CoNMix: Source-Free Domain Adaptation

• CoNMix is a source-free domain adaptation method that integrates nuclear-norm maximization, pseudo-label refinement, and consistency loss to adapt models without access to source data.
• Its three-stage framework leverages teacher-student distillation with MixUp to merge multiple target-specific adaptations into a robust, unified model.
• CoNMix demonstrates improved performance using Vision Transformer backbones and has been successfully reformulated for audio classification under noisy conditions.

CoNMix—short for Consistency with Nuclear-Norm Maximization and MixUp knowledge distillation—is a source-free adaptation framework introduced for source-free single-target domain adaptation (SF-STDA) and source-free multi-target domain adaptation (SF-MTDA), settings in which a model trained on labeled source data must be adapted using only unlabeled target data because the source dataset is unavailable during adaptation (Kumar et al., 2022). In later work on audio classification under background noise, CoNMix was treated as a source-free test-time domain adaptation method and modified for Mel-spectrogram inputs and audio-appropriate augmentations, where the modified version produced the highest classification accuracy among the tested test-time adaptation methods (Shao et al., 21 Jul 2025). The string “CoNMix” is also used in unrelated literatures, notably for dialogue-context mixing in conversational modeling and for the ContaminatedMixt package in robust mixture modeling (Poddar et al., 2022, Punzo et al., 2016).

1. Origin, scope, and problem formulation

The original CoNMix formulation addresses a source-free regime in which a labeled source model is available, but the labeled source samples are not. The paper explicitly targets both Single and Multi target Domain Adaptation (SMTDA) for the source-free paradigm and introduces Source-free Multi-target Domain Adaptation as a novel task (Kumar et al., 2022). The constraint is operational rather than cosmetic: after source training, the source data are discarded, and target adaptation must proceed from the source-trained network, unlabeled target data, target pseudo labels, and target-domain statistics.

The framework is organized into three stages. Stage-1 trains a source model on $(\mathcal{X_S}, \mathcal{Y_S})$. Stage-2 adapts the model to each target domain separately using unlabeled target data and a combination of augmentation consistency, nuclear-norm maximization, pseudo-label refinement, and pseudo-label cross-entropy. Stage-3, used for MTDA, distills knowledge from multiple target-specific teachers into a single student using MixUp Knowledge Distillation (MKD) (Kumar et al., 2022).

This formulation differs from conventional UDA or STDA pipelines because source-target alignment cannot rely on retained source batches, adversarial source-target discrimination, or direct source-target batch-statistic matching. The method therefore places adaptation pressure on target-side regularization and pseudo-label management rather than on joint source-target optimization (Kumar et al., 2022).

2. Stage-2 objective: consistency, nuclear norm, and pseudo-label refinement

For a target domain, CoNMix initializes the target backbone with the source-trained weights, freezes the classifier, and updates the target backbone with SGD. The stage-2 objective is

$$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$

Here, $\mathcal{L}_{NM}$ is the nuclear-norm maximization term, $\mathcal{L}_{CE}^{P_l}$ is pseudo-label cross-entropy, and $\mathcal{L}_{cons}$ is a consistency loss between weak and strong label-preserving augmentations (Kumar et al., 2022).

The consistency component uses weak and strong augmentations of the same target batch. The weak prediction is normalized through an expectation-ratio correction so that the first-order batch statistic matches the first-order overall target statistic, and the normalized weak prediction serves as a soft pseudo-ground-truth for the strong augmentation. The resulting loss is

$\mathcal{L}_{cons} = -\mathbb{E}_{(y_{ts})\in {Y}_{ts}^{B}} \sum_{k=1}^{K}\Hat{y}_{tw}^{k}\log {y}_{ts}^{k}.$

This places CoNMix within the line of consistency-regularized target adaptation methods, but under source-free constraints and with target-statistic correction built into the weak-view target (Kumar et al., 2022).

The nuclear-norm component is the structural regularizer that the paper identifies as especially important. Let $A\in\mathbb{R}^{B\times K}$ be the classification-response matrix for a batch of size $B$ and $K$ classes. CoNMix uses the Frobenius approximation

$\mathcal{L}_{NM} = -\|A\|_{F},$

so minimizing $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$0 maximizes $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$1. The paper states that this encourages sharper, more class-separable predictions while preserving batch diversity, and further argues that nuclear-norm maximization is more effective than information maximization in the source-free setup considered there (Kumar et al., 2022).

