Active Source-Free Domain Adaptation

• ASFDA is a domain adaptation technique that transfers a source-trained model to a target domain using actively selected target labels while keeping source data inaccessible.
• It employs advanced query strategies—such as minimum happy points, uncertainty sampling, and iterative adaptation—to effectively bridge the domain gap.
• Empirical results across tasks like classification, segmentation, and medical imaging show significant performance gains with minimal annotations.

Active Source-Free Domain Adaptation (ASFDA) is a domain adaptation setting in which a source-trained model is transferred to a target domain without access to source data, while a small target-domain annotation budget is used through active selection of the most valuable target samples. In the original formulation, a source-trained model $h_s(x)=g_s(f_s(x))$ is given together with an unlabeled target pool $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$ and a query budget $m\ll N_t$; the aim is to select $\mathcal D_t^L\subset\mathcal D_t$ with $|\mathcal D_t^L|=m$, keep the remainder unlabeled, and train a target model $h_t(x)=g_t(f_t(x))$ that performs well on the target domain under the source-free constraint (Wang et al., 2022). Subsequent formulations express the same objective as minimizing target risk with only source-pretrained parameters, unlabeled target data, and actively queried labels, sometimes in multiple rounds and sometimes in one shot (Lyu et al., 2024, Chen et al., 29 Sep 2025). The topic now spans closed-set classification, segmentation, detection, multimodal supervision, and open-set medical image segmentation (Kothandaraman et al., 2022, Luo et al., 2024, Yang et al., 7 Jun 2026).

1. Definition and formal setting

ASFDA extends standard Source-Free Domain Adaptation (SFDA) by allowing a small number of actively chosen target labels. MHPL states this directly: ASFDA is a more practical scenario than SFDA because it permits actively selecting a few target data to be labeled by experts, while preserving the condition that source data cannot be accessed during adaptation (Wang et al., 2022). LFTL describes the same setting as Source data-Free Active Domain Adaptation (SFADA): a source-pretrained model, an unlabeled target pool $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$, a total annotation budget $B$, and a target model $M_t$ optimized to minimize

$R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$

through iterative query-and-adaptation rounds (Lyu et al., 2024).

A closely related formulation is given by DAM. There, a source-pretrained classifier $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$0 is trained over $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$1 classes, an unlabeled target set $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$2 is available, and a labeling budget $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$3 together with an active query strategy $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$4 selects a subset $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$5. The goal is to adapt a target model $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$6, initialized as $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$7, to minimize

$\mathcal D_t=\{x_i\}_{i=1}^{N_t}$8

using only $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$9, the budgeted labels in $m\ll N_t$0, and the source-model parameters (Chen et al., 29 Sep 2025).

Open-set variants generalize the label-space assumption. ASFOSDA defines source-free open-set domain adaptation for medical image segmentation by allowing unknown target classes $m\ll N_t$1, so that $m\ll N_t$2. In that setting, the target task must handle both known source classes and private target-only structures, and negative transfer can occur if unknown classes are force-aligned to source features (Yang et al., 7 Jun 2026). This establishes that ASFDA is not confined to the closed-set assumption used in early classification benchmarks.

2. Emergence of the paradigm and representative frameworks

MHPL is the first work that explicitly introduces the term “active source free domain adaptation (ASFDA)” and defines the core problem as active target selection under a source-free constraint (Wang et al., 2022). Almost concurrently, SALAD formulates “Source-free Active Label-Agnostic Domain Adaptation” as a setting in which a pre-trained source-domain network must be adapted to a target domain with a small budget for annotation and a shift in the label space, again without access to source data (Kothandaraman et al., 2022). These two 2022 works established two durable themes: active query design and adaptation under missing source data.

Later work broadened both methodology and scope. LFTL formulates SFADA as an iterative query-and-adaptation process and emphasizes “continuous performance improvements through the iterative query-and-adaptation process” (Lyu et al., 2024). ProULearn focuses on “sample exploration” through homogeneity propensity estimation, correlation-based uncertainty, and compact class distributions during adaptation (Pan et al., 23 Jan 2025). DAM introduces a multimodal foundation model into SFADA and explicitly fuses human supervision with Vision-and-Language supervision through dual-focused active supervision and bidirectional distillation (Chen et al., 29 Sep 2025). UGTST and ASFOSDA extend the setting to medical image segmentation, with ASFOSDA explicitly addressing open-set adaptation without source data (Luo et al., 2024, Yang et al., 7 Jun 2026).

