Uncertainty-Aware Pseudo-Labels

• The paper introduces the concept that integrating predictive uncertainty into pseudo-labeling mitigates confirmation bias from high-confidence erroneous predictions.
• It categorizes methods—from hard thresholding and soft weighting to set-valued assignments—to enhance model calibration in semi-supervised and domain adaptation tasks.
• The work leverages strategies such as MC dropout, ensemble methods, and latent dispersion to refine label acceptance and maintain balanced, robust supervision.

Uncertainty-aware pseudo-labels are pseudo-labels whose generation, acceptance, weighting, or refinement is explicitly conditioned on predictive uncertainty rather than on raw confidence alone. In semi-supervised learning, source-free domain adaptation, graph learning, text recognition, dense prediction, positive-unlabeled learning, and semi-supervised regression, this design addresses a common failure mode: partially trained models can produce erroneous but highly confident predictions, and self-training on such predictions induces confirmation bias, degraded calibration, and cumulative error propagation (Rizve et al., 2021, Liu et al., 26 Mar 2025, Sun et al., 17 Oct 2025). The resulting literature treats uncertainty-aware pseudo-labeling not as a single algorithm but as a family of mechanisms: thresholded selection, soft weighting, uncertainty-smoothed label distributions, set-valued labels, region masking, and heteroscedastic regression targets.

1. Conceptual basis and motivation

Classical pseudo-labeling assigns labels to unlabeled samples from a model’s own predictions and retrains on those assignments. The central objection raised across the literature is that model confidence is not, by itself, a reliable indicator of pseudo-label correctness. Neural networks are often miscalibrated, so incorrect predictions can have high confidence; when such predictions are reused as supervision, the model amplifies its own early mistakes. This is the core argument in the uncertainty-aware pseudo-label selection framework UPS, which attributes the underperformance of conventional pseudo-labeling to erroneous high-confidence predictions from poorly calibrated models (Rizve et al., 2021).

The same problem appears in graph domains, but with an additional topological effect. In graph self-training, wrong node pseudo-labels do not remain local: message passing propagates errors over neighborhoods. GUST explicitly frames graph pseudo-labeling as a problem of over-confidence under label sparsity, emphasizing that early wrong labels can be reinforced both by retraining and by adjacency-induced smoothing (Liu et al., 26 Mar 2025). A related graph formulation, UGST, likewise argues that fixed hard pseudo-labels are overly aggressive in noisy graph regions and replaces them with soft, uncertainty-aware assignments refined by an EM-inspired scheme (Wang et al., 26 Mar 2025).

Sequence models expose a different failure mode. In semi-supervised text recognition, one incorrect autoregressive step can corrupt the remainder of a predicted sequence, so high-confidence greedy outputs are particularly brittle. Seq-UPS therefore combines beam-search hypothesis generation with uncertainty estimation over teacher-forced beams, explicitly targeting the combinatorial hypothesis space and error accumulation specific to sequence decoding (Patel et al., 2022).

These formulations share a common premise: pseudo-labels are useful only insofar as their uncertainty is modeled. A plausible implication is that uncertainty-aware pseudo-labeling should be understood less as a filtering heuristic and more as a calibration mechanism for self-supervised supervision.

2. Principal formulations of uncertainty-aware pseudo-labels

Representative methods differ mainly in what object is made uncertainty-aware: the sample, the class assignment, the region, the graph node, or the regression target.

Method	Pseudo-label form	Uncertainty action
UPS	Hard positive and negative pseudo-labels	Confidence-and-uncertainty thresholding
AggMatch	Aggregated soft pseudo-label	Consensus-weighted unsupervised loss
GUST	Soft class distribution $\hat{y}_i$	Mixes prediction with $\mathrm{Uniform}(K)$
Seq-UPS	Sequence pseudo-label from beam search	Hard acceptance by sequence uncertainty
RAPS-based framework	Prediction set $C^*(x,u,\hat{\tau}_{ccal})$	Set-valued pseudo-label construction
Heteroscedastic SSR	Continuous pseudo-label $\hat{y}_j$ with $\sigma_j^2$	Residual weighted by learned variance

A canonical thresholded formulation appears in UPS. For class $c$ of sample $i$, the selection indicator is

$$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$

so classes may be selected either as positives or as negatives. UPS then uses cross-entropy on selected positives and normalized negative cross-entropy when no positive is selected, extending pseudo-labeling to negative pseudo-labels for single-label and multi-label settings (Rizve et al., 2021).

