Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Harmonizing and Merging Source Models for CLIP-based Domain Generalization

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read HAM claims that per-source CLIP fine-tuning with confidence-filtered enrichment, directional alignment, and redundancy-aware merging generalizes better than any single trained model.

desk verdict A sensible learning-then-merging recipe for CLIP-DG, but the core RHM equations are internally inconsistent and the reported numbers are not reproducible as written; fix and resubmit. read the letter →

arxiv 2506.09446 v1 pith:MFKINZA3 submitted 2025-06-11 cs.CV

classification cs.CV
keywords domaingeneralizationCLIPmodelmergingsourceenrichmentparameteralignmentredundancy-awaremulti-sourcetrainingvision-languagemodels
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

The paper proposes HAM, a learning-then-merging strategy for CLIP-based domain generalization. Instead of training one model on all source domains jointly, it fine-tunes a separate CLIP vision encoder for each source domain, enriches each domain with cross-domain samples the model itself judges non-conflicting, aligns update directions across models, and merges the resulting trajectories into one final model. The central claim is that this avoids both sample conflicts (noise, multi-object interference, extreme domain shift) and optimization conflicts (competing gradients in multi-source training), leading to better generalization on unseen domains. On five standard benchmarks, HAM reports 79.0% average accuracy with the ViT-B/16 backbone and 70.9% with ResNet50, outperforming single-model baselines and prior CLIP-based domain generalization methods. If correct, this gives a practical recipe for better unseen-domain accuracy from CLIP while keeping a single model at inference.

What carries the argument

The argument is carried by three modules. Sample Conflict-aware Adaptive Source Enrichment (SAE) computes an adaptive confidence threshold $ au_i = \frac{1}{n}\sum_k \max f_i(x_k)$ on each mini-batch and admits cross-domain samples whose confidence on the current model exceeds it, so each domain trains on useful non-conflicting data. Optimization Conflict-aware Parameter Alignment (OPA) adds a loss $L_{\mathrm{sign}} = \frac{1}{N_L}\sum_l \max(0, -v_i^l \cdot \bar{v}^l)$ that penalizes any layer whose update vector opposes the average update vector across sources. Redundancy-aware Historical Model Merging (RHM) maintains a Beta-weighted moving average of each model over training, then trims small-magnitude entries of the flattened update vector by a global percentile threshold and merges only the surviving entries. RHM turns many per-domain trajectories into a single final model, and the paper's ablations attribute the largest single gain to SAE.

What would settle it

Take one benchmark source domain, flip the labels of its highest-confidence images, run HAM, and check whether the SAE filter admits those confidently wrong samples into other domains and whether the final merged model's target-domain accuracy drops. If the filter passes such samples and performance degrades, the confidence criterion fails as a conflict detector.

Watch

Extended reading notes

Core claim

The paper's central claim is that the two conflicts in multi-source CLIP-based domain generalization are best resolved by separating learning from merging. HAM trains each source model independently on enriched data, with a sign-consistency regularization on model update vectors so that merging does not cancel opposing updates. It then averages each model over its training trajectory with Beta-distributed weights and applies global percentile-based trimming before averaging the trimmed updates across domains back onto the CLIP initialization. On PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet, the paper reports 79.0% average accuracy with ViT-B/16 and 70.9% with ResNet50, exceeding reproduced baselines, and 66.4% average with an ImageNet-pretrained ResNet50, indicating the recipe transfers outside CLIP.

Load-bearing premise

The whole method depends on the belief that the model's own confident predictions are trustworthy guides for which cross-domain samples to add to training.

Editorial extensions

If this is right

  • Unseen-domain accuracy for CLIP can be improved by merging per-source fine-tunes instead of optimizing all sources jointly.
  • The final model is a single vision encoder initialized from CLIP, so deployment needs no ensemble storage or multiple inference passes.
  • On the five benchmark datasets, the method raises average accuracy to 79.0% with ViT-B/16 and 70.9% with ResNet50, with notable gains on datasets with severe domain shift.
  • Because HAM also improves an ImageNet-pretrained ResNet50 to 66.4% average accuracy, the merging recipe is not tied to CLIP's text encoder.
  • Parameter-level merging gives a practical alternative to ensemble learning for domain generalization, requiring one model at deployment time.

