Learning with Noisy & Partial Labels
Validation-Free Hyperparameter Selection for Robust Learning Pipelines
Generated automatically from the limitations stated in 49 papers (CVPR, ICML, NeurIPS), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Across these works, the reported gains hinge on hyperparameters whose optima shift per dataset, noise level, class prior, or architecture, and which are chosen by grid search on held-out clean validation data or expensive ablation. This blocks plug-and-play deployment: a practitioner cannot know whether a method's claimed advantage survives at untuned settings, and comparisons between methods are confounded by unequal tuning budgets. Until hyperparameter settings can be selected from the training data itself (or shown to be inert), robust-learning methods remain fragile recipes rather than tools.
Why it matters
Methods whose hyperparameters auto-adapt to noise level, class prior, and dataset become deployable without validation data, and reported gains become interpretable as robust improvements rather than tuning artifacts.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Hyperparameter sensitivity audits at scale. Re-run a battery of published robust-learning methods across their reported hyperparameter grids on standard benchmarks, measuring the performance variance attributable to tuning versus the claimed improvement over baselines. Deliverable: a public sensitivity leaderboard quantifying which methods' gains exceed their tuning sensitivity.
- 2
Self-tuning criteria from training dynamics. Develop and evaluate estimators (e.g., based on label-noise level estimation, per-sample confidence distributions, or loss-curvature signals) that set key hyperparameters (thresholds, loss weights, warm-up lengths) from observable training statistics alone. Measured by: gap between auto-selected and oracle grid-searched performance across datasets and noise regimes.
- 3
Robust-by-construction hyperparameter intervals. For methods with narrow operating ranges (e.g., candidate-set thresholds), reformulate the objective so performance degrades gracefully outside the tuned range, and measure worst-case performance across the reported grid rather than best-case.
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
Auto-tuning signals may themselves need dataset-specific calibration, recreating the problem one level up; and worst-case-over-grid evaluation may simply show that every method is fragile, yielding a negative result without a constructive path forward.
Sub-problems
Benchmarking and Resolving Hyperparameter Sensitivity in Target-Free Domain Adaptation
Current domain adaptation and generalization algorithms introduce numerous sensitive loss weights, filtering thresholds, and momentum parameters that are routinely tuned using target-domain validation labels. In realistic deployment, target labels are inaccessible, making this selection protocol a violation of the problem setting. Consequently, published performance gains frequently collapse when hyperparameters cannot be oracle-tuned on the unseen target distribution.
Validation-Free Hyperparameter Selection for Parameter-Efficient Model Merging
Current parameter-efficient adaptation and model merging methods introduce structural hyperparameters—such as scaling coefficients ($\alpha, \beta, \lambda$) and component counts ($n, k, K$)—that must be tuned against target validation sets. In zero-shot transfer, out-of-distribution deployment, and cross-task model merging, labeled target validation data is absent by definition, rendering these methods unusable or sensitive to arbitrary defaults. Without validation-free selection criteria, the practical utility of modular model combination remains dependent on the very target-task supervision it seeks to avoid.
Evidence
Each paper's own statement of the limitation, verbatim.
- Towards Multimodal Domain Generalization with Few LabelsCVPR 2026
Relies on multiple sensitive hyperparameter thresholds (e.g., confidence threshold tau, GCE noise parameter q, prototype EMA momentum) and additional translator networks.
- SOS-LoRA: Static Orthogonal-Subspace Low-Rank Adaptation with Fixed Multi-Scale ScalingACL 2026
Introduction of three additional hyperparameters (expert count K, scaling range gamma, and regularization weight lambda) that may require tuning.
- Robust Self-reflective Hashing for Cross-modal Retrieval with Noisy LabelICML 2026
Hyperparameters γ and τ are sensitive — performance degrades when either exceeds a threshold, requiring tuning per dataset
- Revisiting Learning with Noisy Labels: Active Forgetting and Noise SuppressionCVPR 2026
Requires warm-up training and tuned loss weights (beta=0.001, gamma=0.1); hyperparameter beta in particular shows some sensitivity in their sensitivity analysis
- Prior-free Tabular Test-time AdaptationICLR 2026
Requires tuning multiple sensitive hyperparameters online, including entropy thresholding factor zeta, loss weights beta1 and beta2, and subspace dimension m.
- Positive-Unlabeled Learning with Extreme Scarcity of Labeled PositivesICML 2026
Requires three hyperparameters (λ, γ, β) tuned by cross-validation, with degradation for λ>10²
- Partial Label Learning-Inspired Denoising Implicit Feedback for RecommendationSIGIR 2026
Requires tuning dataset-dependent hyperparameters, notably the initial noise prior η, which must be raised to ~0.8 for ML-100K's higher noise ratio
- Neural Collapse in Test-Time AdaptationCVPR 2026
Introduces multiple hyperparameters (weight balance alpha, top-k candidate count, entropy thresholds) that require tuning.
