_i^{(e)} = 1 + \lambda d^{(e)} u_i^{(e)}$ that depends on the sample's cross-sample entropy rank and a temporal decay, keeping the supervision anchored yet adaptive.

What would settle it

Construct a target set where the source expert and the VLM both make confident but identical wrong predictions on a large fraction of samples (for example, a synthetic domain shift that flips a visual attribute the source model over-relies on and the VLM has also learned). If COSMO's entropy-conditioned anchoring prevents the error from being corrected and accuracy stays below a method without anchoring, the reliability of the initial anchor as a reference point is refuted.

Watch

Extended reading notes

Core claim

The central claim is that VLM-guided source-free domain adaptation is best treated as a sample-wise reliability-allocation problem, and that reliability can be operationalized through an entropy-conditioned product of experts. For each sample, COSMO forms an initial consensus from the frozen source expert and the initially prompted VLM, weighting each expert by its predictive concentration relative to uniformity. During adaptation, the target branch and the prompt-adapted VLM re-aggregate their evolving predictions into a dynamic consensus, which is then shifted from the initial anchor by a bounded factor that contracts low-entropy samples and extends high-entropy samples. The paper reports that this procedure retains 78.1% of source-only correct decisions versus 42.8% for direct VLM distillation, while absorbing 91.1% of VLM-only correct decisions, and achieves state-of-the-art accuracy on Office-31, Office-Home, VisDA-C, and DomainNet-126 under matched backbone conditions.

Load-bearing premise

The initial consensus anchor is assumed to be reliable enough that low-entropy target samples should have their evolving predictions pulled back toward it, yet if both the source expert and the more concentrated VLM prediction are confidently wrong on a sample, the anchor encodes that error and CSM actively suppresses correction for that sample.

Editorial extensions

If this is right

  • If COSMO is correct, source-free adaptation systems can integrate semantic knowledge from frozen VLMs without risking that the VLM's confident errors overwrite valid source-derived decisions on individual samples.
  • The entropy-conditioned product of experts provides a closed-form, label-free way to allocate trust between heterogeneous experts per sample, which can be reused in other settings where two predictive models must be fused without calibration.
  • The finding that source-only retention rises from 42.8% to 78.1% with consensus-based supervision implies that the source model's valid evidence is not merely noise to be replaced but a reusable asset during adaptation.
  • Because only the adapted target branch is kept at inference, the co-adaptation machinery adds no inference-time cost, making the approach practical for deployment.
  • The method extends beyond two experts: a three-expert configuration with CLIP and InstructBLIP outperforms a multi-stage distillation baseline, suggesting the consensus formulation scales to additional foundation models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An implicit consequence is that the same anchored-consensus idea could apply to any pair of complementary predictors beyond source models and VLMs, such as an ensemble of architectures with different inductive biases, where the initial anchor would play the role of a prior that prevents catastrophic forgetting of pre-adaptation competence.
  • The cross-sample entropy rank used by CSM is a fully unsupervised curriculum signal; a testable extension would be to replace the rank over the whole target set with an online or batch-local estimate in streaming settings, where the paper already shows batch-local ranks match or exceed the canonical implementation at lower cost.
  • The decision-retention conditions derived for the initial consensus suggest a diagnostic: one could measure, for a given source-VLM pair and dataset, the fraction of samples where the weighted margin condition holds, and predict whether consensus pooling will preserve source-correct decisions before running adaptation.
  • The paper's own framing treats entropy as a proxy for concentration, not correctness; a further testable question is whether a calibrated uncertainty estimate, if available, would improve the same consensus framework beyond what raw entropy achieves.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces COSMO, a source-free domain adaptation method in which a source-initialized target branch and a prompt-adapted VLM are co-adapted through a shared consensus rather than through direct expert-to-expert guidance. At initialization, an entropy-weighted product of experts (reverse-KL barycenter) pools the frozen source expert and the initially prompted VLM into a cached anchor. During adaptation, the current target and VLM predictions are re-aggregated into a dynamic consensus, whose displacement from the anchor is rescaled by Consensus Shift Modulation (CSM) according to the sample's rank in the consensus-entropy distribution, with a temporal decay to identity. The paper reports state-of-the-art results under matched CLIP backbones on Office-31, Office-Home, VisDA-C, and DomainNet-126, together with ablations separating supervision topology, dynamic re-aggregation, CSM, pooling geometry, and objective terms.