Framework	Core mechanisms	Scope
MHPL	minimum happy points, neighbor ambient uncertainty, neighbor diversity relaxation, one-shot querying, neighbor focal loss	ASFDA classification (Wang et al., 2022)
SALAD	GATN and HAL	classification, segmentation and detection (Kothandaraman et al., 2022)
LFTL	Contrastive Active Sampling and Visual Persistence-guided Adaptation	iterative SFADA (Lyu et al., 2024)
ProULearn	homogeneity propensity estimation, correlation index, central correlation loss	SFADA classification (Pan et al., 23 Jan 2025)
DAM	dual-focused active supervision, bidirectional distillation	multimodal SFADA (Chen et al., 29 Sep 2025)
UGTST / ASFOSDA	tiered self-training / Target-refined Self-training, uncertainty and discrepancy-based selection	medical segmentation, including open-set (Luo et al., 2024, Yang et al., 7 Jun 2026)

This progression suggests that ASFDA evolved from a narrowly defined classification scenario into a broader family of source-free active adaptation problems with differing assumptions about label space, supervision sources, and query schedule.

3. Query strategies and notions of informativeness

The earliest explicit ASFDA acquisition principle is MHPL’s notion of “minimum happy” points. A target point is an MH point if it is simultaneously neighbor-chaotic, individual-different, and target-like (Wang et al., 2022). Neighbor-chaotic behavior is quantified by neighbor ambient uncertainty: $m\ll N_t$3 where $m\ll N_t$4 is neighbor entropy over clustering-based pseudo-labels and $m\ll N_t$5 is average feature-space similarity to the $m\ll N_t$6 nearest neighbors. Diversity is imposed by Neighbor Diversity Relaxation, which greedily selects candidates while rejecting those whose nearest neighbor has already been selected. MHPL also adopts one-shot querying: all $m\ll N_t$7 labels are selected before any target fine-tuning (Wang et al., 2022).

LFTL replaces one-shot selection with an iterative criterion based on both uncertainty and persistence. Its Contrastive Active Sampling computes softmax probabilities at each round and forms a contrastively decoded log-probability. The Best-vs-Second Best margin

$m\ll N_t$8

defines informativeness, with smaller values indicating more confusion, while a class-level intransferability term $m\ll N_t$9 down-weights classes that are already easy. The final score is

$\mathcal D_t^L\subset\mathcal D_t$0

and samples are queried in ascending order of $\mathcal D_t^L\subset\mathcal D_t$1 (Lyu et al., 2024). LFTL explicitly defines “informative” samples as those on which the current model remains most uncertain and “persistently challenging” samples as those that remain uncertain even after previous rounds of adaptation (Lyu et al., 2024).

ProULearn develops a different acquisition logic aimed at avoiding noisy outliers. Its homogeneity propensity estimation adapts the Isolation Forest idea: dense regions require more random splits and therefore longer average path-lengths. A correlation index is then computed between feature embeddings, local neighbor probabilities are averaged, local entropy $\mathcal D_t^L\subset\mathcal D_t$2 is calculated, and the final selection score is

$\mathcal D_t^L\subset\mathcal D_t$3

Selection ranks samples by $\mathcal D_t^L\subset\mathcal D_t$4 and excludes already-selected samples’ $\mathcal D_t^L\subset\mathcal D_t$5 neighbors to promote diversity (Pan et al., 23 Jan 2025).

Medical segmentation work emphasizes uncertainty decomposition and spatial redundancy. UGTST estimates global uncertainty per slice through test-time perturbations, pixel-wise entropy, histogram discretization, primary local peak filtering, and a diversity-aware redundancy filter based on K-means$\mathcal D_t^L\subset\mathcal D_t$6 over encoder features (Luo et al., 2024). ASFOSDA separates Class-aware Decomposed Uncertainty into aleatoric uncertainty and epistemic uncertainty, both estimated via test-time augmentation, then applies Class-agnostic Prototype Discrepancy, combining Cross-Domain Discrepancy and Self-Domain Discrepancy, to select the final annotation set (Yang et al., 7 Jun 2026).