A softer alternative appears in GUST, where uncertainty does not discard nodes but directly smooths their pseudo-label distributions. With node uncertainty $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$, GUST defines

$$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$

High uncertainty lowers $\mathrm{Uniform}(K)$0 and flattens $\mathrm{Uniform}(K)$1 toward the uniform distribution, so all unlabeled nodes contribute, but uncertain ones contribute as softened targets rather than hard labels (Liu et al., 26 Mar 2025).

AggMatch adopts yet another formulation. Instead of trusting a single model output, it aggregates pseudo-labels from similar confident instances stored in a class-balanced queue and computes multiple hypotheses by partitioning that queue. Confidence is then obtained from the entropy of votes across hypotheses, and the unsupervised loss is weighted softly rather than gated by a hard threshold. The paper reports that this confidence weighting outperforms uncertainty-based thresholding across $\mathrm{Uniform}(K)$2 by $\mathrm{Uniform}(K)$3–$\mathrm{Uniform}(K)$4 (Kim et al., 2022).

Set-valued formulations replace a single pseudo-label with a predictive set. The regularized conformal prediction framework based on RAPS constructs $\mathrm{Uniform}(K)$5 and then derives pseudo-labels from set membership, combining conformal coverage guarantees with UPS-style uncertainty-and-confidence filtering (Moezzi, 2023). In regression, the analog is heteroscedastic pseudo-labeling: instead of a discrete label and scalar confidence, the model learns an input-dependent variance $\mathrm{Uniform}(K)$6 and uses

$\mathrm{Uniform}(K)$7

so unreliable continuous pseudo-labels are downweighted through a learned heteroscedastic term (Sun et al., 17 Oct 2025).

3. Uncertainty estimation strategies

The literature uses several uncertainty estimators, each tied to a different view of pseudo-label reliability. A prominent family relies on stochastic predictive variation. UPS uses MC Dropout with $\mathrm{Uniform}(K)$8 passes and temperature scaling $\mathrm{Uniform}(K)$9, operationalizing uncertainty as the standard deviation across stochastic probabilities; Seq-UPS uses MC dropout ensembles with dropout rate $C^*(x,u,\hat{\tau}_{ccal})$0 and $C^*(x,u,\hat{\tau}_{ccal})$1 models over teacher-forced beam hypotheses; PUUPL defines aleatoric entropy, total entropy, and epistemic uncertainty through the decomposition $C^*(x,u,\hat{\tau}_{ccal})$2 over an ensemble of predictors (Rizve et al., 2021, Patel et al., 2022, Dorigatti et al., 2022).

A second family treats uncertainty as latent dispersion internal to the representation. GUST places a learned Gaussian over node embeddings,

$C^*(x,u,\hat{\tau}_{ccal})$3

and uses the average variance $C^*(x,u,\hat{\tau}_{ccal})$4 as a unified node-level uncertainty signal, without explicitly separating epistemic and aleatoric components (Liu et al., 26 Mar 2025). In camouflaged object detection, UR-COD uses a CVAE with prior and posterior Gaussian latent variables $C^*(x,u,\hat{\tau}_{ccal})$5 and $C^*(x,u,\hat{\tau}_{ccal})$6, so uncertainty is encoded implicitly in latent distributions and their KL regularization rather than in explicit entropy or variance thresholds (Kajiura et al., 2021).