Significance. If the empirical claims hold, the anchored shared-consensus formulation is a conceptually valuable alternative to directed cross-model distillation in VLM-guided SFDA. The paper has notable strengths: the mathematical derivation of the entropy-conditioned reverse-KL barycenter and its weighted-product-of-experts form (Prop. A.2 and Cor. A.3) is complete and correct; the decision-retention conditions (Props. A.5-A.7) are carefully stated; the algorithm is fully specified with benchmark-level hyperparameters, run-level results, and explicit statements about label use; and the conflict-conditioned retention/absorption analysis directly targets the proposed 'source-derived evidence forgetting' failure. The main weakness is that the headline state-of-the-art claim rests on cross-paper baseline numbers with mismatched source checkpoints and protocols, a comparison gap that is acknowledged in Appendix E.2 but not resolved.

major comments (3)
  1. [§4.2, Tables 1–2; Appendix E.2, Table S5] The headline claim of state-of-the-art performance under matched VLM backbones is not yet supported at the reported precision. COSMO's numbers come from the authors' own source checkpoints and protocol with three seeds, while the VLM-guided baselines (DIFO, ProDe, VSFOT) mostly retain the model-selection, repetition, and rounding conventions of their original papers. Only ProDe on DomainNet-126 is reproduced, and it is a single run with seed 2020. In SFDA, target accuracy is highly sensitive to the source checkpoint and target-adaptation protocol, and 'matched VLM backbone' controls only the CLIP image encoder. With margins as small as 0.4 points on VisDA-C (Table 1) and COSMO's own run-level standard deviation averaging 0.199 points on Office-Home (Table S8), the reported gains cannot be distinguished from protocol variation without paired re-runs. Please either re-run the compared source-free methods under the same source checkpoints and protocol with multiple seeds, or rephrase the contribution as competitive with previously reported numbers and remove the unqualified SOTA wording. The ViT-B/16 DIFO values in Table S9, which are taken from ProDe rather than from the original DIFO paper, should also be flagged as approximate.
  2. [§3.4, Eq. (6); §4.4, Fig. 4] CSM's core premise is that low-consensus-entropy samples should have their anchor-relative displacement contracted, and this premise is only partially validated. The directional validation in Fig. 4 is performed on a single transfer (Office-Home A→C), and the entropy-order informativeness analysis in Fig. 3 is limited to Office-Home. More importantly, the paper does not diagnose the failure mode in which the initial consensus is confidently wrong: because the anchor is formed once from the frozen source expert and the initially prompted VLM, a confidently wrong anchor would cause CSM to suppress correction for exactly the low-entropy samples that most need it. The paper correctly states that entropy is not calibrated correctness, but CSM's rank-based modulation still relies on consensus entropy as a reliability signal. Please add a per-task or per-subset analysis of cases where the initial consensus is wrong but low-entropy, and show whether the proposed contraction still helps; if it does not, qualify the mechanism claim accordingly.
  3. [§4.3, Table 4; Appendix H.2] The design decisions—entropy as the within-sample allocation cue, the direction of CSM, and the fixed coefficients such as λ=0.5—are justified by ablations and sensitivity analyses performed on the same four benchmarks whose final numbers are then reported. This creates a selection-risk concern: the reported component gains are descriptive of the chosen benchmarks rather than predictive of new ones. The paper should explicitly state, in the main text, that these choices were not tuned per transfer and ideally should provide a small validation-based selection procedure or hold out at least one benchmark when justifying each component. As written, the claim that entropy weighting is 'best' (Table 4) is an observation on the evaluated benchmarks, not a general finding.
minor comments (5)
  1. [References; §1] The citation 'Peng et al., 2026' in §1, used for 'correct VLM predictions,' points in the reference list to 'Grkv: Global regression for training-free KV cache compression in long-context llms,' an unrelated paper on KV-cache compression. Please verify and replace this citation.
  2. [Appendix F.4, Table S9] The footnote that the ViT-B/16 DIFO configuration 'was not reported in the original DIFO paper' is important and should also appear near the main-text comparison table, since Table 1 in the main paper lists DIFO without this caveat.
  3. [§4.4, Fig. 5] Figure 5 reports retention/absorption percentages without run-level variance; since Figure 3 includes confidence intervals for the entropy analyses, please add error bars or state explicitly that the conflict-conditioned values are single-run or aggregated without variance.
  4. [§4.4, Fig. 4] The signed-coefficient directional validation is on Office-Home A→C only; the main text should state this explicitly rather than leaving it to the appendix, to prevent readers from inferring multi-benchmark validation of the CSM direction.
  5. [§1 and §4.4] The term 'source-derived evidence forgetting' is introduced informally; a one-sentence formal definition tied to the retention metric RS̄V in Fig. 5 would improve precision.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivation-level circularity; the consensus and modulation steps are explicit, self-contained functions of the branches’ own pre-update predictions, and the only self-citation is unrelated and non-load-bearing.