A common misconception is that ASFDA acquisition is equivalent to plain uncertainty sampling. The literature does not support that reduction. MHPL combines uncertainty with neighbor affinity and diversity; LFTL adds persistence and class-level intransferability; ProULearn adds homogeneity propensity and correlation structure; UGTST adds redundancy filtering; and ASFOSDA adds prototype discrepancy in an open-set setting (Wang et al., 2022, Lyu et al., 2024, Pan et al., 23 Jan 2025, Luo et al., 2024, Yang et al., 7 Jun 2026).

4. Adaptation objectives after querying

Acquisition is only one half of ASFDA; the other half is how the queried labels are exploited together with the unlabeled target set. SALAD frames the problem as a joint optimization of sample selection and knowledge transfer. Its Guided Attention Transfer Network (GATN) enables feature distillation from the pre-trained network to the target network, and its active learning function HAL mines target samples using transfer-ability and uncertainty criteria (Kothandaraman et al., 2022). SALAD’s abstract emphasizes three consequences of this design: task-agnostic operation across classification, segmentation, and detection; handling of shifts in output label space from the pre-trained source network to the target domain; and adaptation without access to source data (Kothandaraman et al., 2022).

MHPL uses a neighbor focal loss to emphasize queried MH points during training. Labeled queried samples receive a cross-entropy term weighted by $\mathcal D_t^L\subset\mathcal D_t$7, pseudo-labeled unlabeled samples receive a lower weight $\mathcal D_t^L\subset\mathcal D_t$8, and the final objective adds entropy minimization and class-balance regularization: $\mathcal D_t^L\subset\mathcal D_t$9 The intended effect is to focus more strongly on the target samples whose labels were actively purchased (Wang et al., 2022).

LFTL’s Visual Persistence-guided Adaptation operates after each query round. Newly labeled samples become anchors, their feature embeddings are maintained in a momentum-updated memory, unlabeled samples are softly assigned to anchors, and adaptation uses three losses: supervised cross-entropy on the labeled set, an anchor-based clustering loss $|\mathcal D_t^L|=m$0, and unlabeled-data entropy minimization $|\mathcal D_t^L|=m$1. The joint objective is

$|\mathcal D_t^L|=m$2

The paper stresses that VPA learns from features of actively selected anchors obtained from previous intermediate models, facilitating feature distribution alignment and active sample exploitation (Lyu et al., 2024).

ProULearn adapts through pseudo-label refinement and compact-class learning. It initializes class centroids from current target predictions, assigns pseudo-labels using refined similarity scores $|\mathcal D_t^L|=m$3, and optimizes a full objective

$|\mathcal D_t^L|=m$4

where $|\mathcal D_t^L|=m$5 is the central correlation loss that draws features toward their assigned class centroids (Pan et al., 23 Jan 2025). The paper argues that this explicitly refines pseudo-labels and compacts class clusters around prototypes.

DAM broadens the supervision model itself. A frozen Vision-and-LLM $|\mathcal D_t^L|=m$6, such as CLIP, supplies multimodal pseudo-label distributions through image and text encoders and learnable category prompts. On the labeled target subset $|\mathcal D_t^L|=m$7, DAM forms dual supervision through human supervision for the target model $|\mathcal D_t^L|=m$8 and prompt-tuning supervision for $|\mathcal D_t^L|=m$9. Fusion then occurs through bidirectional distillation: $h_t(x)=g_t(f_t(x))$0 on top-confidence samples and queried labels, and $h_t(x)=g_t(f_t(x))$1 on the target set with higher weights on queried labels. The target model objective combines distillation, active-label cross-entropy, entropy minimization, and a diversity regularizer, while the ViL model optimizes distillation and prompt supervision; at inference, only $h_t(x)=g_t(f_t(x))$2 is retained, so there are no additional FLOPs from $h_t(x)=g_t(f_t(x))$3 (Chen et al., 29 Sep 2025).