Pseudo labels are produced by an iterative weighted k-means-style procedure similar to SHOT. Initial class centers are computed from target features weighted by model predictions, and labels are assigned by maximum cosine similarity to those centers. Because these pseudo labels are noisy, CoNMix applies pseudo-label refinement (PLR) using temporal ensemble and a cluster-consensus matrix $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$2 across consecutive epochs:

$$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$3

and

$$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$4

The refined labels are then used in the pseudo-label cross-entropy term (Kumar et al., 2022). A central practical point is that CoNMix does not rely on pseudo labels alone; the refinement stage is explicitly designed to reduce label noise before those labels drive the target update.

3. MixUp Knowledge Distillation and the multi-target setting

In SF-MTDA, stage-2 produces one source-free adapted model per target domain. CoNMix then distills these target-specific teachers into a single student through offline knowledge distillation. The student is initialized with ImageNet-pretrained weights, and the teacher-generated pseudo labels are stored for each target domain (Kumar et al., 2022).

The distillation stage uses MixUp to generate virtual intermediate-domain samples across target domains:

$$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$5

The MKD loss is written as

$$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$6

The paper interprets the virtual sample as an intermediate domain acting as an implicit regularizer, preventing overfitting to any one target and improving generalization to unlabeled target domains (Kumar et al., 2022).

A recurrent misconception is to treat CoNMix as merely a target-side self-training scheme with MixUp. The architecture of the method is more specific. MixUp enters only in Stage-3, after one source-free STDA model has already been trained per target domain; the single-target adaptation behavior is governed instead by the interaction among nuclear-norm maximization, pseudo-label refinement, and augmentation consistency (Kumar et al., 2022).

4. Backbone selection, reported performance, and ablation behavior

A major empirical claim of the original paper is that a Vision Transformer backbone is better suited to source-free domain adaptation than a standard CNN such as ResNet-50. The stated reasons are that VT has a global receptive field at every stage, that self-attention can focus more on salient objects and less on background or domain-specific nuisance factors, and that VT features exhibit better domain-transferability and class-discriminability. The paper supports this with smaller $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$7-distance, better t-SNE alignment, and higher benchmark accuracy (Kumar et al., 2022).

The reported results are summarized below.

Setting	Reported result
Office-Home SF-STDA	76.0% avg
Office-31 SF-STDA	89.4% avg
DomainNet SF-STDA	51.6% avg
Office-Home SF-MTDA	75.4% avg
Office-31 SF-MTDA	84.9% avg
Office-Caltech SF-MTDA	96.9% avg
DomainNet SF-MTDA	33.7% avg

On Office-Home SF-STDA, CoNMix (DeiT-S) achieves 76.0% avg, compared with 75.6% for SHOT++*(DeiT-S), 75.1% for SHOT*(DeiT-S), and 74.7% for CDTrans (DeiT-S). On DomainNet SF-STDA, the paper reports 51.6% avg and highlights a 6.0% improvement over prior STDA in its discussion. In SF-MTDA on Office-Home, CoNMix reaches 75.4% avg, compared with 67.4% for source training with DeiT-S, 70.0% for domain aggregation, and 73.1% for SHOT STDA initialization in Stage-3 (Kumar et al., 2022).

The ablation study attributes most of the gain to the nuclear-norm term. Using only $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$8 or only $$\mathcal{L}_{total} = \lambda_1 \mathcal{L}_{NM} + \lambda_2 \mathcal{L}_{CE}^{P_l} + \lambda_3 \mathcal{L}_{cons}.$$9 gives very poor performance; $\mathcal{L}_{NM}$0 is the most important loss component; adding $\mathcal{L}_{NM}$1 to $\mathcal{L}_{NM}$2 improves performance modestly; and the best results occur when all three losses are combined:

$\mathcal{L}_{NM}$3

Within the paper’s own interpretation, this indicates that CoNMix depends less on any single heuristic than on the interaction between prediction sharpening, pseudo-label denoising, and augmentation-stability constraints (Kumar et al., 2022).

5. Audio-domain reformulation in test-time adaptation

A later study investigates CoNMix in a different regime: audio classification under domain shift caused by background noise, using test-time adaptation (TTA) rather than the original image-domain source-free DA setting (Shao et al., 21 Jul 2025). In that work, CoNMix is treated as a source-free test-time domain adaptation method that exploits the entire unlabeled target test set before making predictions. The paper characterizes it as the most sophisticated of the three TTA methods studied and contrasts it with TTT and TENT, which update on mini-batches or single samples.