full rationale

The paper’s derivation chain is not circular by construction. The initial consensus is defined as a closed-form reverse-KL barycenter / weighted product of experts (Eqs. 2–4, Prop. A.2); the dynamic consensus and Consensus Shift Modulation (Eqs. 5–6 and S49–S50) are explicit algebraic transformations of the current target and VLM logits, with the anchor formed once from the frozen source expert and the initially prompted VLM. The IIC/CE objective in Eq. 7 supervises both branches with a detached snapshot, which is a self-training bootstrap rather than a reduction of the reported accuracy to the method’s inputs. The decision-retention conditions in Props. A.5–A.7 are mathematical characterizations of when the weighted product of experts keeps a given top-1 label, and the paper explicitly states they characterize retention, not correctness. The selection of expert-entropy PoE in Table 4 and the CSM direction in Fig. 4 is based on target-benchmark evaluation; this is benchmark-driven design selection, not an equation-level circularity, and is more appropriately classified as a generalization/overfitting concern. Likewise, the E.2 statement that baseline values retain their original model-selection, repetition, and rounding conventions (with only ProDe on DomainNet-126 reproduced) limits the strength of the SOTA claim but is not circularity. One reference, Peng et al. (2026), is a self-citation by co-author Junjie Peng and is cited in Sec. 1 for “correct VLM predictions”; the cited work is actually a KV-cache-compression paper. This citation is incorrect and unrelated, but it is not load-bearing for COSMO’s derivation, so per hard rule 4 it contributes at most a minor non-load-bearing self-citation, not derivation circularity.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method's design rests on two ad hoc modeling choices: entropy-conditioned concentration as a reliability cue and CSM's entropy-rank-based shift modulation. The closed-form consensus derivation is standard math, but the weighting rule and the modulation schedule are not derived from first principles. Several loss coefficients and the modulation strength are hand-set per benchmark, which adds free parameters to the central claim.

free parameters (5)
  • lambda (CSM modulation strength) = 0.5
    Bounds the modulation factor gamma in Eq. (5); fixed across all benchmarks but chosen after directional validation on Office-Home A to C (Fig. 4).
  • alpha (target IIC weight) = 1.3 on Office-31, Office-Home, DomainNet-126; 1.0 on VisDA-C
    Weights target-branch consensus alignment in Eq. (7); benchmark-specific values from Table S2.
  • beta (hard CE weight) = 0.1 on Office-31; 0.4 on Office-Home and DomainNet-126; 0.05 on VisDA-C
    Weights hard-label classification in Eq. (7); set per benchmark.
  • delta (diversity weight) = 1.0 on Office-31 and Office-Home; 0.1 on VisDA-C; 0.5 on DomainNet-126
    Entropy regularization on target batch marginal in Eq. (7); set per benchmark.
  • epsilon (entropy-weight stabilizer) = 1e-5 on Office-31, Office-Home, VisDA-C; 1e-3 on DomainNet-126
    Numerical floor for concentration scores in Eq. (2); set per benchmark.
assumptions (5)
  • domain assumption Closed-set domain adaptation: source and target share the same K classes
    Stated in Sec. 3.1; the method and evaluation depend on identical label sets.
  • ad hoc to paper Entropy (KL to uniform) is an informative proxy for relative expert reliability within a sample
    Invoked in Eq. (2) and CSM Eq. (5); validated post hoc on the same benchmarks, not derived from first principles.
  • ad hoc to paper The fixed initial consensus anchor is a reliable reference that should not be drifted from for low-uncertainty samples
    CSM contraction assumes the anchor preserves valid source evidence; no formal guarantee that the anchor is correct.
  • standard math Weighted reverse-KL barycenter has a unique closed form via Lagrange multiplier (Prop. A.2)
    Standard convex optimization on the simplex; proof given in Appendix A.2.
  • domain assumption IIC mutual information objective provides useful clustering and alignment signal
    Taken from Ji et al. 2019 and used in Eq. (7); not re-derived here.