A third family estimates uncertainty from agreement among multiple structured hypotheses. AggMatch partitions its class-balanced queue into $C^*(x,u,\hat{\tau}_{ccal})$7 disjoint subsets, forms $C^*(x,u,\hat{\tau}_{ccal})$8 aggregated pseudo-label hypotheses, converts them to hard labels, and defines confidence through vote entropy; low entropy corresponds to high consensus (Kim et al., 2022). CSA similarly uses multiple predictors, but it operationalizes uncertainty with Welch’s T-test on the separation between top-1 and top-2 class probabilities,

$C^*(x,u,\hat{\tau}_{ccal})$9

and only high-confidence samples are passed to a global optimal-transport pseudo-label allocation (Nguyen et al., 2022).

Dense prediction work often adopts spatial uncertainty maps. Bidirectional Uncertainty-Aware Region Learning computes pixel- or voxel-wise predictive entropy,

$\hat{y}_j$0

and defines high-uncertainty regions by image-wise entropy percentiles, while UDA for salient object detection estimates dense uncertainty by variance under reversible augmentations and then converts it into pixel weights $\hat{y}_j$1 with $\hat{y}_j$2 (Zhou et al., 11 Feb 2025, Yan et al., 2022).

More specialized estimators target task structure. UPL for imbalance transductive node classification uses Selective Edge Removal and defines uncertainty as the variance of node entropies across $\hat{y}_j$3 perturbed graphs. Pseudo-D, by contrast, treats training dynamics themselves as an uncertainty source: logits are aggregated across epochs and then calibrated with either temperature scaling or class-wise affine calibration to form uncertainty-aware pseudo-labels from neural network training dynamics (Teimuri et al., 2 Feb 2025, Gu et al., 15 Sep 2025).

4. Domain-specific realizations

Graph learning has become a major site for uncertainty-aware pseudo-labeling. GUST combines stochastic node labeling, an EM-like procedure, and a graph regularizer $\hat{y}_j$4. On Cora, CiteSeer, and PubMed, it reports state-of-the-art accuracies of $\hat{y}_j$5 $\hat{y}_j$6, $\hat{y}_j$7 $\hat{y}_j$8, and $\hat{y}_j$9 $\sigma_j^2$0, with the strongest gains when labeled data are extremely sparse (Liu et al., 26 Mar 2025). UGST presents a closely related EM-regularized graph self-training scheme and reports $\sigma_j^2$1 on Cora, $\sigma_j^2$2 on Citeseer, and $\sigma_j^2$3 on Pubmed under $\sigma_j^2$4 labeled data (Wang et al., 26 Mar 2025). For imbalanced transductive node classification, UPL combines Balanced Softmax with uncertainty-aware pseudo-labeling under graph perturbations and reports strong performance across homophilic and heterophilic benchmarks, including $\sigma_j^2$5 balanced accuracy on PubMed and $\sigma_j^2$6 on Chameleon at imbalance ratio $\sigma_j^2$7 (Teimuri et al., 2 Feb 2025).

In sequence recognition, Seq-UPS targets the mismatch between token-level and sequence-level reliability. For each unlabeled image, beam search produces top-$\sigma_j^2$8 hypotheses with $\sigma_j^2$9, MC dropout with $c$0 models estimates teacher-forced character entropies, and a sequence-level uncertainty

$c$1

is thresholded at $c$2. On handwriting with $c$3 labeled IAM, Seq-UPS reports accuracy $c$4 and CER $c$5, while on scene text with $c$6 labeled data it reports total accuracy $c$7, improving over the best SeqCLR baseline by about $c$8 (Patel et al., 2022).

Dense prediction and source-free adaptation favor region-wise or pixel-wise uncertainty. In medical image segmentation, Bidirectional Uncertainty-Aware Region Learning emphasizes high-uncertainty regions on labeled data and low-uncertainty regions on unlabeled data. On ACDC with $c$9 labeled data, it raises Dice from $i$0 to $i$1 when integrated into BCP, and on LA with $i$2 labeled data it raises Dice from $i$3 to $i$4 (Zhou et al., 11 Feb 2025). UPL-SFDA constructs target pseudo-labels from multiple perturbed decoder heads, keeps only pixels satisfying $i$5, and improves average Dice by $i$6, $i$7, and $i$8 percentage points on three source-free medical segmentation tasks (Wu et al., 2023). UP2D extends this line with refined prototype filtering, uncertainty-guided EMA, and quantile-based entropy minimization, reporting state-of-the-art performance on retinal fundus segmentation with improved boundary precision (Bui-Tran et al., 29 Oct 2025).