Reading between the lines

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

  • A testable extension the paper leaves implicit is replacing max-softmax confidence with a calibrated or OOD-aware score, since confidently wrong noisy labels are exactly where the confidence filter is least reliable.
  • The sign loss only penalizes opposing updates; an angular or magnitude-aware penalty could align directions more finely and might reduce sensitivity to the trimming ratio.
  • Global percentile trimming outperformed per-layer trimming in the reported ablations, suggesting redundancy is a property of the whole update vector across layers, which could be tested with Fisher-weighted or Hessian-informed pruning.
  • The recipe implies a generic template for foundation-model domain generalization — per-domain light fine-tuning, trajectory averaging, and sparse parameter merging — that could be tried on other pretrained encoders.
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

4 major / 6 minor

Summary. The paper proposes HAM, a learning-then-merging framework for CLIP-based domain generalization. Instead of training one model on all source domains, HAM fine-tunes a separate CLIP visual encoder per source domain while keeping the text encoder frozen, then merges the source models. Three modules are introduced: (1) SAE, which enriches each domain's mini-batches with cross-domain samples whose confidence under the current model exceeds an adaptive threshold; (2) OPA, a sign-consistency regularizer on parameter update vectors that penalizes deviations from the average update direction; and (3) RHM, which averages historical checkpoints within each source and then merges across sources after trimming redundant parameters. Experiments are reported on PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet, with average accuracies of 70.9% on ResNet50 and 79.0% on ViT-B/16, which the paper claims are state-of-the-art. A further experiment on ImageNet-pretrained ResNet-50 reports 66.4% average accuracy.

Significance. If the method works as described, the paper makes a useful contribution: it provides a concrete framework for resolving sample-level and optimization-level conflicts in multi-source CLIP DG, and its ablation strategy is systematic. The benchmarking is broad, covering five standard DG datasets on two CLIP backbones plus a non-CLIP backbone, and the three modules are ablated separately and jointly. The paper also positions itself clearly relative to model merging and OOD generalization literature. However, the central reproducibility and SOTA claims are currently blocked by an internal inconsistency in the RHM equations, missing uncertainty estimates on small margins, and a reliance on an unvalidated confidence-as-quality assumption in SAE. With those points corrected, the paper would be a solid candidate for the journal.

major comments (4)
  1. [3.4, Eqs. (10)-(13)] The redundancy-aware merging operation is internally inconsistent as written. Eq. (10) defines sigma as the r-th percentile of parameter magnitudes, using a vector tilde v_cat that is never defined in the text; Eq. (11) then forms the mask m = I(v_cat > sigma). Since sigma is a nonnegative percentile of magnitudes, every negative component of v_cat fails the mask and is discarded regardless of its magnitude. Consequently Eq. (13) averages only retained positive updates, and parameters that consistently decreased during fine-tuning are reset to the pretrained value. This is the opposite of the 'largest-magnitude updates' trimming described in the text and is not what Ties-Merging [16] does. Because Table 3 shows RHM contributing 60.0 to 61.4 on TerraInc and 86.5 to 87.3 on OfficeHome, an independent reader cannot implement Eqs. (10)-(13) as written and obtain the reported results. Please define tilde v_cat explicitly and either use a magnitude mask m = I(|v_cat| > sigma) with sign-preserving aggregation, or state the exact operation that was actually run.
  2. [4.2, Table 1 and ViT-B/16 paragraph] The main-results narrative contradicts Table 1 for the ERM baseline. The text states that HAM surpasses ERM (74.3%) by 4.7 percentage points on ViT-B/16, but Table 1 reports ERM* with an average of 75.6 (per-dataset 95.6, 82.2, 83.2, 59.7, 57.3). The 74.3 number appears nowhere in the table, and the gap against the tabulated value is 3.4 points, not 4.7. This discrepancy needs to be fixed in either the text or the table, and all reported deltas should be audited for consistency.
  3. [4.1 and 4.2] No standard deviations are reported despite the statement that all results are the average of seeds {41, 42, 43}. The reported advantages over the strongest ViT-B/16 baselines are small: 0.4 points over CLIPood (78.6) and 0.5 points over AlignCLIP (78.5). Without variance estimates, these margins cannot be distinguished from seed noise, which undermines the central state-of-the-art claim. Please report per-dataset standard deviations for the three seeds, and also report the selected hyperparameters (learning rate, total steps, trimming ratio r, sign-loss weight lambda, and Beta parameter beta) for each dataset, since the paper currently gives only search ranges. Code release would also be necessary to make the method reproducible.
  4. [3.4, Eq. (3) and Table 2] The SAE module relies on the assumption that the current model's confidence identifies non-conflicting cross-domain samples. This assumption is not established in the paper. Label noise and multi-object interference can produce confidently wrong predictions, and extreme domain shift can make a model confidently overfit. The assumption is load-bearing because SAE is the largest single contributor in the ablation: TerraIncognita accuracy rises from 54.8% to 61.4% when SAE is added (Table 2). Please provide evidence on the quality of the selected samples, for example precision/recall against a clean validation set, a comparison with oracle filtering, or an ablation using random selection at the same enrichment rate. Without such evidence, the claimed mechanism for the improvement is not supported.
minor comments (6)
  1. [3.4, Eq. (7)] The moving-average recursion uses gamma_k in the second term of the right-hand side, but the index should be gamma_t (the weight of the current step). As written, the recurrence is ill-formed and cannot be implemented as stated.
  2. [4.2, ResNet50 paragraph] The text says CLIPCEIL++ achieves 70.7% on ResNet50, but Table 1 lists CLIPCEIL++* as 69.9%. One of these numbers is wrong, and the comparison text should be aligned with the table.
  3. [5, Conclusions] The claim that HAM is 'the first attempt at model merging in the domain generalization field' is too strong given that the paper's own Related Work section discusses SWAD [12], Model Soup [13], and other merging-based OOD generalization methods. Please soften the claim to a more precise statement about the specific learning-then-merging formulation.
  4. [3.4, Eq. (10)] The vector tilde v_cat is used in Eq. (10) before it is defined. Even if the intended meaning is the elementwise absolute value of v_cat, this should be stated explicitly in the main text rather than left implicit.
  5. [Algorithm 1, line 17] The phrasing 'Extract the concatenated model update vectors of v_cat' is grammatically unclear. The intended meaning is likely 'Extract all model update vectors and concatenate them into v_cat'; please reword.
  6. [Table 2] The row labels 'Single source only' and 'w/ SAE' with checkmarks and crosses are cryptic. Please specify explicitly what each checkmark and cross denotes, for instance by adding a legend or replacing the symbols with 'without SAE' and 'with SAE'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: external benchmarks validate HAM; module heuristics are not self-referential derivations.