how reviews work

0 comments
Cite this review

Pith. "Pith review of COSMO: Consensus-Driven Shift Modulation for Source-Free Domain Adaptation." pith.science (2026). https://pith.science/paper/TJXTLOSA

@misc{pith2026260804604,
  author       = {Pith},
  title        = {Pith review of: COSMO: Consensus-Driven Shift Modulation for Source-Free Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TJXTLOSA}},
  note         = {Machine review of arXiv:2608.04604}
}
read the original abstract

Source-free domain adaptation (SFDA) adapts a source-trained model to an unlabeled target domain without source data, a practical setting under privacy or storage constraints. Yet its self-generated supervision can reinforce source bias under substantial domain shifts. Pretrained vision-language models (VLMs) offer complementary semantic knowledge, but the relative reliability of the source model and VLM varies across target samples. Existing cross-model guidance does not explicitly account for this variation and may overwrite valid source-derived evidence under conflict, a failure we term source-derived evidence forgetting. We formulate VLM-guided SFDA as a sample-wise reliability-allocation problem and propose Consensus-Driven Shift Modulation (COSMO). COSMO replaces expert-to-expert guidance with co-adaptation through an anchored shared consensus. It first forms a sample-specific initial consensus that favors the more concentrated prediction. During adaptation, COSMO re-aggregates both branches' evolving evidence and regulates how far the resulting consensus moves from its initial anchor based on consensus uncertainty and training progress. This keeps the shared supervision anchored yet adaptive. Across four benchmarks, COSMO achieves state-of-the-art performance under matched VLM backbones. Further analyses indicate that it better balances the retention of valid source-derived evidence with the absorption of complementary VLM evidence.

Figures

Figures reproduced from arXiv: 2608.04604 by the authors.

Figure 1
Figure 1. Conceptual comparison of VLM-guided SFDA. Ellipses summarize target-domain predictions; bars show an illustrative sample. (a) Cross-model guidance may overwrite valid source-derived evidence under expert conflict. (b) COSMO instead learns from an evolving shared consensus, retaining source￾derived evidence while integrating complementary VLM evidence. target structure through pseudo-labels, prototypes, neighborhoods… view at source ↗
Figure 2
Figure 2. Overview of COSMO. (a) For each target sample, the frozen source expert and initially prompted VLM form a cached consensus anchor, while the current target and prompt-adapted VLM branches form a dynamic consensus. CSM modulates its anchor-relative displacement, and the detached result supervises both branches. (b) Entropy-conditioned weighting yields a weighted product of experts that emphasizes more concentrated pr… view at source ↗
Figure 3
Figure 3. Informativeness of the two entropy cues on Office-Home. (a) Source-expert win rate among samples where exactly one initial expert is correct, binned by ∆H = HV − HS; the dashed line denotes the overall win rate. (b) Consensus error rate across consensus-entropy quintiles before and after adapta￾tion. Curves show 12-shift means with 95% confidence intervals; boxed values report macro AUROC. Labels are used only for p… view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Conflict-conditioned outcomes on Office [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 4
Figure 4. Figure 4: Directional validation of CSM on Office [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 50 canonical work pages

  1. [1]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    Domain impression: A source data free domain adaptation method , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    work page

  2. [2]