Show all 49 papers
- Dynamic Label Noise Suppression with Optimal Teacher Pool for Facial Expression RecognitionCVPR 2026
Gains over strongest baselines are small (roughly 0.6-1.5 points at most noise levels), within a range that could be sensitive to hyperparameter choices (pool size, forced-update period, confidence decay, smoothing bounds all tuned)
- Disambiguation and Structured Sparsity CLASS: Deep Partial Label Feature Selection with Cluster-GuidedKDD 2026
Hyperparameter sensitivity: results depend notably on lambda (sparsity) and M (gate ratio), which require grid search over {0.01–100}
- Class-Conditional Distribution Balancing for Group Robust ClassificationICML 2026
Hyperparameters (split ratio γ, clipping threshold T) are selected via empirical ablation and T depends on the number of classes; robust annotation-free tuning strategies remain future work
- SEGA: Shaping Semantic Geometry for Robust Hashing under Noisy SupervisionNeurIPS 2025
Requires hyperparameter tuning of the percentile threshold q_r, which performs best only at moderate values (0.3–0.5); too high over-filters clean samples, too low leaves noise unfiltered
- RankMatch: A Novel Approach to Semi-Supervised Label Distribution Learning Leveraging Rank Correlation between LabelsNeurIPS 2025
Performance depends on hyperparameter λ: drops noticeably at λ=0.005 (especially Emotion6) and degrades slightly at λ=0.1, requiring tuning in the 0.01–0.05 range
- Point-to-Region Loss for Semi-Supervised Point-Based Crowd CountingCVPR 2025
Performance depends on hyperparameters τ, η (best 0.7), and the unlabeled-loss ratio α; a too-large α causes training failure
- Noise Separation guided Candidate Label Reconstruction for Noisy Partial Label LearningICLR 2025
Hyper-parameters (λ, β, warm-up epoch e_w, end epoch e_end) are tuned on a held-out clean validation set split from training data, which may not be available in practice
- Minimal Interaction Seperated Tuning: A New Paradigm for Visual AdaptationCVPR 2025
The hyperparameter k must be selected per task from a validation set, adding tuning burden and requiring held-out data
- Image Retrieval Learning Resistant Binary Descriptors Against Noise for EfficientSIGIR 2025
Requires tuning thresholds tau_D and tau_V and hyperparameters lambda and tau, with performance degrading above tau=0.15
- Heterogeneous Label Shift: Theory and AlgorithmICML 2025
Performance is sensitive to hyperparameters α and β, with best results confined to {0.5, 1} in the authors' sensitivity analysis
- Diverse Prototypical Ensembles Improve Robustness to Subpopulation ShiftICML 2025
Adds ensemble-related hyperparameters (temperature τ, IPS weight α, number of prototypes) and ~2 minutes training per prototype
- D2SP: Dynamic Dual-Stage Purification Framework for Dual Noise Mitigation in Vision-based Affective Recognition.CVPR 2025
Hyperparameters (pruning threshold λ, correction threshold τ, stability window t) must be tuned per dataset (e.g., λ=0.7 for FERV39k but 0.6 for DFEW; t=3 vs 2), reducing plug-and-play generality
- Curriculum Debiasing: Toward Robust Parameter-Efficient Fine-Tuning Against Dataset BiasesACL 2025
Introduces hyperparameters (alpha, beta) whose tuning normally requires validation data, which is often unavailable in OOD settings, though the authors report reduced sensitivity
- Confidence Difference Reflects Various Supervised Signals in Confidence-Difference ClassificationICML 2025
Requires tuning two extra hyperparameters (threshold θ, regularization weight α) whose optima shift with class prior (θ=0.4 for π=0.5 vs θ=0.2-0.3 for π=0.2; α sensitive on CIFAR-10)
- Revive Re-weighting in Imbalanced Learning by Density Ratio EstimationNeurIPS 2024
Requires a warm-up/temperature hyperparameter (momentum coefficient m affects accuracy, Fig. 3) that must be tuned; validation is limited to image classification long-tailed benchmarks
- Rethinking Weight Decay for Robust Fine-Tuning of Foundation ModelsNeurIPS 2024
λ still requires cross-validation (values used span 0.5 to 2.2 across tasks), so hyper-parameter tuning is not fully eliminated
- MELoRA: Mini-Ensemble Low-Rank Adapters for Parameter-Efficient Fine-TuningACL 2024
Introduces a new hyperparameter n whose optimum varies by dataset and task (n=2 optimal on GLUE, n=4-8 on INSTRUCTEVAL), requiring extra tuning; the authors note performance drops for overly large n due to overfitting
- Learning from Noisy Labels via Conditional Distributionally Robust OptimizationNeurIPS 2024
Requires warm-up with majority-vote labels and hyperparameter ϵ must be tuned within a narrow range (0, 1/K)
- Improving 3D Semi-supervised Learning by Effectively Utilizing All Unlabelled DataECCV 2024
Performance is highly sensitive to hyperparameters: at learning rate 0.0005 accuracy collapses to 12%, and loss weight factors change accuracy by several points
- HINTED: Hard Instance Enhanced Detector with Mixed-Density Feature Fusion for Sparsely-Supervised 3D Object DetectionCVPR 2024
Ablation of the loss weights is highly sensitive: e.g., γ1=0 gives 63.1 Avg vs 72.9 at γ1=0.1, indicating performance depends on careful hyperparameter tuning