The paper describes the original CoNMix idea as combining several complementary adaptation signals rather than relying on a single heuristic such as entropy minimization. It uses a teacher-student paradigm in which a pre-trained model is copied into a teacher for the target domain, pseudo-labels are generated for the entire target set, and a student model is adapted using a weighted mixture of losses encouraging confident predictions, consistency under augmentation, and agreement with pseudo-labels. The single-target adaptation objective is written as

$\mathcal{L}_{NM}$4

For audio, however, the method is not used exactly as originally proposed. Waveforms are first converted into Mel-spectrograms so that the inputs can be processed like 2D images; image-style augmentations are replaced with audio-appropriate ones such as time shifting; and the pseudo-label term is modified after the authors observed that the original pseudo-label loss could produce negative adaptation effects, especially on SpeechCommands (Shao et al., 21 Jul 2025).

The audio-specific modifications are dataset-dependent. For AudioMNIST, the pseudo-label loss is removed entirely during adaptation because the authors found that it hurt adaptation. For SpeechCommands and its variants, the original softmax-based pseudo-label cross-entropy is replaced with a log-softmax / negative log-likelihood formulation, yielding

$\mathcal{L}_{NM}$5

The target-domain test sets are corrupted with one-second clips of background noise at 10 dB and 3 dB SNR using the noises doing the dishes (DD), exercise bike (EB), and running tap (RT). The study evaluates AudioMNIST (AM) and SpeechCommands V1 (SC), and also constructs SpeechCommands Random (SCR) and SpeechCommands Numbers (SCN) (Shao et al., 21 Jul 2025).

The reported results are strong. On AudioMNIST, modified CoNMix yields error rates of 6.27% under DD-10, 12.85% under DD-3, 5.31% under EB-10, 15.01% under EB-3, 5.37% under RT-10, and 12.75% under RT-3. On SpeechCommands with 10 dB DD noise, Tent Adaptation and Norm Adaptation reach 77.32% and 82.74% error, respectively, while TTT reaches 9.17% and modified CoNMix reaches 8.49%. Under 3 dB DD noise on SC, the error rates are 82.75%, 85.87%, 19.93%, and 14.47% for Tent Adaptation, Norm Adaptation, TTT, and CoNMix, respectively (Shao et al., 21 Jul 2025).

The paper further reports that TENT and TTT can exhibit negative adaptation on SpeechCommands, meaning the adapted error is worse than the no-adaptation baseline, whereas CoNMix generally improves over the baseline. It attributes the superior performance of CoNMix to two factors: adaptation using the entire test set for multiple epochs, which is less stochastic than per-batch or per-sample updates, and a more expressive architecture based on a Vision Transformer with a ResNet50-based embedding network and transformer blocks. The authors also conclude that the pseudo-label component must be treated carefully in audio tasks: if it is poorly aligned with the domain shift, it can harm adaptation rather than help it (Shao et al., 21 Jul 2025). This suggests that CoNMix’s multi-signal adaptation framework is portable across modalities only when the pseudo-label objective and augmentation design are reformulated to match modality-specific failure modes.

6. Terminological ambiguity and unrelated uses

The name “CoNMix” does not denote a single research object across all arXiv usage represented here. In conversational modeling, ConMix is a token-level augmentation strategy within DialAug for retrieval-based dialogue systems. For a context $\mathcal{L}_{NM}$6, the method samples another context $\mathcal{L}_{NM}$7 from the batch and a Bernoulli mask $\mathcal{L}_{NM}$8, then constructs

$\mathcal{L}_{NM}$9

with $\mathcal{L}_{CE}^{P_l}$0 in the experiments, so that roughly 30% of tokens are replaced while EOT markers are preserved. This ConMix is used together with an auxiliary contrastive loss and yields Recall@1 gains of +0.8% on Ubuntu V2, +1.9% on Advising DSTC7, +1.0% on Ubuntu DSTC7, and +2.3% on Taskmaster over the plain bi-encoder baseline (Poddar et al., 2022). Methodologically, this work concerns robust conversational response ranking, not source-free domain adaptation.

In robust model-based clustering and classification, “CoNMix” is also used to refer to ContaminatedMixt, an R package for fitting mixtures of multivariate contaminated normal distributions. Each cluster is modeled by a good subcomponent and a bad subcomponent with the same mean but inflated covariance for the bad points. The package fits 14 parsimonious covariance structures, uses an Expectation-Conditional Maximization algorithm, supports model selection with criteria including AIC, BIC, and ICL, and enables automatic detection of mild outliers / bad points via posterior probabilities (Punzo et al., 2016). This usage belongs to the literature on robust clustering and classification under elliptically contoured groups, not to domain adaptation.

For researchers, the practical implication is straightforward: in current arXiv usage, CoNMix most commonly denotes the source-free DA framework of 2022 when the topic is domain adaptation, but similar capitalization can also refer to a dialogue augmentation operator or to contaminated-normal mixture modeling in statistics.