Research map / Label-efficient and robust learning / OOD and anomaly detection
Robust fine-tuning for OOD generalization
Fine-tunes vision-language models while preserving pretrained robustness using energy and calibration objectives.
56 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in OOD and anomaly detection, keyword cluster “ood detection · distribution ood · distribution detection”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
20 papers in 2023–24 and 27 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 9.4% to 10.4%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2020 · Likelihood Ratios and Generative Classifiers for Unsupervised Out-of-Domain Detection in Task Oriented Dialog (AAAI 2020 · 40 citations)
- 2021 · MASKER: Masked Keyword Regularization for Reliable Text Classification (AAAI 2021 · 28 citations)
- 2022 · Estimating Example Difficulty Using Variance of Gradients (CVPR 2022 · 59 citations)
- 2023 · Masked Images Are Counterfactual Samples for Robust Fine-Tuning (CVPR 2023 · 16 citations)
- 2024 · Debiasing Multimodal Sarcasm Detection with Contrastive Learning (AAAI 2024 · 35 citations)
- 2025 · On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · 9 citations)
- 2026 · Cross-modal Proxy Evolving for OOD Detection with Vision-Language Models (AAAI 2026 · 17 citations)
Earliest papers here
- Likelihood Ratios and Generative Classifiers for Unsupervised Out-of-Domain Detection in Task Oriented Dialog (AAAI 2020 · 40 citations)
- Towards Maximizing the Representation Gap between In-Domain & Out-of-Distribution Examples (NeurIPS 2020 · 19 citations)
- MASKER: Masked Keyword Regularization for Reliable Text Classification (AAAI 2021 · 28 citations)
What has been tried
44 papers here report 464 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 6%, hurt 23%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add image modality | MSD IID test · accuracy | 87.67 -> 88.95 | Helped | Debiasing Multimodal Sarcasm Detection with Contrastive Learning (AAAI 2024 · 35 citations) |
| Add decoupled background OOD augmentation | Four traditional OOD datasets (Avg) · AUROC | 92.97 -> 94.65 | Helped | Overcoming Shortcut Problem in VLM for Robust Out-of-Distribution Detection (CVPR 2025 · 8 citations) |
| add basic digital augmentations | ImageNet-C · accuracy | 51.0 -> 51.7 | Helped | Unexplored Faces of Robustness and Out-of-Distribution: Covariate Shifts in Environment and Sensor Domains (CVPR 2024 · 3 citations) |
| Remove highly similar LAION samples to match ImageNet-Train similarity for ImageNet-Sketch | ImageNet-Sketch · top-1 accuracy | 50.21 -> 44.78 | Hurt | Does CLIP’s generalization performance mainly stem from high train-test similarity? (ICLR 2024 · 3 citations) |
| Remove optimization on generated OOD image features (L_c) | Setup-I, 32-shot · Avg OOD accuracy | 62.24 -> 63.07 | Hurt | CRoFT: Robust Fine-Tuning with Concurrent Optimization for OOD Generalization and Open-Set OOD Detection (ICML 2024 · 3 citations) |
| add cross-domain estimation tuning term (L*crd) to SupCon feature-tuning | YellowR1, λ=0.90 · RMSall | 1.3697 -> 0.3085 | Helped | Cross-Rejective Open-Set SAR Image Registration (CVPR 2025 · 2 citations) |
| Reduce the number of sampled frequency subsets from CF-1 to CF-2.10 | CIFAR-10 · computation time | ∼200× reduction | Mixed | Do ImageNet-trained Models Learn Shortcuts? The Impact of Frequency Shortcuts on Generalization (CVPR 2025 · 2 citations) |
| Add orthogonality constraint to MCL + EMA SD | Avg. shifts · accuracy | 62.28 -> 62.55 | Helped | Towards Calibrated Robust Fine-Tuning of Vision-Language Models (NeurIPS 2024 · 2 citations) |
| Use sliding-window selection rather than reservoir sampling | Synthetic · average reward | Helped | AdaO2B: Adaptive Online to Batch Conversion for Out-of-Distribution Generalization (AAAI 2025 · 1 citation) | |
| Use feature split f=0.5 instead of f=1.0 | Threshold Classifier pairing (m=0.25) · detection rate | 0.58 -> 1.0 | Helped | Identifying Predictions That Influence the Future: Detecting Performative Concept Drift in Data Streams (AAAI 2025 · 1 citation) |
| add class-conditional feature generator (no self-distillation) | Base-to-new ablation, 11 datasets · New accuracy | 63.22 -> 69.02 | Helped | Overcoming the Pitfalls of Vision-Language Model Finetuning for OOD Generalization (ICLR 2024 · 1 citation) |
| Replace the fully educated DFS scratchpad with an inductive scratchpad | Cycle task OOD · OOD generalization | Helped | How Far Can Transformers Reason? The Globality Barrier and Inductive Scratchpad (NeurIPS 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 19 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet-1k · 4 of 19 papers
- CIFAR-10 · 4 of 19 papers
- ImageNet-R · 3 of 19 papers
- iNaturalist · 3 of 19 papers
- Texture · 3 of 19 papers
- CelebA · 3 of 19 papers
- ImageNet-v2 · 2 of 19 papers
- ImageNet-C · 2 of 19 papers
Baselines
- DNN · 2 of 19 papers
- MCM · 2 of 19 papers
- NegLabel · 2 of 19 papers
- CoOp · 2 of 19 papers
- LoCoOp · 2 of 19 papers
- P-CBM · 2 of 19 papers
- DetectGPT · 2 of 19 papers
- Fast-DetectGPT · 2 of 19 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 47 papers; from OpenAlex and the papers' affiliations).
- Apple · 3 papers
- Yonsei University · 3 papers
- Shandong University · 2 papers
- Swiss Federal Institute of Technology Lausanne · 2 papers
- EPFL · 2 papers
- University of California, Berkeley · 2 papers
- Lawrence Livermore National Labs · 2 papers
- Shanghai Jiao Tong University · 2 papers
- Carnegie Mellon University · 2 papers
- Harvard University · 2 papers
Use it
Download BibTeX of 33 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Provable and feature-based OOD detection (122 papers)
- Spurious-aware OOD detection (109 papers)
- Out-of-distribution generalization theory (71 papers)
- Uncertainty and misclassification detection (56 papers)
Elsewhere in the topic
- Unified multi-class anomaly detection (86 papers)
- Anomaly detection with contaminated data (64 papers)
- Deep one-class and tabular anomaly detection (50 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).