- DuPL: Dual Student with Trustworthy Progressive Learning for Robust Weakly Supervised Semantic SegmentationCVPR 2024
Hyperparameter sensitivity: performance degrades when the final threshold τh(T) drops below 0.55 (excess noise) and ANF warm-up must be tuned (best at 8000 iterations; too early filters correct labels, too late after noise memorization)
- Candidate Pseudolabel Learning: Enhancing Vision-Language Models by Prompt Tuning with Unlabeled DataICML 2024
Hyperparameter sensitivity: when β exceeds 0.99 many instances get empty candidate sets and are discarded, reducing utilization of unlabeled data; best operating range is narrow (α 0.45–0.75, β 0.95–0.97)
- ActiveDC: Distribution Calibration for Active FinetuningCVPR 2024
Hyperparameters (α per dataset, λ, ξ, number of generated features) are dataset-specific and tuned empirically; performance drops sharply when mis-specified (e.g., ξ=0.3 drops CIFAR10 to 70.2 from 87.3; 4× generation drops CIFAR100 to 57.1)
- Two-Way Multi-Label LossCVPR 2023
Hyperparameter T_P=4 was tuned on MS-COCO; results with T_P→0 (hard min) degrade sharply, so the method depends on soft-margin temperature choice
- Towards Effective Visual Representations for Partial-Label LearningCVPR 2023
Requires hand-tuned hyperparameters (temperature, moving-average factors, mixup Beta parameter, dynamic loss weighting schedule).
- Simple and Fast Group Robustness by Automatic Feature ReweightingICML 2023
Still requires a group-annotated validation set for hyperparameter tuning and early stopping, and more groups demand more such labels
- RoPAWS: Robust Semi-supervised Representation Learning from Uncurated DataICLR 2023
Effectiveness depends on three hyperparameters (r, tau_prior, k) set empirically per architecture; robustness of these choices across settings is only explored in appendices
- Revisiting Domain-Adaptive 3D Object Detection by Reliable, Diverse and Class-balanced Pseudo-LabelingICCV 2023
Hyperparameters (δ_cde=0.6, d=5, sampling schedule Sr/Sg) were hand-tuned and only sensitivity-tested on limited ranges across the three tasks
- Model Calibration in Dense Classification with Adaptive Label PerturbationICCV 2023
Requires a held-out validation set and retraining each epoch with regulariser computations; hyperparameters η and λ need tuning (λ outside 500–10,000 or η outside ~0.0002–0.005 degrade or destabilise results)
- Mitigating Memorization of Noisy Labels by Clipping the Model PredictionICML 2023
Hyperparameter sensitivity: performance varies substantially with 1/τ, and the clipping may affect gradients for samples whose logits exceed the threshold
- MADG: Margin-based Adversarial Learning for Domain GeneralizationNeurIPS 2023
Uses test-domain validation for hyperparameter selection, which accesses target-like data; the margin ρ must be tuned (ρ̂=1.5 best among tested values)
- Learning in Imperfect Environment: Multi-Label Classification with Long-Tailed Distribution and Partial LabelsICCV 2023
Requires multiple hyperparameters (alpha, tau, loss weights, learning-rate decays for three parallel models) whose optimal values are tuned per setting; ablations show performance decays either side of alpha~2 and tau~0.7
- Label Information Bottleneck for Label EnhancementCVPR 2023
Performance depends on hyperparameters alpha and beta tuned over a grid ({0.001,...,10}) per dataset
- LA-Net: Landmark-Aware Learning for Reliable Facial Expression Recognition under Label NoiseICCV 2023
Performance is sensitive to hyperparameters: optimal K shifts between clean (K=12) and noisy (K=8) data, and confidence threshold δ, decay ω, and loss weights require grid search
- Holistic Label Correction for Noisy Multi-Label ClassificationICCV 2023
The threshold δ̂ must be pre-specified (searched in {0.25–0.45} using 10% of noisy training data as validation), introducing a hyperparameter sensitivity not validated outside this range
- Harnessing Hard Mixed Samples with Decoupled RegularizerNeurIPS 2023
Introduces hyperparameters (η, and ξ,t for the BCE variant) whose optimal values depend on mixup type and architecture, requiring tuning on new tasks
- Fuzzy Positive Learning for Semi-Supervised Semantic SegmentationCVPR 2023
Performance depends on tuning the single hyperparameter T, which must be set negatively correlated with the amount of labeled data (e.g., T=0.95 best at 1/32 labels, T=0.85-0.9 best at 1/8)
- Ambiguity-Resistant Semi-Supervised Learning for Dense Object DetectionCVPR 2023
Hyperparameters (τ_neg=0.1, dynamic τ_pos, similarity thresholds) are tuned empirically; sensitivity analysis is limited
- Active Negative Loss Functions for Learning with Noisy LabelsNeurIPS 2023
Requires tuning α, β, and regularization coefficient δ, with performance sensitive to these choices
- Balancing Methods for Multi-label Text Classification with Long-Tailed Class DistributionEMNLP 2021
Introduces five distinct hyperparameters (alpha, beta, gamma, lambda, kappa) that may require significant tuning across different domains.