full rationale

This is an empirical method paper. The central claims are evaluated on five external DG benchmarks against published baselines, so the results are not defined in terms of the method's own outputs. SAE (Eqs. 2-3) uses the current model's confidence to select enrichment samples; this is a self-training heuristic, but the final accuracy is measured on unseen target domains and is not a tautology. OPA (Eq. 5) regularizes update directions relative to the average of other models; this couples training but does not make the final merged model equal to an input. RHM (Eqs. 8-13) implements magnitude-based trimming and merging, adapted from external work; any sign/threshold typo is a reproducibility concern, not circularity. No load-bearing argument reduces to self-citation: external references such as Ties-Merging, Task Arithmetic, and SWA are independent, parameter-free support. Thus the appropriate circularity score is 0.

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

The method rests on standard DG assumptions and on several unvalidated modeling choices. Hyperparameters lambda, r, learning rate, and total training steps are tuned on validation but not reported per dataset, and the Beta parameter for historical weighting is not reported. No new physical or conceptual entities are introduced.

free parameters (5)
  • lambda (sign loss weight) = range {0.1, 0.5, 1.0}; per-dataset values not reported
    Eq. (15) combines L_CE and lambda * L_sign, and sensitivity analysis is only shown on TerraIncognita.
  • r (redundancy trimming ratio) = range {0.2, 0.4, 0.6, 0.8}; per-dataset values not reported
    Eq. (10)-(11) uses this percentile to mask parameter updates, and sensitivity shows strong performance drops at r=0.9.
  • learning rate = searched in {5e-6, 1e-5}; per-dataset not reported
    AdamW training of each source model depends on this choice.
  • total training steps = searched in {500, 3000, 5000, 10000}; per-dataset not reported
    Determines trajectory length and the historical averaging window in Eq. (7).
  • Beta distribution parameter beta = not reported
    Eq. (7) uses gamma_t = Beta(beta,beta)(...) for trajectory weighting, but no value or search range is given anywhere in the paper.
assumptions (5)
  • domain assumption Source and target domains share one label space
    Used by Eq. (1) and by the enrichment step, which transfers labels across domains. This is the standard DG setup and is not verified empirically.
  • ad hoc to paper Confidence of the current model identifies non-conflicting cross-domain samples
    Eq. (3) uses max(f_i(x)) > tau_i to select samples. The paper provides no evidence that confidence detects label noise or multi-object interference, which it itself lists as conflict types in Fig. 1(a).
  • ad hoc to paper Average update direction is a valid alignment target
    Eq. (5) penalizes opposing signs relative to the detached average update. The paper does not analyze whether this erases useful domain-specific updates.
  • domain assumption Largest-magnitude parameters are the important ones; others are redundant
    RHM relies on this in Eq. (10)-(11), supported only by general pruning references [51,52], not by an experiment on merged CLIP models.
  • domain assumption CLIP text encoder can remain frozen and its zero-shot space is a good basis
    All source encoders are initialized from CLIP's theta_0 and the text encoder is fixed. This is standard for CLIP-DG and is not tested as a variable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Harmonizing and Merging Source Models for CLIP-based Domain Generalization." pith.science (2026). https://pith.science/paper/MFKINZA3