In medical segmentation, self-training is the dominant post-query adaptation strategy. UGTST uses a two-stage tiered self-training schedule: first, the source model is adapted on manually labeled slices together with source-model pseudo-labels for the remaining target data; then new pseudo-labels are generated with the augmentation ensemble, and the model is retrained on mixed true and pseudo labels (Luo et al., 2024). ASFOSDA similarly performs supervised fine-tuning on queried labels, generates pseudo-labels for the remaining unlabeled set, and then minimizes a joint Dice-plus-cross-entropy objective over both labeled and pseudo-labeled subsets (Yang et al., 7 Jun 2026).

5. Task coverage, label-space assumptions, and domain-specific variants

ASFDA was first studied on standard visual classification benchmarks, including Office-Home, Office-31, and VisDA-2017 or VisDA-C (Wang et al., 2022, Lyu et al., 2024, Pan et al., 23 Jan 2025). Even within that initial regime, methodological assumptions diverged: MHPL adopts one-shot querying before fine-tuning, whereas LFTL explicitly iterates querying and adaptation across rounds (Wang et al., 2022, Lyu et al., 2024).

SALAD is notable for broadening the task definition beyond classification. Its experiments cover digits classification on MNIST, SVHN, and VISDA, synthetic-to-real image segmentation from GTA5 to CityScapes, and document layout detection from PubLayNet to DSSE (Kothandaraman et al., 2022). It is also explicitly label-agnostic in the sense that it can handle shifts in output label space from the pre-trained source network to the target domain (Kothandaraman et al., 2022). This suggests that active source-free adaptation is not intrinsically tied to a fixed shared label set.

Medical image segmentation introduced a different set of constraints. UGTST studies cross-center prostate MRI segmentation with a 2D U-Net source model, target domains from BIDMC and HK, and a one-round annotation budget of $h_t(x)=g_t(f_t(x))$4 of slices (Luo et al., 2024). ASFOSDA generalizes the problem to open-set volumetric medical image segmentation, where target domains include private classes of new anatomical structures or pathologies not present in the source data. Its evaluation uses MSD Spleen $h_t(x)=g_t(f_t(x))$5 FLARE, FLARE $h_t(x)=g_t(f_t(x))$6 AMOS CT, and Abdominal MR $h_t(x)=g_t(f_t(x))$7 AMOS MR, with both 3D U-Net and Swin UNETR backbones (Yang et al., 7 Jun 2026).

Multimodal variants add another dimension. DAM argues that prior SFADA methods rely exclusively on data-driven pseudo-labels and sparse human labels, while multimodal or external knowledge is not deeply fused. It therefore integrates multimodal supervision from a ViL model with scarce human annotations to form a dual supervisory signal (Chen et al., 29 Sep 2025). This shifts ASFDA from purely model-internal uncertainty and clustering mechanisms toward supervision fusion with foundation models.

Another common misconception is that ASFDA is necessarily a closed-set classification problem with a conventional pseudo-labeling loop. The published record includes label-space shift in SALAD, open-set medical segmentation in ASFOSDA, and multimodal foundation-model supervision in DAM (Kothandaraman et al., 2022, Yang et al., 7 Jun 2026, Chen et al., 29 Sep 2025).

6. Empirical record, efficiency, and research directions

Quantitatively, early classification results established the basic usefulness of the paradigm. With a $h_t(x)=g_t(f_t(x))$8 label budget, MHPL reports average target accuracies of $h_t(x)=g_t(f_t(x))$9 on Office-Home, $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$0 on Office-31, and $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$1 on VisDA-2017, outperforming the listed SFDA, active learning, and active DA baselines in its evaluation protocol (Wang et al., 2022). SALAD reports improvements of $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$2, across datasets and tasks, over prior adaptation methods that assume access to large amounts of annotated source data for adaptation (Kothandaraman et al., 2022).

LFTL emphasizes both accuracy and computational efficiency. On VisDA-C with $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$3 labels, it achieves $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$4, compared with SFUDA SOTA $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$5 for SHOT++ and ADA SOTA $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$6 for LADA, which uses source; at $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$7 labels it reaches $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$8, a $T_u^{(0)}=\{x_i\}_{i=1}^{n_u^{(0)}}$9 gain over MHPL and $B$0 over LADA. On Office-Home it rises from $B$1 at $B$2 labels to $B$3 at $B$4 labels, and on Office-31 from $B$5 to $B$6 over the same budgets. The same paper reports that adaptation on VisDA-C completes in approximately $B$7 versus $B$8 for prior ADA or SFUDA methods such as SHOT and LADA (Lyu et al., 2024).

