Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics

Seongjun Lee; Changhee Lee

cs.AIOct 2, 2026

Safeguarding Mutual Correction in Source-Free Domain Adaptation via Cut Statistics

Authors: Seongjun Lee, Changhee Lee

Organizations: Department of Artificial Intelligence Korea University Seoul, Republic of Korea

Abstract

Source-Free Domain Adaptation (SFDA) aims to adapt a source-pretrained model to an unlabeled target domain without access to the original source domain. While early single-model approaches rely on self-refinement, they are inherently susceptible to confirmation bias and struggle to correct their own systematic errors. To overcome this limitation, recent methods introduce Vision-Language (ViL) models as external knowledge sources. However, these approaches operate in a largely unidirectional paradigm, using the ViL model primarily to supervise the source-pretrained model. This overlooks a key structural property: the two models exhibit distinct failure modes -- where one produces an incorrect prediction, the other may produce a correct one, creating a natural opportunity for mutual correction within the target domain. Yet, without ground-truth labels, identifying which model is correct on any given sample is non-trivial, and naively exchanging predictions risks propagating errors across models. To address this challenge, we propose SafeCut, a novel approach that leverages the cut statistic as a label-free measure of prediction reliability to gate cross-model supervision. Our approach dynamically controls both the direction and strength of supervision based on relative reliability, selectively amplifying true corrections while suppressing miscorrections on a per-sample basis. We further provide theoretical justification showing that this reliability-gated mechanism guarantees a net-positive correction signal. Extensive experiments across diverse SFDA benchmarks demonstrate that SafeCut achieves state-of-the-art performance, highlighting the effectiveness of safeguarding mutual correction in SFDA via cut statistics.

Figures & tables

Figure 1: Reliability-based Mutual Correction. The model with lower cut corrects the other (Sample 1: $\theta_{v}\!\to\!\theta_{t}$ , Sample 2: $\theta_{t}\!\to\!\theta_{v}$ ).

Panel of Figure 2 — Figure 2: Cut statistic as a reliability proxy. For each adaptation task, the left panel shows that prediction accuracy decreases (approximately) monotonically with the cut statistic $z$ (computed via ( 9 )) for both models. The right panel shows the sample density in the $(z_{t},z_{v})$ plane for $\mathcal{S}_{c\bar{c}}$ (blue) and $\mathcal{S}_{\bar{c}c}$ (orange), which are concentrated in $\Delta(x)\geq\gamma_{0}$ and $\Delta(x)\leq-\xi_{0}$ , respectively.

Method	Venue	A $\to$ D	A $\to$ W	D $\to$ A	D $\to$ W	W $\to$ A	W $\to$ D	Avg.
Source	–	79.1	76.5	59.9	95.5	61.4	98.8	78.5
SHOT	ICML20	94.4	88.4	73.2	96.7	74.9	99.8	87.9 /88.6
NRC	NIPS21	92.4	94.1	75.6	97.2	74.6	98.6	88.7 /89.4
AaD	NIPS22	87.6	86.5	58.9	90.6	58.7	90.6	78.8 /89.9
AdaCon	CVPR22	87.8	86.3	71.3	96.7	67.2	97.6	84.5 /–
CoWA	ICML22	88.0	90.9	74.6	98.1	75.8	99.4	87.8 /90.3

Table 1: Closed-set SFDA results (Acc, %) on Office-31 .

Method	Venue	Office-Home													VisDA
		Ar $\to$ Cl	Ar $\to$ Pr	Ar $\to$ Rw	Cl $\to$ Ar	Cl $\to$ Pr	Cl $\to$ Rw	Pr $\to$ Ar	Pr $\to$ Cl	Pr $\to$ Rw	Rw $\to$ Ar	Rw $\to$ Cl	Rw $\to$ Pr	Avg.	Sy $\to$ Re
Source	–	43.6	67.0	73.9	49.9	60.2	62.6	51.7	40.9	72.6	64.1	46.3	78.1	59.2	49.1
SHOT	ICML20	55.2	78.5	81.3	68.9	79.0	78.8	68.2	54.3	81.7	73.5	59.1	83.6	71.8 /71.8	82.3 /82.9
NRC	NIPS21	54.7	77.1	79.4	64.7	76.2	75.3	63.3	53.6	78.8	70.2	56.5	82.5	69.4 /72.2	83.4 /85.9
AaD	NIPS22	57.7	78.5	80.8	68.0	79.6	78.9	66.5	56.6	81.6	70.9	57.3	83.7	71.7 /72.7	78.8 /88.0
AdaCon	CVPR22	49.5	76.1	77.0	63.5	73.2	74.4	60.7	47.1	77.9	67.8	49.7	80.3	66.4 /–	67.2 /78.7

Table 2: Comparison on the Office-Home and VisDA datasets (Acc, %).