@misc{pith2026250609446,
  author       = {Pith},
  title        = {Pith review of: Harmonizing and Merging Source Models for CLIP-based Domain Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFKINZA3}},
  note         = {Machine review of arXiv:2506.09446}
}
read the original abstract

CLIP-based domain generalization aims to improve model generalization to unseen domains by leveraging the powerful zero-shot classification capabilities of CLIP and multiple source datasets. Existing methods typically train a single model across multiple source domains to capture domain-shared information. However, this paradigm inherently suffers from two types of conflicts: 1) sample conflicts, arising from noisy samples and extreme domain shifts among sources; and 2) optimization conflicts, stemming from competition and trade-offs during multi-source training. Both hinder the generalization and lead to suboptimal solutions. Recent studies have shown that model merging can effectively mitigate the competition of multi-objective optimization and improve generalization performance. Inspired by these findings, we propose Harmonizing and Merging (HAM), a novel source model merging framework for CLIP-based domain generalization. During the training process of the source models, HAM enriches the source samples without conflicting samples, and harmonizes the update directions of all models. Then, a redundancy-aware historical model merging method is introduced to effectively integrate knowledge across all source models. HAM comprehensively consolidates source domain information while enabling mutual enhancement among source models, ultimately yielding a final model with optimal generalization capabilities. Extensive experiments on five widely used benchmark datasets demonstrate the effectiveness of our approach, achieving state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2506.09446 by the authors.

Figure 1
Figure 1. Illustration of dual conflicts in domain generalization: (a) Sample conflicts arise from four distinct types of data [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The proposed framework for CLIP-based domain generalization. It consists of three key modules: 1) Sample Conflict [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis on TerraInc of the weight [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: t-SNE [65] visualization on feature points of HAM and CLIP pre-trained models across different classes and domains. Different colors indicate different classes or domains. ture discrimination across classes. This enhanced cluster￾ing suggests that HAM better aligns vis…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stay Unique, Stay Efficient: Preserving Model Personality in Multi-Task Merging

    cs.LG 2025-12 conditional novelty 5.0 of 10

    DTS approximates task-specific model deltas with truncated SVD plus grouped thresholding and scaling, achieving near-individual multi-task accuracy with ~1% extra storage per task.

Reference graph

Works this paper leans on

77 extracted references · 64 canonical work pages · cited by 1 Pith paper

  1. [16]

    Ties-merging: Resolving interference when merg- ing models

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin A Raffel, and Mohit Bansal. Ties-merging: Resolving interference when merg- ing models. InProc. NeurIPS, pages 7093–7115, 2023

  2. [1]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InProc. ICML, pages 8748–8763. PmLR, 2021

  3. [2]

    Clipceil: Domain generaliza- tion through clip via channel refinement and image-text alignment

    Xi Yu, Shinjae Yoo, and Yuewei Lin. Clipceil: Domain generaliza- tion through clip via channel refinement and image-text alignment. InProc. NeurIPS, pages 4267–4294, 2024. 14 Yuhe Ding et al

  4. [3]

    Alignclip: navigating the misalignments for robust vision-language generalization.Ma- chine Learning, 114(3):1–19, 2025

    Zhongyi Han, Gongxu Luo, Hao Sun, Yaqian Li, Bo Han, Ming- ming Gong, Kun Zhang, and Tongliang Liu. Alignclip: navigating the misalignments for robust vision-language generalization.Ma- chine Learning, 114(3):1–19, 2025

  5. [4]

    Domain generalization via invariant feature representation

    Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In Proc. ICML, pages 10–18. PMLR, 2013

  6. [5]