ProULearn reports further improvements on four benchmark datasets. At $B$9 labels, it reaches $M_t$0 on Office-31 versus $M_t$1 for MHPL, $M_t$2 on Office-Home versus $M_t$3, $M_t$4 on DomainNet-126 versus $M_t$5, and $M_t$6 on VisDA-2017 versus $M_t$7 (Pan et al., 23 Jan 2025). Its ablations attribute gains to the central correlation loss, the HPE-based sample-selection mechanism, and the use of correlation rather than Euclidean or cosine distance in the distribution metric (Pan et al., 23 Jan 2025).

DAM reports that under a $M_t$8 labeling budget, CoreSet + DAM attains $M_t$9 on Office-31, $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$0 on Office-Home, and $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$1 on VisDA-C, while MHPL + DAM attains $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$2, $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$3, and $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$4, respectively. The paper summarizes these as consistent $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$5 absolute gains over strong SFADA baselines and states that DAM outperforms even methods that access source data or that use only single-path pseudo-labels (Chen et al., 29 Sep 2025).

In medical segmentation, UGTST reports a marked improvement with a mere $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$6 annotation budget, specifically an average Dice score enhancement of $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$7 and $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$8 in two target domains compared with state-of-the-art methods, reaching $R_T(M_t)=\mathbb E_{x\sim T}[\ell(M_t(x),y)]$9 Dice on BIDMC and $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$00 on HK, with HD$\mathcal D_t=\{x_i\}_{i=1}^{N_t}$01 values of $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$02 and $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$03, respectively (Luo et al., 2024). ASFOSDA reports, with a 3D U-Net backbone and $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$04 budget, $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$05 DSC on MSD$\mathcal D_t=\{x_i\}_{i=1}^{N_t}$06FLARE, $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$07 on FLARE$\mathcal D_t=\{x_i\}_{i=1}^{N_t}$08AMOS CT, and $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$09 on MR$\mathcal D_t=\{x_i\}_{i=1}^{N_t}$10AMOS MR, each with $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$11 versus all other SFDA and ASFDA methods in its comparison. The same work states that ASFOSDA recovers approximately $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$12 of fully supervised performance with only $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$13 labels and improves DSC by about $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$14 over the next best ASFDA method on Swin UNETR (Yang et al., 7 Jun 2026).

The main methodological tension in the literature is between one-shot querying and iterative query-and-adaptation. MHPL and ProULearn favor one-time selection, whereas LFTL and DAM explicitly reuse intermediate models during adaptation (Wang et al., 2022, Pan et al., 23 Jan 2025, Lyu et al., 2024, Chen et al., 29 Sep 2025). Another tension concerns supervision sources: classical ASFDA relies on queried labels and pseudo-labels, while DAM argues for deep fusion of sparse human labels with multimodal priors (Chen et al., 29 Sep 2025). Future directions identified within the literature include joint optimization of the active sampling strategy $\mathcal D_t=\{x_i\}_{i=1}^{N_t}$15 within the DAM loop, extension to open-set or partial-label SFADA, exploration of other foundation models such as ALIGN or adapter-based architectures, extension of ProULearn’s propensity concept to text and speech, incorporation of active augmentation of queried examples, and end-to-end optimization of the tree ensemble and feature extractor (Chen et al., 29 Sep 2025, Pan et al., 23 Jan 2025).

Taken together, the literature presents ASFDA as a family of source-free adaptation problems in which the marginal value of a small annotation budget is amplified by structured query criteria and adaptation objectives. The field began with minimum happy points and task-agnostic source-free active transfer, and it has expanded toward iterative persistence-aware querying, correlation-based sample exploration, multimodal bidirectional distillation, and open-set medical segmentation (Wang et al., 2022, Kothandaraman et al., 2022, Lyu et al., 2024, Pan et al., 23 Jan 2025, Chen et al., 29 Sep 2025, Luo et al., 2024, Yang et al., 7 Jun 2026).