Uncertainty-Guided Hybrid Framework

• The uncertainty-guided hybrid framework integrates heterogeneous modeling strategies by embedding explicit uncertainty estimates to guide loss weighting, gating, and runtime decisions.
• It leverages diverse uncertainty representations—such as GP posterior variance, entropy measures, and conformal prediction—to optimize performance across tasks like remote sensing, robotics, and medical imaging.
• Empirical results indicate enhanced accuracy and selective computation benefits, although challenges remain in calibration consistency and reconciling varied uncertainty semantics.

“Uncertainty-guided hybrid framework” denotes a class of computational architectures in which explicit uncertainty estimates are embedded into the interaction among heterogeneous modeling components rather than appended as a purely diagnostic output. In recent arXiv literature, the phrase has been attached to physics-constrained Gaussian process regression with deep kernels, dual-level planning-and-reasoning agents for web interaction, uncertainty-gated perception stacks in remote sensing and medical imaging, hybrid industrial monitoring pipelines wrapped by conformal prediction, and post-hoc control layers for frozen robotic policies (Chang et al., 2022, Zhang et al., 20 Apr 2026, Chen et al., 17 Apr 2026). Earlier hybrid formulations combined symbolic truth-maintenance with numeric evidence calculus, showing that the underlying idea predates current deep learning systems (D'Ambrosio, 2013).

1. Conceptual scope and defining properties

Across the cited works, “hybrid” does not refer to one fixed algorithm. It refers instead to explicit coupling between distinct computational regimes: data-driven and physics-constrained learning, planning and search, 2D and 3D segmentation, rule-based and AI-based redaction, or symbolic and numeric reasoning. “Uncertainty-guided” likewise has a broad operational meaning: the uncertainty signal may weight losses, gate refinement, choose a planner, reject unsafe outputs, trigger human review, or quantify coverage. This suggests that the concept is best understood as an architectural pattern rather than a single model family.

Representative formulation	Hybrid composition	Uncertainty signal
"A hybrid data driven-physics constrained Gaussian process regression framework with deep kernel for uncertainty quantification" (Chang et al., 2022)	GPR with deep kernel + Boltzmann–Gibbs physics likelihood	GP posterior and physics loss
"WebUncertainty: Dual-Level Uncertainty Driven Planning and Reasoning For Autonomous Web Agent" (Zhang et al., 20 Apr 2026)	adaptive planning + ConActU-guided MCTS reasoning	task uncertainty, epistemic uncertainty, aleatoric uncertainty
"StreamSplit: Continuous Audio Representation Learning via Uncertainty-Guided Adaptive Splitting" (Quan et al., 26 May 2026)	streaming contrastive learning + server-side Hybrid Loss + PPO splitter	GMM-posterior entropy, latency-energy reward
"CloudMamba: An Uncertainty-Guided Dual-Scale Mamba Network for Cloud Detection in Remote Sensing Imagery" (Yang et al., 8 Apr 2026)	CNN–Mamba coarse segmenter + uncertainty-guided refiner	per-pixel $U(x,y)=1-2|P_c(x,y)-0.5|$
"A Hybrid Intelligent Framework for Uncertainty-Aware Condition Monitoring of Industrial Systems" (Ahang et al., 10 Apr 2026)	feature-level or model-level fusion + conformal layer	coverage, prediction-set size, abstention
"ReconVLA: An Uncertainty-Guided and Failure-Aware Vision-Language-Action Framework for Robotic Control" (Chen et al., 17 Apr 2026)	frozen VLA + conformalized quantile regression + Mahalanobis detector	action-level conformal bound, state anomaly score

A recurrent design principle is the separation of a primary predictive mechanism from a secondary mechanism that estimates reliability, then feeds that estimate back into computation. In some systems the feedback is differentiable and part of training; in others it is a runtime control signal.