    Learn to preserve and diversify: Parameter-efficient group with orthog- onal regularization for domain generalization

    Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthog- onal regularization for domain generalization. InProc. ECCV, pages 198–216. Springer, 2024

  7. [6]

    Seeking consistent flat minima for better domain generalization via refining loss landscapes

    Aodi Li, Liansheng Zhuang, Xiao Long, Minghong Yao, and Shafei Wang. Seeking consistent flat minima for better domain generalization via refining loss landscapes. InarXiv preprint arXiv:2412.13573, 2025

  8. [7]

    Leveraging vision-language models for improving domain generalization in image classification

    Sravanti Addepalli, Ashish Ramayee Asokan, Lakshay Sharma, and R Venkatesh Babu. Leveraging vision-language models for improving domain generalization in image classification. InProc. CVPR, pages 23922–23932, 2024

Show all 77 references
  1. [8]

    Soft prompt generation for domain generalization

    Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. InProc. ECCV, pages 434–450. Springer, 2024

  2. [9]

    Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. Model merging in llms, mllms, and beyond: Methods, theories, applications and opportunities. arXiv preprint arXiv:2408.07666, 2024

  3. [10]

    Towards efficient pareto set approximation via mixture of experts based model fusion.arXiv preprint arXiv:2406.09770, 2024

    Anke Tang, Li Shen, Yong Luo, Shiwei Liu, Han Hu, and Bo Du. Towards efficient pareto set approximation via mixture of experts based model fusion.arXiv preprint arXiv:2406.09770, 2024

  4. [11]

    Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

    Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Averaging weights leads to wider optima and better generalization.arXiv preprint arXiv:1803.05407, 2018

  5. [12]

    Swad: Domain generalization by seeking flat minima

    Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Se- unghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. InProc. NeurIPS, pages 22405–22418, 2021

  6. [13]

    Model soups: averaging weights of multiple fine-tuned models im- proves accuracy without increasing inference time

    Mitchell Wortsman, Gabriel Ilharco, Samir Ya Gadre, Re- becca Roelofs, Raphael Gontijo-Lopes, Ari S Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, et al. Model soups: averaging weights of multiple fine-tuned models im- proves accuracy without increasing ...

  7. [14]

    Ensemble learning.The handbook of brain theory and neural networks, 2(1):110–125, 2002

    Thomas G Dietterich et al. Ensemble learning.The handbook of brain theory and neural networks, 2(1):110–125, 2002

  8. [15]

    Editing models with task arithmetic.arXiv preprint arXiv:2212.04089, 2022

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Editing models with task arithmetic.arXiv preprint arXiv:2212.04089, 2022

  9. [17]

    Domain generalization: A survey.IEEE transactions on pat- tern analysis and machine intelligence, 45(4):4396–4415, 2022

    Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey.IEEE transactions on pat- tern analysis and machine intelligence, 45(4):4396–4415, 2022

  10. [18]

    Domain generalization with small data.International Journal of Computer Vision, 132(8):3172–3190, 2024

    Kecheng Chen, Elena Gal, Hong Yan, and Haoliang Li. Domain generalization with small data.International Journal of Computer Vision, 132(8):3172–3190, 2024

  11. [19]

    Xinyue Huo, Lingxi Xie, Hengtong Hu, Wengang Zhou, Houqiang Li, and Qi Tian. Domain-agnostic priors for semantic segmenta- tion under unsupervised domain adaptation and domain general- ization.International Journal of Computer Vision, 132(9):3954– 3976, 2024

  12. [20]

    Csdg-fas: Closed-space domain generalization for face anti-spoofing.International Journal of Computer Vision, 132(11):4866–4879, 2024

    Keyao Wang, Guosheng Zhang, Haixiao Yue, Yanyan Liang, Mouxiao Huang, Gang Zhang, Junyu Han, Errui Ding, and Jing- dong Wang. Csdg-fas: Closed-space domain generalization for face anti-spoofing.International Journal of Computer Vision, 132(11):4866–4879, 2024

  13. [21]

    Do- main generalization with adversarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Do- main generalization with adversarial feature learning. InProc. CVPR, pages 5400–5409, 2018

  14. [22]

    Deep domain generalization via conditional invariant adversarial networks

    Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. InProc. ECCV, pages 624–639, 2018

  15. [23]

    Multi- adversarial discriminative deep domain generalization for face presentation attack detection