Target: Ar						Target: Cl						Target: Pr						Target: Rw
		$\theta_{1}$		$\theta_{2}$				$\theta_{1}$		$\theta_{2}$				$\theta_{1}$		$\theta_{2}$				$\theta_{1}$		$\theta_{2}$
$\theta_{1}$	$\theta_{2}$	Bef.	Aft.	Bef.	Aft.	$\theta_{1}$	$\theta_{2}$	Bef.	Aft.	Bef.	Aft.	$\theta_{1}$	$\theta_{2}$	Bef.	Aft.	Bef.	Aft.	$\theta_{1}$	$\theta_{2}$	Bef.	Aft.	Bef.	Aft.
Cl	Pr	49.9	72.2	51.7	72.2	Ar	Pr	43.6	59.8	40.9	59.9	Ar	Cl	67.0	81.3	60.2	81.2	Ar	Cl	73.9	83.1	62.6	83.1
Cl	Rw	49.9	75.8	64.2	75.8	Ar	Rw	43.6	61.7	46.3	61.6	Cl	Rw	60.2	86.4	78.1	86.4	Ar	Pr	73.9	84.3	72.6	84.2
Pr	Rw	51.7	74.4	64.2	74.4	Pr	Rw	40.9	60.4	46.3	60.4	Ar	Rw	67.0	83.0	78.1	83.0	Cl	Pr	62.6	83.9	72.6	83.9

Table 4: Mutual correction between two ResNet-50 models from different source domains (Office-Home, Acc, %). Bef./Aft.: accuracy before/after adaptation.

Figure 3: Correctness pattern (top) and average cut trajectories (bottom) on Office-Home Cl $\to$ Pr. Samples are grouped by their initial state and tracked throughout adaptation.

Figure 4: Peer complementarity. Prediction agreement vs. adaptation gain in the dual-ResNet setting of Table 4 .

Method	Target	CLIP
Source	64.14	50.01
ProDe	76.27	50.00
$\mathcal{L}_{\mathrm{c}}\!=\!1,\mathcal{L}_{\mathrm{a}}\!=\!0$	77.29	77.20
$\mathcal{L}_{\mathrm{c}}\!=\!1,\mathcal{L}_{\mathrm{a}}\!=\!1$	87.96	87.78
SafeCut	94.96	94.20

Table 6: SFDA on Camelyon17 (Acc, %).

Method	Target	CLIP
Source	39.0	25.4
ProDe	35.2	35.3
SafeCut	46.0	45.9

Table 7: EyePACS

\to

APTOS2019 (Acc, %).

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Venue	A $\to$ D	A $\to$ W	D $\to$ A	D $\to$ W	W $\to$ A	W $\to$ D	Avg.
Source	–	79.1	76.5	59.9	95.5	61.4	98.8	78.5
SHOT	ICML20	94.0	90.1	74.7	98.4	74.3	99.9	88.6
NRC	NIPS21	96.0	90.8	75.3	99.0	75.0	100.0	89.4
AaD	NIPS22	96.4	92.1	75.0	99.1	76.5	100.0	89.9
CoWA	ICML22	94.4	95.2	76.2	98.5	77.6	99.8	90.3
ELR	ICLR23	93.8	93.3	76.2	98.0	76.9	100.0	89.6

Appendix

Table 8: Comparison against paper-reported best results on Office-31 (Acc, %).

