Domain Adaptation & Generalization
Domain-Invariant Learning Without Domain Labels or Target-Domain Data at Training Time
Generated automatically from the limitations stated in 10 papers (NeurIPS, ICCV, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current methods across these settings all condition on information that disappears in deployment: explicit domain labels for training samples, knowledge of the domain count, or a pool of unlabeled target-domain data available during training. When source domain identity is unknown, only one source is available, or target data cannot be collected before the model is fixed, these methods are simply unusable — not merely degraded. This blocks the realistic scenario where a single labeled dataset is all one has, and the test distribution is only discovered at inference time.
Why it matters
Domain generalization becomes applicable when domain identity is unannotated, the domain count is unknown, or only one labeled source exists — the common case in practice rather than the benchmark case.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Infer domain structure automatically: cluster training samples using model-internal disagreement or feature-space density to produce pseudo-domain partitions, then run domain-invariant learning over the discovered partitions with no explicit labels. Measure whether discovered partitions recover ground-truth domains (adjusted mutual information) and whether invariance over inferred partitions matches invariance over true labels on standard benchmarks, including single-source splits.
- 2
Train a domain-label-free objective that requires no per-domain decoders or domain count — e.g., invariance enforced via sample-level reweighting or contrasts that do not scale with the number of partitions — and measure the point at which performance degrades as the number of latent domains grows from 1 to 50.
- 3
Test all of the above under a strict protocol: no unlabeled target data, no domain labels, single or unknown source count, reporting the gap to methods that are granted each privilege, to quantify exactly which piece of supervision buys how much.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
The inferred partitions may be too noisy or orthogonal to true domain structure, so invariance over them buys nothing beyond a plain single-model baseline — in which case domain labels turn out to carry information that cannot be recovered from the data alone.
Sub-problems
Few-Shot and Zero-Shot Learning Under Extreme Unlabeled Target Data Scarcity and Mismatch
Current semi-supervised and transductive few-shot methods depend on access to thousands of unlabeled in-domain target samples during training or adaptation. In practical deployment scenarios, such as rare disease diagnostics or proprietary edge deployments, collecting even moderate pools of unlabeled target data is impossible or violates privacy constraints. Consequently, existing models degrade unpredictably or fail entirely when evaluated in truly data-starved or severely mismatched target regimes.
Evidence
Each paper's own statement of the limitation, verbatim.
- Gradient-Guided Annealing for Domain GeneralizationCVPR 2025
Requires domain labels at training time, so it is inapplicable to single-source DG
- Generalized Category Discovery under Domain Shift: A Frequency Domain PerspectiveNeurIPS 2025
Requires access to unlabeled data from the unknown domain during training, so it does not apply to domain generalization settings
- Approximate Domain Unlearning for Vision-Language ModelsNeurIPS 2025
Assumes domain labels are available for all training samples; missing labels are only addressed via simple pseudo-labeling experiments in an appendix.
- A Generalized Label Shift Perspective for Cross-Domain Gaze EstimationNeurIPS 2025
Requires unlabeled target domain samples during training, so it does not generalize to unseen domains without target data
- LEDetection: A Simple Framework for Semi-Supervised Few-Shot Object DetectionAISTATS 2024
Requires a large pool of unlabeled images; evaluated only with COCO-unlabeled2017, and the VOC experiments already use out-of-domain COCO images, so behavior when unlabeled data is unavailable or highly mismatched is unclear
- Towards Unsupervised Domain Generalization for Face Anti-SpoofingICCV 2023
Requires unlabeled data from multiple source domains during pretraining and identity-wise data splits, which may not be available in all deployment settings
- Selective Annotation Makes Language Models Better Few-Shot LearnersICLR 2023
Requires a pool of unlabeled in-domain data to select from; evaluation relies on subsampled pools of at least 3K instances, and behavior with very small unlabeled pools is untested
- DomainDrop: Suppressing Domain-Sensitive Channels for Domain GeneralizationICCV 2023
Requires domain labels for source training samples, so it does not apply when source domain identity is unknown
Show all 10 papers
- Cycle-consistent Masked AutoEncoder for Unsupervised Domain GeneralizationICLR 2023
Requires knowing the number of domains in advance, since one domain-specific decoder per training domain must be trained; scaling to many domains is unclear
- CDFSL-V: Cross-Domain Few-Shot Learning for VideosICCV 2023
Requires access to the unlabeled training split of the target domain during training, which may not be available in practice
Nearest existing work
- Domain Generalization via Entropy RegularizationNeurIPS 2020
- Adaptive Methods for Real-World Domain GeneralizationCVPR 2021
- Learning to Generalize across Domains on Single Test SamplesICLR 2022
- Exploiting Domain-Specific Features to Enhance Domain GeneralizationNeurIPS 2021
- Domain Invariant Representation Learning with Domain Density TransformationsNeurIPS 2021
- Domain Generalization by Learning and Removing Domain-specific FeaturesNeurIPS 2022
- Semi-Supervised Domain Generalization with Known and Unknown ClassesNeurIPS 2023
- Domain Generalization with Interpolation RobustnessACML 2023
- On Learning Invariant Representations for Domain AdaptationICML 2019
- Generalization on Unseen Domains via Inference-Time Label-Preserving Target ProjectionsCVPR 2021
- Learning Invariant Representations and Risks for Semi-Supervised Domain AdaptationCVPR 2021
- Learning to Balance Specificity and Invariance for In and Out of Domain GeneralizationECCV 2020
- Balancing Discriminability and Transferability for Source-Free Domain AdaptationICML 2022
- Domain Separation NetworksNeurIPS 2016
- Partial Adversarial Domain AdaptationECCV 2018
Related open problems
Domain Adaptation & Generalization
Cross-Architecture Robustness and Evaluation of Domain Adaptation in Object Detection
Current domain adaptation and domain generalization methods for object detection are almost exclusively evaluated on two-stage anchor-based architectures, primarily Faster R-CNN with ResNet backbones. Because these adaptation mechanisms often couple directly to region proposal networks, RoI pooling representations, or batch normalization layers, it is completely unknown whether reported domain adaptation gains transfer to one-stage detectors (e.g., FCOS, YOLO) or query-based transformers (e.g., DETR variants). Consequently, practitioners deploying modern transformer or anchor-free architectures have no empirical guidance on whether established domain adaptation methods function outside Faster R-CNN.
Domain Adaptation & Generalization
Systematic Evaluation of Data Augmentation Techniques Across Disparate Distribution Shift Types
Data augmentation methods developed for out-of-distribution robustness are overwhelmingly hand-engineered to simulate surface-level texture and appearance variations (e.g., color perturbations and synthetic corruptions). When these methods are deployed under other distribution shift types—such as subpopulation shifts, geometric transformations, or domain shifts in specialized fields like medical imaging—they frequently fail or degrade accuracy below standard Empirical Risk Minimization (ERM). Because augmentation techniques have not been systematically evaluated across these differing shift categories, practitioners have no reliable way to anticipate when a given augmentation strategy will harm downstream generalization.
Domain Adaptation & Generalization
Cold-Start Streaming Domain Adaptation and Discovery Without Labeled Base Sessions or Offline Pre-Passes
Current continual domain adaptation and streaming open-world discovery methods structurally depend on an offline initialization phase—either requiring a fully labeled source base dataset or an offline full-dataset inference pass to compute initial prototypes and representations. In truly dynamic environments where data arrives purely as an online stream and domain shifts occur from the first observation, collecting labeled base data or executing multi-pass offline initialization is impossible. Consequently, existing frameworks cannot be deployed in pure cold-start streaming regimes.