2. Uncertainty representations and formal signals

One major lineage uses fully probabilistic surrogates. In the physics-constrained deep-kernel GPR framework, the physics-only likelihood is written as

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$

and the hybrid objective becomes

$$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$

At test time, the predictive distribution is Gaussian,

$$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$

A related GP-based formulation for MR-guided radiotherapy uses the standard GP posterior mean and variance together with a scalar uncertainty statistic

$$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$

which is thresholded for rejection (Chang et al., 2022, Huttinga et al., 2022).

A second lineage uses entropy-like or heuristic confidence scores. WebUncertainty normalizes candidate-action confidences by

$$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$

defines normalized predictive entropy

$$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$

and then sets

$EU=1-E, \qquad AU=H_{\rm norm}\times E.$

StreamSplit uses the entropy of a local GMM posterior,

$U_t=-\sum_{c=1}^C p(c|z_t)\log p(c|z_t),$

as part of the RL state for split decisions. CloudMamba adopts a deterministic per-pixel ambiguity proxy,

$U(x,y)=1-2|P_c(x,y)-0.5|,$

so that $$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$0 yields maximum uncertainty and confident logits near $$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$1 or $$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$2 yield low uncertainty (Zhang et al., 20 Apr 2026, Quan et al., 26 May 2026, Yang et al., 8 Apr 2026).

A third lineage emphasizes calibration guarantees. In the industrial monitoring framework, nonconformity is defined by

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$3

and the conformal prediction set is

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$4

ReconVLA uses conformalized quantile regression with one-sided error bound

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$5

and chooses the action with minimal predicted bound. GUIDE attaches an evidential meta-model to a frozen classifier, forming Dirichlet concentrations

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$6

with epistemic uncertainty

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$7

These constructions differ materially from heuristic entropy maps because they are designed either to guarantee coverage under exchangeability or to learn when uncertainty should be expressed (Ahang et al., 10 Apr 2026, Chen et al., 17 Apr 2026, Barker et al., 29 Sep 2025).

A fourth lineage decomposes total predictive variance or propagates mixed uncertainty classes. HybridFlow writes

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$8

with a conditional normalizing flow for aleatoric uncertainty and a probabilistic predictor for epistemic uncertainty. DiffHybrid-UQ uses the law of total variance inside a differentiable hybrid solver and propagates mean and variance through nonlinear operators by unscented transformation. The decoupled M-PDEM framework instead treats epistemic parameters as part of an augmented random space, with

$$P(y|X,\mathcal P)=Z(\beta,X,\mathcal P)^{-1}\exp[-\beta L_{\rm phy}(y,X)],$$9

so that conditional response densities can be evolved numerically for mixed aleatory and epistemic inputs (Katwyk et al., 6 Oct 2025, Akhare et al., 2023, Luo et al., 11 Sep 2025).

3. Modes of hybridization

A large subset of the literature uses hybridization to reconcile physical structure with data efficiency. The deep-kernel GPR framework combines observed data with PDE residual information through a Boltzmann–Gibbs term. The composite-plate impact framework combines an identified First-Order Shear Deformation Theory model with multi-fidelity GPR for localization and adaptive regularization for force reconstruction. Chu and Qian’s turbulence work uses a lightweight 1D-CNN to modulate the Eigenspace Perturbation Method while preserving realizability of the Reynolds-stress correction. The module-based multi-physics framework allows different stochastic modules to use intrusive, non-intrusive, or semi-intrusive UQ while exchanging generalized polynomial chaos information. DiffHybrid-UQ similarly embeds known PDE operators as fixed differentiable layers and learns unknown components and physical parameters with Bayesian model averaging (Chang et al., 2022, Xiao et al., 13 Jul 2025, Chu et al., 7 Nov 2025, Mittal et al., 2014, Akhare et al., 2023).