Method	Venue	Ar $\to$ Cl	Ar $\to$ Pr	Ar $\to$ Rw	Cl $\to$ Ar	Cl $\to$ Pr	Cl $\to$ Rw	Pr $\to$ Ar	Pr $\to$ Cl	Pr $\to$ Rw	Rw $\to$ Ar	Rw $\to$ Cl	Rw $\to$ Pr	Avg.
Source	–	43.6	67.0	73.9	49.9	60.2	62.6	51.7	40.9	72.6	64.1	46.3	78.1	59.2
SHOT	ICML20	57.1	78.1	81.5	68.0	78.2	78.1	67.4	54.9	82.2	73.3	58.8	84.3	71.8
NRC	NIPS21	57.7	80.3	82.0	68.1	79.8	78.6	65.3	56.4	83.0	71.0	58.6	85.6	72.2
AaD	NIPS22	59.3	79.3	82.1	68.9	79.8	79.5	67.2	57.4	83.1	72.1	58.5	85.4	72.7
CoWA	ICML22	56.9	78.4	81.0	69.1	80.0	79.9	67.7	57.2	82.4	72.8	60.5	84.5	72.5
ELR	ICLR23	58.4	78.7	81.5	69.2	79.5	79.3	66.3	58.0	82.6	73.4	59.8	85.1	72.6

Appendix

Table 9: Comparison against paper-reported best results on Office-Home (Acc, %).

Method	Venue	C $\to$ P	C $\to$ R	C $\to$ S	P $\to$ C	P $\to$ R	P $\to$ S	R $\to$ C	R $\to$ P	R $\to$ S	S $\to$ C	S $\to$ P	S $\to$ R	Avg.
Source	–	44.7	59.9	47.5	53.3	75.3	46.2	55.1	62.6	46.4	55.2	50.8	59.6	54.7
ELR	ICLR23	77.2	91.2	68.5	79.8	92.7	67.5	75.6	82.2	65.7	85.1	77.6	89.3	79.4
DIFO-V	CVPR24	76.6	87.2	74.9	80.0	87.4	75.6	80.8	77.3	75.5	80.5	76.7	87.3	80.0
ProDe-V	ICLR25	83.2	92.4	79.0	85.0	92.3	79.3	85.5	83.1	79.1	85.5	83.4	92.4	85.0
Ours	NIPS26	82.8	92.3	81.8	86.2	92.2	81.0	86.9	82.9	81.7	86.6	83.0	92.3	85.8

Appendix

Table 10: Comparison against paper-reported best results on DomainNet-126 (Acc, %).

Method	Venue	plane	bcycl	bus	car	horse	knife	mcycl	person	plant	sktbrd	train	truck	Avg.
Source	–	60.7	21.7	50.9	68.5	71.8	5.3	86.3	20.2	67.1	43.2	83.3	10.6	49.1
SHOT	ICML20	94.3	88.5	80.1	57.3	93.1	94.9	80.7	80.3	91.5	89.1	86.3	58.2	82.9
NRC	NIPS21	96.8	91.3	82.4	62.4	96.2	95.9	86.1	80.6	94.8	94.1	90.4	59.7	85.9
AaD	NIPS22	97.4	90.5	80.8	76.2	97.3	96.1	89.8	82.9	95.5	93.0	92.0	64.7	88.0
AdaCon	CVPR22	97.0	84.7	84.0	77.3	96.7	93.8	91.9	84.8	94.3	93.1	94.1	49.7	86.8
CoWA	ICML22	96.2	89.7	83.9	73.8	96.4	97.4	89.3	86.8	94.6	92.1	88.7	53.8	86.9

Appendix

Table 11: Comparison against paper-reported best results on VisDA-C (per-class Acc, %).

Method	Venue	plane	bcycl	bus	car	horse	knife	mcycl	person	plant	sktbrd	train	truck	Per-class
Source	–	60.7	21.7	50.9	68.5	71.8	5.3	86.3	20.2	67.1	43.2	83.3	10.6	49.1
SHOT	ICML20	94.9	86.9	79.7	54.8	93.6	94.1	78.5	81.4	90.6	90.0	84.5	58.3	82.3
NRC	NIPS21	97.2	91.5	83.5	39.3	96.4	96.2	82.8	79.7	93.8	93.3	91.2	55.8	83.4
AaD	NIPS22	97.5	91.8	70.6	45.5	92.3	90.0	65.6	65.6	95.0	88.4	87.5	56.0	78.8
AdaCon	CVPR22	88.4	69.1	78.5	55.2	85.4	91.6	79.7	55.8	80.3	48.4	64.7	8.9	67.2
CoWA	ICML22	95.5	84.2	83.0	68.9	94.3	96.9	88.8	83.1	91.3	89.0	85.5	51.5	84.3

Appendix

Table 12: Full per-class results on VisDA-C (Synthesis

\rightarrow

Real, acc, %).