    IEEE Transactions on Circuits and Systems for Video Technology , volume=

    VDM-DA: Virtual domain modeling for source data-free domain adaptation , author=. IEEE Transactions on Circuits and Systems for Video Technology , volume=. 2021 , publisher=

    work page 2021

  3. [3]

    Machine Learning , volume=

    Generation, augmentation, and alignment: A pseudo-source domain based method for source-free domain adaptation , author=. Machine Learning , volume=. 2024 , publisher=

    work page 2024

  4. [4]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Source-free domain adaptation via distribution estimation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  5. [5]

    Pattern recognition , volume=

    Source-free domain adaptation with class prototype discovery , author=. Pattern recognition , volume=. 2024 , publisher=

    work page 2024

  6. [6]

    International journal of computer vision , volume=

    Source-free domain adaptation via target prediction distribution searching , author=. International journal of computer vision , volume=. 2024 , publisher=

    work page 2024

  7. [7]

    International conference on machine learning , pages=

    Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  8. [8]

    Advances in neural information processing systems , volume=

    Exploiting the intrinsic neighborhood structure for source-free domain adaptation , author=. Advances in neural information processing systems , volume=

Show all 60 references
  1. [9]

    2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=

    Self-supervised noisy label learning for source-free unsupervised domain adaptation , author=. 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages=. 2022 , organization=

    2022

  2. [10]

    The Eleventh International Conference on Learning Representations , year=

    When Source-Free Domain Adaptation Meets Learning with Noisy Labels , author=. The Eleventh International Conference on Learning Representations , year=

  3. [11]

    International conference on machine learning , pages=

    Confidence score for source-free unsupervised domain adaptation , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  4. [12]

    Advances in neural information processing systems , volume=

    Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data , author=. Advances in neural information processing systems , volume=

  5. [13]

    Advances in Neural Information Processing Systems , volume=

    Divide and contrast: Source-free domain adaptation via adaptive contrastive learning , author=. Advances in Neural Information Processing Systems , volume=

  6. [14]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Guiding pseudo-labels with uncertainty estimation for source-free unsupervised domain adaptation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  7. [15]

    Advances in Neural Information Processing Systems , volume=

    Attracting and dispersing: A simple approach for source-free domain adaptation , author=. Advances in Neural Information Processing Systems , volume=

  8. [16]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Class relationship embedded learning for source-free unsupervised domain adaptation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  9. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Source-free domain adaptation with frozen multimodal foundation model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  10. [18]

    International Journal of Computer Vision , volume=

    Source-free domain adaptation guided by vision and vision-language pre-training , author=. International Journal of Computer Vision , volume=. 2025 , publisher=

    2025

  11. [19]

    The Thirteenth International Conference on Learning Representations , year=

    Proxy Denoising for Source-Free Domain Adaptation , author=. The Thirteenth International Conference on Learning Representations , year=

  12. [20]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vision-Language Model Guided Source-Free Domain Adaptation via Optimal Transport , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  13. [21]

    2025 , eprint=

    Collaborative Learning with Multiple Foundation Models for Source-Free Domain Adaptation , author=. 2025 , eprint=

    2025

  14. [22]

    European conference on computer vision , pages=

    Adapting visual category models to new domains , author=. European conference on computer vision , pages=. 2010 , organization=

    2010

  15. [23]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Deep hashing network for unsupervised domain adaptation , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  16. [24]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , pages=

    Visda: A synthetic-to-real benchmark for visual domain adaptation , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops , pages=

  17. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Semi-supervised domain adaptation via minimax entropy , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [26]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Moment matching for multi-source domain adaptation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  19. [27]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Generalized source-free domain adaptation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  20. [28]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  21. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Padclip: Pseudo-labeling with adaptive debiasing in clip for unsupervised domain adaptation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  22. [30]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =

    Singha, Mainak and Pal, Harsh and Jha, Ankit and Banerjee, Biplab , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =. 2023 , pages =

    2023

  23. [31]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Prompt-based distribution alignment for unsupervised domain adaptation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  24. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Domain-agnostic mutual prompting for unsupervised domain adaptation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  25. [33]

