Learning with Noisy & Partial Labels
Hyperparameter-Free Generalization: Benchmarking and Predicting Sensitivity of New Hyperparameters Across Tasks
Generated automatically from the limitations stated in 23 papers (CVPR, ACL, ICLR), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
A tuning-free evaluation standard that reveals which methods' gains are real versus artifacts of per-dataset search, and enables principled deployment in validation-scarce settings (OOD, few-shot, cross-domain transfer).
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Build a "tuning sensitivity audit" protocol: for 15–20 methods with known new hyperparameters, re-evaluate each across ≥10 datasets under (a) default/fixed values, (b) per-dataset tuning, and (c) transfer of tuned values across datasets and architectures. Measure the gain attributable to tuning vs. the method itself, producing a public leaderboard of hyperparameter sensitivity. Startable immediately with published code.
- 2
Develop sensitivity predictors: fit simple meta-models over dataset statistics (size, noise level, label quality, domain shift) to predict near-optimal hyperparameter values without validation data; evaluate on the transfer setting from (1) by measuring regret relative to tuned optima.
- 3
Formalize "default robustness" as an evaluation criterion — an interval around each hyperparameter within which performance degrades <1% — and test whether existing methods can be re-tuned once to a robust default that works everywhere, reporting where no such default exists.
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 sensitivity audit may simply confirm what everyone already anecdotally knows ("tuning matters") without yielding transferable predictors, and meta-models may fail because optimal hyperparameters depend on interactions too specific to dataset statistics to generalize.
Sub-problems
Validation-Free Hyperparameter Estimation for Model Merging and Modular Adaptation
Current parameter-efficient adaptation and model merging methods rely on tuning structural hyperparameters—such as scaling coefficients ($\lambda, \alpha$), rank thresholds ($k$), and expert counts ($K$)—using held-out labeled validation data. In zero-shot deployment, out-of-distribution transfer, and decentralized adaptation, such labeled validation splits do not exist. Practitioners are consequently forced to rely on arbitrary default hyperparameters, which frequently leads to catastrophic interference, sub-optimal skill combination, or severe overfitting.
Self-Regulating Auxiliary Hyperparameters for Sample-Constrained Bayesian Optimization
Bayesian optimization is designed for sample-efficient optimization of expensive black-box functions where evaluations are strictly budget-constrained. When advanced BO methods introduce internal auxiliary hyperparameters (such as decay rates, coupling weights, and softmax temperatures) that require manual tuning or grid search, the meta-tuning cost directly defeats the premise of sample-efficient optimization. In single-run or physical experimental campaigns, practitioners cannot afford pilot trials to calibrate the optimizer itself. Consequently, advanced BO formulations remain fragile or unusable outside the specific benchmark suites on which their internal parameters were tuned.
Target-Free Hyperparameter Selection for Few-Shot and Zero-Shot Transfer
Few-shot and zero-shot algorithms are highly sensitive to hyperparameters like activation thresholds, temperature scalings, and projection dimensions, but standard workflows tune these against labeled target-domain validation sets. In genuine zero-shot and extreme few-shot deployments, target validation labels are unavailable by definition, forcing practitioners to use arbitrary literature defaults that severely degrade downstream performance. This creates a structural evaluation gap where reported benchmarks implicitly rely on label leakage to achieve reported performance.
Evidence
Each paper's own statement of the limitation, verbatim.
- 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.
- 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
- Neuron-Aware Active Few-Shot Learning for LLMsACL 2026
Hyperparameters K (top-k activated neurons) and τ are tuned per model-dataset pair (e.g., K ranges from 500 to 8000 across settings), suggesting sensitivity and the need for tuning budget
- Incorporating Expert Priors into Bayesian Optimization via Dynamic Mean DecayICLR 2026
Introduces additional hyperparameters that require configuration: prior mean scaling (A), offset (B), decay rate (λ), and initialization sampling ratio (ρ).
- Discrete Compositional Generation via General Soft Operators and Robust Reinforcement LearningICLR 2026
Introduces additional hyperparameters (interpolation parameter q, softmax temperature alpha, and regularizer scale omega) requiring grid search over domain-specific settings.
- Counterfactual Credit Guided Bayesian OptimizationAISTATS 2026
Introduces multiple tunable hyperparameters (credit weight lambda, sensitivity tau, half-life decay M, MC samples K, neighbor count H) that must be set relative to the evaluation budget.
- Optimizing the Unknown: Black Box Bayesian Optimization with Energy-Based Model and Reinforcement LearningNeurIPS 2025
Requires tuning of coupling hyperparameters (λ, γ, α); the scale mismatch between f and E_θ can cause failures unless normalization and adaptive λ are used
- 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
Show all 23 papers
- Few-Shot Recognition via Stage-Wise Retrieval-Augmented FinetuningCVPR 2025
Hyperparameters are fixed literature defaults and the authors acknowledge they are not necessarily optimal per task; no per-task tuning mechanism exists yet
- 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
- 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
- Project and Probe: Sample-Efficient Adaptation by Interpolating Orthogonal FeaturesICLR 2024
Hyperparameter tuning (including d) uses a target validation set; selecting hyperparameters without domain-specific target data remains unsolved
- 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
- Towards Effective Visual Representations for Partial-Label LearningCVPR 2023
Requires hand-tuned hyperparameters (temperature, moving-average factors, mixup Beta parameter, dynamic loss weighting schedule).
- 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
- 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
- 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
- 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
- MetaPerturb: Transferable Regularizer for Heterogeneous Tasks and ArchitecturesNeurIPS 2020
- Quick-Tune: Quickly Learning Which Pretrained Model to Finetune and HowICLR 2024
- Improved Regularization and Robustness for Fine-tuning in Neural NetworksNeurIPS 2021
- Meta-learning Hyperparameter Performance Prediction with Neural ProcessesICML 2021
- Efficient Transfer Learning Method for Automatic Hyperparameter TuningAISTATS 2014
- Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly RetrainingNeurIPS 2025
- Robust Fine-Tuning of Deep Neural Networks with Hessian-based Generalization GuaranteesICML 2022
- Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-SchedulingICLR 2026
- Hyperparameter Learning via Distributional TransferNeurIPS 2019
- Co-Tuning for Transfer LearningNeurIPS 2020
- PATS: Sensitivity-aware Noisy Learning for Pretrained Language ModelsEMNLP 2022
- Towards a Unified View of Parameter-Efficient Transfer LearningICLR 2022
- Holistic Transfer: Towards Non-Disruptive Fine-Tuning with Partial Target DataNeurIPS 2023
- HyperSTAR: Task-Aware Hyperparameters for Deep NetworksCVPR 2020
- Efficiently Tuned Parameters Are Task EmbeddingsEMNLP 2022
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
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.