    Rui Shao, Xiangyuan Lan, Jiawei Li, and Pong C Yuen. Multi- adversarial discriminative deep domain generalization for face presentation attack detection. InProc. CVPR, pages 10023– 10031, 2019

  16. [24]

    Unified deep supervised domain adaptation and generalization

    Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gian- franco Doretto. Unified deep supervised domain adaptation and generalization. InProc. ICCV, pages 5715–5725, 2017

  17. [25]

    Respecting do- main relations: Hypothesis invariance for domain generalization

    Ziqi Wang, Marco Loog, and Jan Van Gemert. Respecting do- main relations: Hypothesis invariance for domain generalization. In2020 25th International Conference on Pattern Recognition (ICPR), pages 9756–9763. IEEE, 2021

  18. [26]

    Addressing model vulner- ability to distributional shifts over image transformation sets

    Riccardo V olpi and Vittorio Murino. Addressing model vulner- ability to distributional shifts over image transformation sets. In Proc. ICCV, pages 7980–7989, 2019

  19. [27]

    Sebastian Otálora, Manfredo Atzori, Vincent Andrearczyk, Amjad Khan, and Henning Müller. Staining invariant features for improv- ing generalization of deep convolutional neural networks in com- putational pathology.Frontiers in bioengineering and biotechnol- ogy, 7:198, 2019

  20. [28]

    Domain randomiza- tion and pyramid consistency: Simulation-to-real generalization without accessing target domain data

    Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni- Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomiza- tion and pyramid consistency: Simulation-to-real generalization without accessing target domain data. InProc. ICCV, pages 2100– 2110, 2019

  21. [29]

    Semi- supervised domain generalization with stochastic stylematch.In- ternational Journal of Computer Vision, 131(9):2377–2387, 2023

    Kaiyang Zhou, Chen Change Loy, and Ziwei Liu. Semi- supervised domain generalization with stochastic stylematch.In- ternational Journal of Computer Vision, 131(9):2377–2387, 2023

  22. [30]

    Style neophile: Constantly seeking novel styles for domain generaliza- tion

    Juwon Kang, Sohyun Lee, Namyup Kim, and Suha Kwak. Style neophile: Constantly seeking novel styles for domain generaliza- tion. InProc. CVPR, pages 7130–7140, 2022

  23. [31]

    Learning to generalize: Meta-learning for domain generalization

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. InProc. AAAI, volume 32, 2018

  24. [32]

    Shape-aware meta- learning for generalizing prostate mri segmentation to unseen do- mains

    Quande Liu, Qi Dou, and Pheng-Ann Heng. Shape-aware meta- learning for generalizing prostate mri segmentation to unseen do- mains. InProc. MICCAI, pages 475–485. Springer, 2020

  25. [33]

    Episodic training for domain generaliza- tion

    Da Li, Jianshu Zhang, Yongxin Yang, Cong Liu, Yi-Zhe Song, and Timothy M Hospedales. Episodic training for domain generaliza- tion. InProc. ICCV, pages 1446–1455, 2019

  26. [34]

    Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification

    Yuyang Zhao, Zhun Zhong, Fengxiang Yang, Zhiming Luo, Yao- jin Lin, Shaozi Li, and Nicu Sebe. Learning to generalize unseen domains via memory-based multi-source meta-learning for person re-identification. InProc. CVPR, pages 6277–6286, 2021

  27. [35]

    Domain adaptive ensemble learning.IEEE Transactions on Image Pro- cessing, 30:8008–8018, 2021

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Domain adaptive ensemble learning.IEEE Transactions on Image Pro- cessing, 30:8008–8018, 2021

  28. [36]

    Deep domain generalization with structured low-rank constraint.IEEE Transactions on Image Pro- cessing, 27(1):304–313, 2017

    Zhengming Ding and Yun Fu. Deep domain generalization with structured low-rank constraint.IEEE Transactions on Image Pro- cessing, 27(1):304–313, 2017

  29. [37]

    Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets

    Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets. IEEE Transactions on Medical Imaging, 39(12):4237–4248, 2020

  30. [38]

    Self-supervised learning across domains.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(9):5516–5528, 2021

    Silvia Bucci, Antonio D’Innocente, Yujun Liao, Fabio M Carlucci, Barbara Caputo, and Tatiana Tommasi. Self-supervised learning across domains.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 44(9):5516–5528, 2021. Harmonizing and Merging Source Models for CLIP...