	SafeCut (Ours)				ProDe
Dataset	Avg. KNN	Avg. Train	Avg. Infer.	Total	Total
Office-31	0.05 s	8.92 s	6.57 s	0.78 h	1.11 h
Office-Home	$<$ 0.01 s	17.85 s	15.83 s	5.62 h	5.87 h
DomainNet-126	0.25 s	72.21 s	44.86 s	19.55 h	29.58 h
VisDA	0.40 s	160.80 s	80.70 s	2.02 h	3.49 h

Appendix

Table 13: Computational cost analysis between SafeCut (ours) and ProDe.

Figure 5: Accuracy vs. cut statistic $z$ for all 12 Office-Home tasks (CLIP peer).

Figure 6: Accuracy vs. cut statistic $z$ for all 12 DomainNet-126 tasks (CLIP peer).

Figure 7: Accuracy vs. cut statistic $z$ for all 12 Office-Home source-pair combinations (dual-ResNet peer).

Figure 8: Joint cut density of $\mathcal{S}_{c\bar{c}}$ and $\mathcal{S}_{\bar{c}c}$ for all 12 Office-Home tasks (CLIP peer).

Figure 9: Joint cut density of $\mathcal{S}_{c\bar{c}}$ and $\mathcal{S}_{\bar{c}c}$ for all 12 DomainNet-126 tasks (CLIP peer).

Figure 10: Joint cut density of $\mathcal{S}_{c\bar{c}}$ and $\mathcal{S}_{\bar{c}c}$ for all 12 Office-Home source-pair combinations (dual-ResNet peer).

Figure 11: Temporal stability of the cut–accuracy monotonicity. Mean accuracy versus cut statistic $z$ for the target model (solid) and CLIP peer (dashed) at four checkpoints (Epoch 1, $T/3$ , $2T/3$ , $T$ ) on three Office-Home tasks. The monotonically decreasing trend and the relative ordering between the two curves are preserved throughout adaptation, while the target curve progressively converges toward the peer.

Figure 12: Correctness pattern transition (Epoch 0 vs. 50) for all 12 Office-Home tasks (CLIP peer).

Figure 13: Correctness pattern transition (Epoch 0 vs. 50) for all 12 DomainNet-126 tasks (CLIP peer).

Figure 14: Correctness pattern transition (Epoch 0 vs. 50) for all 12 Office-Home source-pair combinations (dual-ResNet peer, $\theta_{1}$ ).

Target	Source 1	Source 2	Agreement (%)	Gain (%p)
Cl	Ar	Pr	41.1	17.6
Ar	Cl	Pr	43.3	21.3
Cl	Ar	Rw	46.1	16.7
Cl	Pr	Rw	46.4	16.8
Ar	Cl	Rw	48.4	18.7
Ar	Pr	Rw	53.6	16.5

Appendix

Table 14: Effect of peer complementarity in the dual-ResNet setting of Table 4 . Agreement denotes the prediction agreement between the two models before adaptation, and Gain denotes their average accuracy improvement after SafeCut adaptation.

Optimization
	O31	OH	DN	VisDA	Cam17
Epochs	30	50	50	30	20
Batch size	64	192	192	192	128
LR	5e-3	1e-2	1e-2	5e-3	5e-3
Prompt LR	5e-3	5e-4	5e-4	1e-3	5e-4
Cut statistic

Appendix

Table 15: Implementation details. Red : identical across all applicable benchmarks; underline : identical in most benchmarks.

Hyperparameter	Search Space
Epochs $E$	$\{1,\ 2,\ 3,\ 4,\ 5,\ 10\}$
$\lambda_{\mathrm{div}}$	$\{0.05,\ 0.1\}$
$\lambda_{\mathrm{iic}}$	$\{0.3,\ 0.5,\ 0.6,\ 0.7,\ 1.0\}$
LR	$\{10^{-3},\ 3{\times}10^{-3},\ 5{\times}10^{-3},\ 10^{-2},$ $\ 0.05,\ 0.06,\ \ldots,\,\ 0.25,\ 0.30,\ 0.35,\ 0.40\}$
Prompt LR	$\{10^{-5},\ 5{\times}10^{-5},\ 10^{-4},\ 5{\times}10^{-4},$ $\ 10^{-3},\ 5{\times}10^{-3},\ 0.01,\ 0.02,\ \ldots,\ 0.10\}$
Prompt init	a_photo_of_a (fixed)

Appendix

Table 16: Hyperparameter search space for ProDe on Camelyon17.