Uncertainty-aware pseudo-labeling has also proven effective in source-free domain adaptation beyond segmentation. The uncertainty-aware mean teacher for 3D object detection weights ROI classification losses by inverse predictive variance, $i$9, and improves Moderate/3D performance from $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$0 to $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$1 on Waymo $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$2 KITTI and from $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$3 to $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$4 on nuScenes $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$5 KITTI relative to a mean teacher without uncertainty weighting (Hegde et al., 2021). In salient object detection, uncertainty-aware self-training from synthetic clean labels achieves average improvements over EDNL of about $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$6 in $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$7, $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$8 in weighted $$g_{i,c} = \mathbf{1}\{ u(p_{i,c}) \le \kappa_p \}\mathbf{1}\{ p_{i,c} \ge \tau_p \} + \mathbf{1}\{ u(p_{i,c}) \le \kappa_n \}\mathbf{1}\{ p_{i,c} \le \tau_n \},$$9, and a $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$0 MAE reduction across six benchmarks (Yan et al., 2022).

5. Calibration, imbalance, and selection dynamics

A recurring result is that uncertainty-aware pseudo-labels are closely tied to calibration. UPS reports that filtering by uncertainty reduces Expected Calibration Error on the selected subset and materially improves pseudo-label accuracy over confidence-only selection (Rizve et al., 2021). PUUPL reaches the same conclusion in positive-unlabeled learning: it uses ensemble-based epistemic uncertainty to select balanced positive and negative pseudo-labels, reports substantial gains in highly imbalanced settings, and reduces test ECE by at least $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$1 versus naive pseudo-labeling (Dorigatti et al., 2022). Pseudo-D addresses calibration even more directly by constructing pseudo-labels from neural network training dynamics and then calibrating the aggregated logits with either global temperature scaling or class-wise affine calibration. On echocardiography, it reports study-level ECE $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$2 and AURC $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$3, the best among the listed methods (Gu et al., 15 Sep 2025).

Imbalance introduces an additional selection problem: thresholds that work for head classes often suppress tail classes. UDTS makes this explicit by using class-wise dynamic thresholds

$$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$4

where $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$5 is an EMA estimate of class competence and $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$6 is a global baseline. This makes pseudo-label selection stage-aware and class-aware, and the method reports at least approximately $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$7, $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$8, $$u_i := \bar{\sigma}_i^2 = \frac{1}{d}\sum_{\ell=1}^d \sigma_{i,\ell}^2$$9, and $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$0 accuracy improvements over FixMatch on CIFAR10-LT, CIFAR100-LT, STL-10-LT, and TissueMNIST, respectively (Yang et al., 2024).

A different response to threshold brittleness appears in UES. Rather than discarding uncertain pseudo-labels, it models their utility with long-tailed sample weights and head weights derived from ensemble disagreement. Even unreliable pseudo-labels retain a nonzero contribution. Combined with DualPose, this yields improvements of $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$1, $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$2, and $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$3 in PCK on Sniffing, FLIC, and LSP, and with FixMatch it improves accuracy by $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$4 on CIFAR-10 with $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$5 labeled samples and $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$6 on CIFAR-100 with $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$7 labeled samples (Wu et al., 13 Mar 2025).

Optimal-transport formulations target the same issue from a global assignment perspective. CSA first applies an uncertainty gate based on Welch’s T-test and then solves an entropically regularized optimal transport problem over the accepted high-confidence subset, with class quotas and reject slack. This replaces local greedy pseudo-label acceptance with a global allocation that explicitly addresses class balance and overconfidence (Nguyen et al., 2022).