  31. [39]

    Efficient domain generalization via common-specific low-rank decomposi- tion

    Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposi- tion. InProc. ICML, pages 7728–7738. PMLR, 2020

  32. [40]

    Improving sample efficiency in model- free reinforcement learning from images

    Denis Yarats, Amy Zhang, Ilya Kostrikov, Brandon Amos, Joelle Pineau, and Rob Fergus. Improving sample efficiency in model- free reinforcement learning from images. InProc. AAAI, vol- ume 35, pages 10674–10681, 2021

  33. [41]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. InProc. ICLR, 2022

  34. [42]

    Evolutionary optimization of model merging recipes.Nature Ma- chine Intelligence, pages 1–10, 2025

    Takuya Akiba, Makoto Shing, Yujin Tang, Qi Sun, and David Ha. Evolutionary optimization of model merging recipes.Nature Ma- chine Intelligence, pages 1–10, 2025

  35. [43]

    Jointly training large autoregressive multimodal models

    Emanuele Aiello, Lili Yu, Yixin Nie, Armen Aghajanyan, and Bar- las Oguz. Jointly training large autoregressive multimodal models. arXiv preprint arXiv:2309.15564, 2023

  36. [44]

    Diffusion soup: Model merging for text-to-image diffusion models

    Benjamin Biggs, Arjun Seshadri, Yang Zou, Achin Jain, Aditya Golatkar, Yusheng Xie, Alessandro Achille, Ashwin Swami- nathan, and Stefano Soatto. Diffusion soup: Model merging for text-to-image diffusion models. InProc. ECCV, pages 257–274. Springer, 2024

  37. [45]

    Adamerging: Adaptive model merging for multi-task learning.arXiv preprint arXiv:2310.02575, 2023

    Enneng Yang, Zhenyi Wang, Li Shen, Shiwei Liu, Guibing Guo, Xingwei Wang, and Dacheng Tao. Adamerging: Adaptive model merging for multi-task learning.arXiv preprint arXiv:2310.02575, 2023

  38. [46]

    Robust fine-tuning of zero-shot models

    Mitchell Wortsman, Gabriel Ilharco, Jong Wook Kim, Mike Li, Simon Kornblith, Rebecca Roelofs, Raphael Gontijo Lopes, Han- naneh Hajishirzi, Ali Farhadi, Hongseok Namkoong, et al. Robust fine-tuning of zero-shot models. InProc. CVPR, pages 7959– 7971, 2022

  39. [47]

    Model ratatouille: Recycling diverse models for out-of-distribution generalization

    Alexandre Ramé, Kartik Ahuja, Jianyu Zhang, Matthieu Cord, Léon Bottou, and David Lopez-Paz. Model ratatouille: Recycling diverse models for out-of-distribution generalization. InProc. ICML, pages 28656–28679. PMLR, 2023

  40. [48]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. InProc. ECCV, pages 456–473, 2018

  41. [49]

    Adamatch: A unified approach to semi-supervised learning and domain adaptation.arXiv preprint arXiv:2106.04732, 2021

    David Berthelot, Rebecca Roelofs, Kihyuk Sohn, Nicholas Car- lini, and Alex Kurakin. Adamatch: A unified approach to semi-supervised learning and domain adaptation.arXiv preprint arXiv:2106.04732, 2021

  42. [50]

    Clipood: Generalizing clip to out-of- distributions

    Yang Shu, Xingzhuo Guo, Jialong Wu, Ximei Wang, Jianmin Wang, and Mingsheng Long. Clipood: Generalizing clip to out-of- distributions. InInternational Conference on Machine Learning, pages 31716–31731. PMLR, 2023

  43. [51]

    Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks.Journal of Machine Learning Research, 22(241):1–124, 2021

    Torsten Hoefler, Dan Alistarh, Tal Ben-Nun, Nikoli Dryden, and Alexandra Peste. Sparsity in deep learning: Pruning and growth for efficient inference and training in neural networks.Journal of Machine Learning Research, 22(241):1–124, 2021

  44. [52]

    Evaluating pruning methods

    Georg Thimm and Emile Fiesler. Evaluating pruning methods. In Proc. ISANN, pages 20–25, 1995

  45. [53]

    Deeper, broader and artier domain generalization

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. InProc. ICCV, pages 5542–5550, 2017

  46. [54]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. InProc. CVPR, pages 1521–1528, 2011

  47. [55]

    Deep hashing network for unsuper- vised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsuper- vised domain adaptation. InProc. CVPR, pages 5018–5027, 2017

  48. [56]

    Moment matching for multi-source do- main adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source do- main adaptation. InProc. ICCV, pages 1406–1415, 2019