    International conference on machine learning , pages=

    Unsupervised domain adaptation by backpropagation , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  26. [34]

    Advances in neural information processing systems , volume=

    Conditional adversarial domain adaptation , author=. Advances in neural information processing systems , volume=

  27. [35]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adaptive adversarial network for source-free domain adaptation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  28. [36]

    Neural Networks , volume=

    Semantic consistency learning on manifold for source data-free unsupervised domain adaptation , author=. Neural Networks , volume=. 2022 , publisher=

    2022

  29. [37]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Invariant information clustering for unsupervised image classification and segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  30. [38]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Contrastive test-time adaptation , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  31. [39]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    Domain adaptation via prompt learning , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2023 , publisher=

    2023

  32. [40]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Tinyclip: Clip distillation via affinity mimicking and weight inheritance , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  33. [41]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Clip-kd: An empirical study of clip model distillation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  34. [42]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mobileclip: Fast image-text models through multi-modal reinforced training , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  35. [43]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Customkd: Customizing large vision foundation for edge model improvement via knowledge distillation , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  36. [44]

    2024 , url=

    Size Wu and Wenwei Zhang and Lumin Xu and Sheng Jin and Xiangtai Li and Wentao Liu and Chen Change Loy , booktitle=. 2024 , url=

    2024

  37. [45]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Feng, Qianhan and Li, Wenshuo and Lin, Tong and Chen, Xinghao , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  38. [46]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Sameni, Sepehr and Kafle, Kushal and Tan, Hao and Jenni, Simon , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  39. [47]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Jang, Jinseong and Ma, Chunfei and Lee, Byeongwon , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  40. [48]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Li, Zheng and Li, Xiang and Fu, Xinyi and Zhang, Xin and Wang, Weiqiang and Chen, Shuo and Yang, Jian , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  41. [49]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Kim, Sanghwan and Xiao, Rui and Georgescu, Mariana-Iuliana and Alaniz, Stephan and Akata, Zeynep , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  42. [50]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Cao, Jiajun and Zhang, Yuan and Huang, Tao and Lu, Ming and Zhang, Qizhe and An, Ruichuan and Ma, Ningning and Zhang, Shanghang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  43. [51]

    Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,

    Towards Dynamic-Prompting Collaboration for Source-Free Domain Adaptation , author =. Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence,. 2024 , month =. doi:10.24963/ijcai.2024/182 , url =

    2024 doi

  44. [52]

    2025 , eprint=

    ViLAaD: Enhancing "Attracting and Dispersing'' Source-Free Domain Adaptation with Vision-and-Language Model , author=. 2025 , eprint=

    2025

  45. [53]

    Neurocomputing , volume=

    Language-guided alignment and distillation for source-free domain adaptation , author=. Neurocomputing , volume=. 2025 , publisher=

    2025

  46. [54]

    Pattern Recognition , pages=

    Source-Free Domain Adaptation via Multimodal Space-Guided Alignment , author=. Pattern Recognition , pages=. 2025 , publisher=

    2025

  47. [55]

    2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Empowering Source-Free Domain Adaptation via MLLM-Guided Reliability-Based Curriculum Learning , author=. 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2026 , organization=

    2026

  48. [56]

    2026 , eprint=

    Source-Free Domain Adaptation with Vision-Language Prior , author=. 2026 , eprint=

    2026

  49. [57]

    Instruct

    Wenliang Dai and Junnan Li and Dongxu Li and Anthony Tiong and Junqi Zhao and Weisheng Wang and Boyang Li and Pascale Fung and Steven Hoi , booktitle=. Instruct. 2023 , url=

    2023

  50. [58]

    2023 , eprint=

    Towards General Text Embeddings with Multi-stage Contrastive Learning , author=. 2023 , eprint=

    2023

  51. [59]

    Journal of Machine Learning Research , author =

    Visualizing. Journal of Machine Learning Research , author =. 2008 , pages =

    2008

  52. [60]

    arXiv preprint arXiv:2605.31105 , year=

    GRKV: Global Regression for Training-Free KV Cache Compression in Long-Context LLMs , author=. arXiv preprint arXiv:2605.31105 , year=

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.