- A Partial Label Metric Learning Algorithm for Class Imbalanced DataACML 2021
Requires dataset-specific hyperparameter tuning for the neighborhood size k, regularization strength lambda, and LDD constant c.
Nearest existing work
- Improved Regularization and Robustness for Fine-tuning in Neural NetworksNeurIPS 2021
- Instance-adaptive training with noise-robust losses against noisy labelsEMNLP 2021
- PATS: Sensitivity-aware Noisy Learning for Pretrained Language ModelsEMNLP 2022
- Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-SchedulingICLR 2026
- Why Is Prompt Tuning for Vision-Language Models Robust to Noisy Labels?ICCV 2023
- Noise against noise: stochastic label noise helps combat inherent label noiseICLR 2021
- Robust training with ensemble consensusICLR 2020
- Hyperparameter Sensitivity in Deep Outlier Detection: Analysis and a Scalable Hyper-Ensemble SolutionNeurIPS 2022
- Understanding and Utilizing Deep Neural Networks Trained with Noisy LabelsICML 2019
- Distilling Effective Supervision From Severe Label NoiseCVPR 2020
- Robust Training under Label Noise by Over-parameterizationICML 2022
- Intrinsic Gradient Suppression for Label-Noise Prompt Tuning in Vision–Language ModelsICML 2026
- Coresets for Robust Training of Deep Neural Networks against Noisy LabelsNeurIPS 2020
- Label Noise in Adversarial Training: A Novel Perspective to Study Robust OverfittingNeurIPS 2022
- Scalable Gradient-Based Tuning of Continuous Regularization HyperparametersICML 2016
Related open problems
Learning with Noisy & Partial Labels
Eliminating the Auxiliary Clean-Label Anchor Required to Bootstrap Training under Imperfect Supervision
Across weak-supervision methods — noisy labels, partial labels, single positives, model merging, node classification — the same structural dependency recurs: the method only works if some external supervised anchor exists, whether a verified clean reference set, a small labeled validation split, class-frequency priors from a labeled pool, or a warm-up model trained on trustworthy data. In the regime where imperfect labeling actually occurs (annotation too expensive or impossible), that anchor is exactly what is missing, so these methods cannot be deployed there at all. The result is a circular dependency: robust learning from weak labels presupposes the labels one is trying to avoid collecting. Today there is no principled way to replace the anchor with signals derivable from the weak data itself.
Learning with Noisy & Partial Labels
Hyperparameter-Free Generalization: Benchmarking and Predicting Sensitivity of New Hyperparameters Across Tasks
Nearly every new method in this space ships with two to five new hyperparameters whose optimal values are dataset- or architecture-dependent, and the reported results are only achievable after per-task grid search on held-out validation data. This inflates effective cost, undermines plug-and-play claims, and silently biases the literature toward methods whose authors had tuning budget. Worse, when validation data is scarce (OOD, few-shot, cross-domain), the reported optima may not exist at transfer time at all — yet no one measures how much of a method's claimed gain survives under zero-tuning conditions.
Learning with Noisy & Partial Labels
Robust Learning under Incomplete, Evolving, and Annotator-Heterogeneous Label Spaces
Nearly every label-learning method in circulation assumes a closed world: the class set is known a priori, fixed across training and deployment, fully covered by the annotation, and interpreted identically by every annotator. When any of these fail — a class missing from the taxonomy, new classes arriving in a stream, annotators with different accuracies or selection biases, or legacy datasets with only collapsed majority labels — the methods either fail silently or degrade without diagnostic. Today there is no principled way to even measure how much of a model's error comes from label-space misspecification versus model capacity, so practitioners cannot tell whether to fix the taxonomy, re-annotate, or retrain.