  49. [57]

    In search of lost domain generalization.arXiv preprint arXiv:2007.01434, 2020

    Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization.arXiv preprint arXiv:2007.01434, 2020

  50. [58]

    Decoupled weight decay regu- larization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regu- larization. InProc. ICLR, 2019

  51. [59]

    Invariant risk minimization.arXiv preprint arXiv:1907.02893, 2019

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization.arXiv preprint arXiv:1907.02893, 2019

  52. [60]

    Invariant information bottleneck for domain generalization

    Bo Li, Yifei Shen, Yezhen Wang, Wenzhen Zhu, Dongsheng Li, Kurt Keutzer, and Han Zhao. Invariant information bottleneck for domain generalization. InProc. AAAI, pages 7399–7407, 2022

  53. [61]

    Domain generalization by mutual-information regularization with pre-trained models

    Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. InProc. ECCV, pages 440–457, 2022

  54. [62]

    Context-aware robust fine-tuning.International Journal of Computer Vision, 132(5):1685–1700, 2024

    Xiaofeng Mao, Yufeng Chen, Xiaojun Jia, Rong Zhang, Hui Xue, and Zhao Li. Context-aware robust fine-tuning.International Journal of Computer Vision, 132(5):1685–1700, 2024

  55. [63]

    An image is worth 16x16 words: Transformers for image recogni- tion at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa De- hghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recogni- tion at scale. InProc. ICLR, 2021

  56. [64]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProc. CVPR, pages 770–778, 2016

  57. [65]

    Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne.Journal of machine learning research, 9(11), 2008

  58. [66]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. InProc. ICLR, 2018

  59. [67]

    Deep coral: Correlation alignment for deep domain adaptation

    Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. InProc. ECCV Workshops, pages 443–450, 2016

  60. [68]

    Reducing domain gap by reducing style bias

    Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proc. CVPR, pages 8690–8699, 2021

  61. [69]

    Gradient matching for domain generalization

    Yuge Shi, Jeffrey Seely, Philip Torr, Siddharth N, Awni Hannun, Nicolas Usunier, and Gabriel Synnaeve. Gradient matching for domain generalization. InProc. ICLR, 2022

  62. [70]

    Selfreg: Self-supervised contrastive regularization for domain generalization

    Daehee Kim, Youngjun Yoo, Seunghyun Park, Jinkyu Kim, and Jaekoo Lee. Selfreg: Self-supervised contrastive regularization for domain generalization. InProc. ICCV, pages 9619–9628, 2021

  63. [71]

    Exploit- ing domain-specific features to enhance domain generalization

    Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploit- ing domain-specific features to enhance domain generalization. In Proc. NeurIPS, pages 21189–21201, 2021

  64. [72]

    Mixstyle neural networks for domain generalization and adaptation.Inter- national Journal of Computer Vision, 132(3):822–836, 2024

    Kaiyang Zhou, Yongxin Yang, Yu Qiao, and Tao Xiang. Mixstyle neural networks for domain generalization and adaptation.Inter- national Journal of Computer Vision, 132(3):822–836, 2024

  65. [73]

    Under- standing hessian alignment for domain generalization

    Sobhan Hemati, Guojun Zhang, Amir Estiri, and Xi Chen. Under- standing hessian alignment for domain generalization. InProc. ICCV, pages 19004–19014, 2023

  66. [74]

    Balanced direction from multifarious choices: Arithmetic meta-learning for domain generalization

    Xiran Wang, Jian Zhang, Lei Qi, and Yinghuan Shi. Balanced direction from multifarious choices: Arithmetic meta-learning for domain generalization. InProc. CVPR, 2025

  67. [75]

    LFME: A simple framework for learning from mul- tiple experts in domain generalization

    Liang Chen, Yong Zhang, Yibing Song, Zhiqiang Shen, and Lingqiao Liu. LFME: A simple framework for learning from mul- tiple experts in domain generalization. InProc. NeurIPS, 2024

  68. [76]

    Learning intrin- sic invariance within intra-class for domain generalization.IEEE Transactions on Multimedia, pages 1–14, 2025

    Chaoyang Zhou, Zengmao Wang, and Bo Du. Learning intrin- sic invariance within intra-class for domain generalization.IEEE Transactions on Multimedia, pages 1–14, 2025

  69. [77]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei- Fei. Imagenet: A large-scale hierarchical image database. InProc. CVPR, pages 248–255, 2009

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.