Learning with Noisy & Partial Labels
Robust Learning under Incomplete, Evolving, and Annotator-Heterogeneous Label Spaces
Generated automatically from the limitations stated in 13 papers (NeurIPS, EMNLP, ICML), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Learning systems that degrade gracefully instead of silently when the label world shifts, and a diagnostic that tells practitioners whether their annotation protocol or their model is the bottleneck.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build a misspecification stress-test harness: take standard benchmark datasets, systematically inject the three failure modes (delete classes, stream in novel classes, simulate annotators with varying skill and class-selection bias), and measure the error decomposition of representative noisy/partial-label methods as each assumption is violated. Deliverable within a month: a quantified sensitivity curve per method.
- 2
Develop an estimator that treats the effective label set as latent and partially observed: infer which classes are present but unnamed from the structure of annotator disagreements, and validate by measuring recovery of held-out held-back classes on datasets where the full taxonomy is known.
- 3
Extend the estimator to the online setting with an explicit change-point signal for "a new class has appeared," measuring detection delay and false-alarm rate on streaming benchmarks.
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
Annotator disagreement patterns may simply not carry enough signal to distinguish a genuinely novel latent class from ordinary annotator confusion, in which case the estimator recovers nothing beyond what a confusion matrix already shows.
Sub-problems
Learning Subjective Emotion Distributions Without Per-Annotator Identifiers
Current perspectivist and annotator-aware emotion models require disaggregated, multi-annotator raw votes tied to specific annotator identifiers to model subjective disagreement. However, the majority of existing large-scale emotion and sentiment benchmarks only release aggregated summary statistics such as mean intensity scores or majority-vote labels. Consequently, modern perspectivist techniques cannot be trained on or applied to legacy benchmark corpora, leaving supervised models to treat genuine human subjectivity as label noise and cap predictive performance.
Open-World Partial and Noisy Label Learning with Dynamic Class Discovery
Current weakly and partially supervised learning methods strictly presuppose a closed world where all semantic classes are known a priori and label distributions remain static. When applied to real-world streaming data where novel categories continuously emerge without pre-assigned identifiers or prior frequency counts, these models force unseen classes into known categories, corrupting existing representations. As a consequence, weakly supervised systems cannot be deployed in open-world, long-running environments without periodic, expensive complete retraining from scratch on globally curated label taxonomies.
Evidence
Each paper's own statement of the limitation, verbatim.
- KaLOS finds Consensus: A Meta-Algorithm for Evaluating Inter-Annotator Agreement in Complex Vision TasksCVPR 2026
Requires multi-rater metadata, so it cannot provide post-hoc quality assurance on legacy single-annotator datasets without additional annotation
- Suicidal Posts Detection System Incorporating Psychological Risk FactorsACML 2025
Reliance on manual professional annotation for training data which is time-consuming and difficult to scale.
- RD-MCSA: A Multi-Class Sentiment Analysis Approach Integrating In-Context Classification Rationales and DemonstrationsEMNLP 2025
Moderate absolute performance: supervised models trained on tens of thousands of samples often fail to exceed 80% accuracy; annotation subjectivity means different annotators assign different labels to the same sample.
- ComRank: Ranking Loss for Multi-Label Complementary Label LearningNeurIPS 2025
Relies on an instance-dependent assumption p(ȳ|x,Y)=p(ȳ|x), which the authors admit is a tractable approximation and may fail when annotators cannot exclude labels without knowing the full relevant set
- Learning to Complement and to Defer to Multiple UsersECCV 2024
Assumes all users are homogeneous (same skill), which is unrealistic since users have distinct classification accuracies; this avoids combinatorial explosion but limits real-world applicability
- IWBVT: Instance Weighting-based Bias-Variance Trade-off for CrowdsourcingNeurIPS 2024
Relies on the assumption that integrated labels equal true labels after instance weighting; if label integration is poor, the bias-variance trade-off can amplify wrong labels
- Collaborative Refining for Learning from Inaccurate LabelsNeurIPS 2024
Restricted to binary classification: in the multi-class setting, when all annotators are wrong on a disagreement sample, LRD is forced to select one of the incorrect labels, degrading performance.
- Streaming Active Learning with Deep Neural NetworksICML 2023
Assumes the number of label classes is known a priori and cannot handle growing label spaces in evolving environments.
Show all 13 papers
- Predict and Use: Harnessing Predicted Gaze to Improve Multimodal Sarcasm DetectionEMNLP 2023
Inter-annotator agreement is only moderate (Fleiss' kappa 0.41), and the approach is benchmarked on a single dataset, limiting generalizability
- Estimating the Uncertainty in Emotion Attributes using Deep Evidential RegressionACL 2023
Requires raw per-annotator labels; datasets that release only averaged or majority-vote labels (e.g. some as in Poria et al. 2019) cannot be used.
- All Points Matter: Entropy-Regularized Distribution Alignment for Weakly-supervised 3D SegmentationNeurIPS 2023
Assumes the available sparse labels cover all semantic classes present in the data; fails when classes are missing or novel
- Balancing Methods for Multi-label Text Classification with Long-Tailed Class DistributionEMNLP 2021
Requires static knowledge of class frequencies from the training set, making it incompatible with online learning or streaming data.
- Complementary-Label Learning for Arbitrary Losses and ModelsICML 2019
Assumes complementary labels are chosen uniformly at random among non-true classes, failing in settings with annotator class-selection bias.
Nearest existing work
- Collaborative Refining for Learning from Inaccurate LabelsNeurIPS 2024
- Learning From Noisy Labels by Regularized Estimation of Annotator ConfusionCVPR 2019
- FREDIS: A Fusion Framework of Refinement and Disambiguation for Unreliable Partial Label LearningICML 2023
- ULAREF: A Unified Label Refinement Framework for Learning with Inaccurate SupervisionICML 2024
- Error-Bounded Correction of Noisy LabelsICML 2020
- Learning a Deep ConvNet for Multi-Label Classification With Partial LabelsCVPR 2019
- Imprecise Label Learning: A Unified Framework for Learning with Various Imprecise Label ConfigurationsNeurIPS 2024
- IWBVT: Instance Weighting-based Bias-Variance Trade-off for CrowdsourcingNeurIPS 2024
- PNP: Robust Learning From Noisy Labels by Probabilistic Noise PredictionCVPR 2022
- Learning from Noisy Labels via Conditional Distributionally Robust OptimizationNeurIPS 2024
- Unsupervised Label Noise Modeling and Loss CorrectionICML 2019
- Multi-Label Classification With Partial Annotations Using Class-Aware Selective LossCVPR 2022
- Noisy Label Learning with Instance-Dependent Outliers: Identifiability via Crowd WisdomNeurIPS 2024
- Learning with Noisy LabelsNeurIPS 2013
- Detecting annotation noise in automatically labelled dataACL 2017
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
Validation-Free Hyperparameter Selection for Robust Learning Pipelines
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.
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.