Source-Free Domain Adaptation

Source-Free Domain Adaptation transfers models to new data distributions without access to the original training examples.

See all 69 papers

May 4, 2026

Rethinking the Need for Source Models: Source-Free Domain Adaptation from Scratch Guided by a Vision-Language Model

Source-Free Domain Adaptation (SFDA) adapts source models to target domains without accessing source data, addressing privacy and transmission issues. However, existing methods still initialize from a source pre-trained model and thus are not truly source-free. Recent works have introduced Vision-Language (ViL) models to guide the adaptation process, in these methods, we observe that for the same target domain, different source models yield minimal variation in final results, indicating the source model itself has limited impact. Motivated by this, we propose ViL-Only Domain Adaptation (VODA) , a stricter setting that eliminates all dependencies on source domain, relying solely on a randomly initialized model, a ViL model, and unlabeled target data. We analyze the adaptation dynamics of VODA and introduce Two-Stage Denoised-Region Distillation (TS-DRD) , a two-stage framework that first warms up the model with ViL guidance, then seek a Denoised-Region inherent in both the ViL and adapting model, yielding cleaner supervision for distillation. Experiments on Office-Home, VisDA, and DomainNet-126 show that under VODA, TS-DRD achieves competitive or superior performance to existing SFDA methods that still use source models, demonstrating its effectiveness and the potential of the VODA setting.

Zhou Bingtao, Xiang Mian, Ning Qian

Sichuan University, No.24 South Section 1, Yihuan Road, Chengdu, 610065, Sichuan, China · Hubei Minzu University, 39 Xueyuan Road, Enshi, 445000, Hubei, China

Source-Free Domain Adaptation VLM Adaptation

Ask PDF

Aug 5, 2026

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

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.

Bo Li, Junjie Peng, Xiaohua Xie +1

School of Computer Science and Engineering, Sun Yat-sen University

Source-Free Domain Adaptation VLM Adaptation

Ask PDF

Apr 20, 2026

Source-Free Domain Adaptation with Vision-Language Prior

Source-Free Domain Adaptation (SFDA) seeks to adapt a source model, which is pre-trained on a supervised source domain, for a target domain, with only access to unlabeled target training data. Relying on pseudo labeling and/or auxiliary supervision, conventional methods are inevitably error-prone. To mitigate this limitation, in this work we for the first time explore the potentials of off-the-shelf vision-language (ViL) multimodal models (e.g., CLIP) with rich whilst heterogeneous knowledge. We find that directly applying the ViL model to the target domain in a zero-shot fashion is unsatisfactory, as it is not specialized for this particular task but largely generic. To make it task-specific, we propose a novel DIFO++ approach. Specifically, DIFO++ alternates between two steps during adaptation: (i) Customizing the ViL model by maximizing the mutual information with the target model in a prompt learning manner, (ii) Distilling the knowledge of this customized ViL model to the target model, centering on gap region reduction. During progressive knowledge adaptation, we first identify and focus on the gap region, where enclosed features are entangled and class-ambiguous, as it often captures richer task-specific semantics. Reliable pseudo-labels are then generated by fusing predictions from the target and ViL models, supported by a memory mechanism. Finally, gap region reduction is guided by category attention and predictive consistency for semantic alignment, complemented by referenced entropy minimization to suppress uncertainty. Extensive experiments show that DIFO++ significantly outperforms the state-of-the-art alternatives. Our code and data are available at https://github.com/tntek/DIFO-Plus.

Song Tang, Yunxiang Bai, Wenxin Su +3

Institute of Machine Intelligence (IMI), University of Shanghai for Science and Technology, Shanghai, China · Technical Aspects of Multimodal Systems (TAMS) Group, Department of Informatics, Universität Hamburg, Hamburg, Germany · European Molecular Biology Laboratory, Heidelberg, Germany +2

Source-Free Domain Adaptation Cross-Modal Knowledge Distillation VLM Adaptation

Ask PDF

Abstract

Source-Free Domain Adaptation

Explore similar work

Rethinking the Need for Source Models: Source-Free Domain Adaptation from Scratch Guided by a Vision-Language Model

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

Source-Free Domain Adaptation with Vision-Language Prior