Another subset uses uncertainty to localize where additional computation should be spent. CloudMamba performs coarse segmentation, computes a per-pixel uncertainty map, and applies a second-stage refiner only to hard regions. The glioma framework uses a spherical-projection 2D nnU-Net to generate voxel-wise entropy, ranks 3D windows by cumulative uncertainty, and applies localized 3D refinement only within those windows. The DICOM de-identification system splits into a metadata path and a pixel path, then uses uncertainty-aware Faster R-CNN proposals plus OCR and NER to determine whether automated masking is acceptable or whether manual review is required. SCOUT couples probabilistic scene-graph construction to active traversal so that ambiguous objects are revisited while unseen free space is explored (Yang et al., 8 Apr 2026, Yang et al., 21 Jul 2025, Naddeo et al., 31 Jul 2025, Mao et al., 4 Jun 2026).

A third subset makes uncertainty an explicit control variable in decision-making. WebUncertainty separates task-level planning uncertainty from action-level reasoning uncertainty, using the former to choose between explicit and implicit planning and the latter to shape Monte Carlo Tree Search. GUIDE for robot navigation augments the MDP state with local uncertainty bounds from Task-Specific Uncertainty Maps and penalizes violations during Soft Actor-Critic training. ReconVLA wraps a frozen VLA policy with action-level conformal bounds and state-level Mahalanobis detection. RL for complex model transformations treats human advice as subjective-logic opinions, calibrates belief, disbelief, uncertainty, and base rate, and fuses them into the policy via Belief Constraint Fusion (Zhang et al., 20 Apr 2026, Puthumanaillam et al., 20 May 2025, Chen et al., 17 Apr 2026, Dagenais et al., 25 Jun 2025).

A historically earlier formulation expresses hybridization in symbolic-numeric terms. D’Ambrosio combines an Assumption-based Truth Maintenance System with Dempster–Shafer-style support calculations by treating the ATMS as a symbolic algebra system for uncertainty management. In that construction, symbolic labels encode dependence structure, while numeric support is deferred to query time (D'Ambrosio, 2013).

4. Training, gating, and rejection mechanisms

At the objective level, hybrid frameworks often sum or multiply losses arising from different modeling commitments. The physics-constrained GPR system minimizes a single loss combining physics fit, quadratic GP data fit, and log-determinant regularization. StreamSplit’s server-side refinement uses a Hybrid Loss composed of a task loss, a Sliced-Wasserstein term aligning the embedding marginal to a uniform prior on the hypersphere, and a Laplacian regularizer preserving temporal smoothness under dropped or delayed frames. DiffHybrid-UQ combines heteroscedastic Gaussian likelihoods with deep-ensemble Bayesian learning and unscented-transformation propagation through nonlinear known operators (Chang et al., 2022, Quan et al., 26 May 2026, Akhare et al., 2023).

At the gating level, uncertainty frequently determines whether a second computational path is invoked. CloudMamba forms an acceptance mask

$$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$0

and routes only the complement to the refiner. The glioma framework computes cumulative uncertainty

$$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$1

over sliding 3D windows and retains high-scoring windows under an overlap constraint of 40% of the kernel volume. StreamSplit makes a split decision every $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$2 frames, using average uncertainty, CPU utilization, and upload bandwidth to choose the split point $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$3. SCOUT scores candidate viewpoints by

$$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$4

thereby balancing semantic certainty gain, geometric coverage gain, and travel cost (Yang et al., 8 Apr 2026, Yang et al., 21 Jul 2025, Quan et al., 26 May 2026, Mao et al., 4 Jun 2026).