A common misconception is that uncertainty-aware pseudo-labeling is synonymous with stricter filtering. The literature does not support that equivalence. Some methods, such as UPS and Seq-UPS, rely on hard selection, but others treat uncertainty as a weighting or smoothing signal rather than a rejection criterion. This suggests that the decisive variable is not whether uncertain samples are used, but how their supervision is modulated.

6. Limitations, misconceptions, and future directions

The principal limitation is that uncertainty estimation itself can be wrong. GUST notes that if $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$8 is miscalibrated, the method may under-smooth or over-smooth pseudo-labels; in very noisy or adversarial graphs, topology-aware regularization can still propagate errors (Liu et al., 26 Mar 2025). URL+CRL makes an analogous point for segmentation: erroneous pseudo-labels often concentrate in high-uncertainty regions, but directly removing such regions discards valuable supervision. In its ACDC ablation, removing high-uncertainty pseudo-labels yields Dice $$\alpha_i = \frac{1}{1 + \exp(\gamma \cdot \bar{\sigma}_i^2)}, \qquad \hat{y}_i = \alpha_i \cdot p_i + (1-\alpha_i)\cdot \mathrm{Uniform}(K).$$9, whereas downweighting them yields $\mathrm{Uniform}(K)$00 (Zhou et al., 11 Feb 2025).

Threshold sensitivity remains unresolved in several settings. Seq-UPS reports that too small a threshold $\mathrm{Uniform}(K)$01 selects too few samples and too large a threshold admits noisy labels, with $\mathrm{Uniform}(K)$02 giving a good trade-off in its experiments (Patel et al., 2022). Conformal approaches improve set calibration, but their coverage guarantees depend on exchangeability and can degrade under distribution shift, class imbalance, or calibration drift across pseudo-labeling iterations (Moezzi, 2023). Long-tailed methods such as UDTS and UES reduce dependence on a single global threshold, yet both still require tuning of class-wise dynamics or weighting schedules (Yang et al., 2024, Wu et al., 13 Mar 2025).

Computational overhead is another recurring concern. MC dropout, ensemble inference, multi-hypothesis voting, repeated beam evaluation, and second-order optimization all increase cost relative to naive pseudo-labeling. This is explicit in methods such as Seq-UPS, PUUPL, the uncertainty-aware mean teacher for 3D detection, and heteroscedastic semi-supervised regression, although several papers note that the additional cost is moderate relative to their backbones or is confined to pseudo-label generation phases (Dorigatti et al., 2022, Hegde et al., 2021, Sun et al., 17 Oct 2025).

A second misconception is that uncertainty-aware pseudo-labels primarily estimate epistemic uncertainty. In fact, the literature mixes several uncertainty notions. GUST uses a unified latent variance without explicit epistemic-versus-aleatoric decomposition; URL+CRL uses predictive entropy; Pseudo-D emphasizes aleatoric uncertainty captured through training dynamics; heteroscedastic SSR explicitly models input-dependent noise $\mathrm{Uniform}(K)$03 (Liu et al., 26 Mar 2025, Zhou et al., 11 Feb 2025, Gu et al., 15 Sep 2025, Sun et al., 17 Oct 2025). The absence of a single dominant definition suggests that “uncertainty-aware” is a functional rather than ontological label: it denotes any pseudo-labeling mechanism that uses uncertainty signals to reduce training noise.

Future directions stated across the papers include more expressive priors and variational objectives for disentangling uncertainty types, uncertainty-aware adjacency or message-passing transformations in graphs, learnable region- and boundary-aware denoising, adaptive curricula for thresholds, class-aware calibration under imbalance, multi-sample predictive integration, and extension to dynamic or temporal graphs, speech recognition, machine translation, and other structured prediction problems (Liu et al., 26 Mar 2025, Patel et al., 2022, Zhou et al., 11 Feb 2025). A plausible implication is that the field is moving away from binary pseudo-label acceptance and toward calibrated supervision distributions whose uncertainty is task-structured, class-aware, and stage-dependent.