At the rejection and oversight level, uncertainty becomes a safety filter. The MR-guided radiotherapy system sets a threshold $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$5 from the training-set percentile of $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$6 and rejects inferred deformation vector fields when $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$7. ReconVLA fits a Gaussian to safe states, computes Mahalanobis distance

$$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$8

and triggers intervention when $$\mathcal L(\theta)=\beta L_{\rm phy}(y,X)+y^\top(K_\theta+\sigma^2I)^{-1}y+\log\det(K_\theta+\sigma^2I).$$9; it also executes the action sample with minimal conformal upper bound. The DICOM pipeline accepts low-variance detector proposals for OCR and NER, but flags high-uncertainty cases for a human-in-the-loop interface. The industrial monitoring framework evaluates coverage, average prediction-set size, and abstention rate, making uncertainty part of the output contract rather than an afterthought (Huttinga et al., 2022, Chen et al., 17 Apr 2026, Naddeo et al., 31 Jul 2025, Ahang et al., 10 Apr 2026).

At the search and policy-shaping level, the uncertainty signal directly modulates exploration. WebUncertainty uses ConActU priors in PUCT-style selection and uses $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$0-dependent penalties in simulation. GUIDE for navigation replaces the original reward with

$$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$1

where $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$2 penalizes local violation of task-specific uncertainty tolerance. In RL for complex model transformations, advice uncertainty $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$3 is calibrated from state distance or manually set, converted together with belief and disbelief into a subjective-logic opinion, and fused into the policy before standard policy-gradient learning (Zhang et al., 20 Apr 2026, Puthumanaillam et al., 20 May 2025, Dagenais et al., 25 Jun 2025).

5. Representative empirical results

Domain	Paper	Reported result
Elliptic PDE surrogate	"A hybrid data driven-physics constrained Gaussian process regression framework with deep kernel for uncertainty quantification" (Chang et al., 2022)	validation MSE fell to $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$4 on the Exponential-of-GRF case; below $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$5 on the channelised field
Autonomous web agent	"WebUncertainty: Dual-Level Uncertainty Driven Planning and Reasoning For Autonomous Web Agent" (Zhang et al., 20 Apr 2026)	WebArena, GPT-4-Turbo: $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$6 SR vs $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$7; WebVoyager: $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$8 vs $$y_\ast|x_\ast,X,Y\sim \mathcal N(\mu_\ast,\Sigma_\ast).$$9
Streaming edge audio	"StreamSplit: Continuous Audio Representation Learning via Uncertainty-Guided Adaptive Splitting" (Quan et al., 26 May 2026)	latency $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$0 ms $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$1 ms; bandwidth $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$2 KB $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$3 KB; energy $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$4 mJ $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$5 mJ; AudioSet balanced $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$6
Cloud detection	"CloudMamba: An Uncertainty-Guided Dual-Scale Mamba Network for Cloud Detection in Remote Sensing Imagery" (Yang et al., 8 Apr 2026)	GF1_WHU: mIoU $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$7, F1 $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$8, OA $$\sigma_t=\sum_{j=1}^{N_{\mathcal Y}} s_j^2[\Sigma_t]_{jj},$$9; Levir_CS: mIoU $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$0, F1 $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$1, OA $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$2
Industrial condition monitoring	"A Hybrid Intelligent Framework for Uncertainty-Aware Condition Monitoring of Industrial Systems" (Ahang et al., 10 Apr 2026)	Parallel model-level + stacking (LogR): $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$3; Feature-level (Base+Lag+Residual) + stacking (LogR): $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$4; average set size $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$5 at $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$6 for Base+Lag+Residual
MR-guided radiotherapy	"Gaussian Processes for real-time 3D motion and uncertainty estimation during MR-guided radiotherapy" (Huttinga et al., 2022)	inference frame rate up to $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$7 Hz including data acquisition and reconstruction; 75th-percentile EPE below $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$8mm in silico
Impact monitoring in composites	"Physics-guided impact localisation and force estimation in composite plates with uncertainty quantification" (Xiao et al., 13 Jul 2025)	MF GPR with $$p_i=\frac{c_i}{\sum_{j=1}^K c_j}, \qquad E=\frac{1}{K}\sum_{i=1}^K c_i,$$9 experimental impacts: mean localization error $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$0 mm; adaptive $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$1 yields peak-force errors $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$2
VLA robotic control	"ReconVLA: An Uncertainty-Guided and Failure-Aware Vision-Language-Action Framework for Robotic Control" (Chen et al., 17 Apr 2026)	average success: default $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$3 $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$4, mean action $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$5, ReconVLA (CQR) $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$6; SMD AUC $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$7 on real UR5 experiments
Glioma segmentation	"A Voxel-Wise Uncertainty-Guided Framework for Glioma Segmentation Using Spherical Projection-Based U-Net and Localized Refinement in Multi-Parametric MRI" (Yang et al., 21 Jul 2025)	Dice scores: ET $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$8, TC $$H_{\rm norm}=-\frac{1}{\log K}\sum_{i=1}^K p_i\log p_i,$$9, WT $EU=1-E, \qquad AU=H_{\rm norm}\times E.$0

These reported results cluster around limited-label regimes, hard-region refinement, out-of-distribution detection, and real-time safety filtering. This suggests that the most visible benefits of uncertainty-guided hybridization often appear not in average-case prediction alone, but in selective computation, calibration, and failure management.

6. Limitations, misconceptions, and open problems

One common misconception is that all “uncertainty-guided” systems estimate the same object. The surveyed literature uses sharply different semantics: entropy of a local GMM posterior, evidential concentration mass, p-box envelopes in augmented random space, and subjective-logic uncertainty are not interchangeable quantities, even when they share the same vocabulary of confidence or ambiguity. A second misconception is that hybridization implies end-to-end differentiable Bayes-consistent inference; many frameworks instead use deterministic heuristics, calibrated set prediction, or manually specified uncertainty mappings as the control signal (Quan et al., 26 May 2026, Barker et al., 29 Sep 2025, Luo et al., 11 Sep 2025, Dagenais et al., 25 Jun 2025).

Evaluation caveats are explicit in several papers. The deep-kernel physics-constrained GPR work reports that no direct baseline GPR or PINN comparison was reported. In industrial monitoring, parallel fusion can become conservative if residual classifiers dominate soft votes, whereas weighted voting or stacking mitigates this. In CloudMamba, the full-dataset two-stage gain is small—single-stage mIoU $EU=1-E, \qquad AU=H_{\rm norm}\times E.$1 versus two-stage mIoU $EU=1-E, \qquad AU=H_{\rm norm}\times E.$2—while the benefit is larger on the top 10% “hard” samples, where mIoU rises from $EU=1-E, \qquad AU=H_{\rm norm}\times E.$3 to $EU=1-E, \qquad AU=H_{\rm norm}\times E.$4 and F1 from $EU=1-E, \qquad AU=H_{\rm norm}\times E.$5 to $EU=1-E, \qquad AU=H_{\rm norm}\times E.$6. In the turbulence framework of Chu and Qian, no formal uncertainty bands or probabilistic calibration metrics were produced, and only the TKE magnitude was corrected while anisotropy shape and orientation were left unchanged (Chang et al., 2022, Ahang et al., 10 Apr 2026, Yang et al., 8 Apr 2026, Chu et al., 7 Nov 2025).

The open problems identified by the papers are correspondingly heterogeneous. For GP-based radiotherapy, more expressive motion models and multi-task GP structure are proposed to address low-rank and independent-output assumptions. For HybridFlow, proposed extensions include Bayesian flows or flow ensembles, alternative generative models, and online or continual-learning variants. For turbulence UQ, the stated extensions are Bayesian or ensemble CNNs, full eigenspace correction, and direct incorporation of physics-consistency regularizers. Taken together, these directions suggest that future uncertainty-guided hybrid frameworks will likely move toward stronger calibration guarantees, richer decomposition of aleatoric and epistemic components, and tighter integration between uncertainty estimation and adaptive computation (Huttinga et al., 2022, Katwyk et al., 6 Oct 2025, Chu et al., 7 Nov 2025).