mlsanityData to 7 Oct 2026

Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distribution

arXiv 2202.10054 ↗First posted 21 Feb 2022cs.LG cs.CVcited by 722

Not on the map. Why?

Cited by 70 in Multimodal LLMs / vision-language models, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

When transferring a pretrained model to a downstream task, two popular methods are full fine-tuning (updating all the model parameters) and linear probing (updating only the last linear layer -- the "head"). It is well known that fine-tuning leads to better accuracy in-distribution (ID). However, in this paper, we find that fine-tuning can achieve worse accuracy than linear probing out-of-distribution (OOD) when the pretrained features are good and the distribution shift is large. On 10 distribution shift datasets (Breeds-Living17, Breeds-Entity30, DomainNet, CIFAR → STL, CIFAR10.1, FMoW, ImageNetV2, ImageNet-R, ImageNet-A, ImageNet-Sketch), fine-tuning obtains on average 2% higher accuracy ID but 7% lower accuracy OOD than linear probing. We show theoretically that this tradeoff between ID and OOD accuracy arises even in a simple setting: fine-tuning overparameterized two-layer linear networks. We prove that the OOD error of fine-tuning is high when we initialize with a fixed or random head -- this is because while fine-tuning learns the head, the lower layers of the neural network change simultaneously and distort the pretrained features. Our analysis suggests that the easy two-step strategy of linear probing then full fine-tuning (LP-FT), sometimes used as a fine-tuning heuristic, combines the benefits of both fine-tuning and linear probing. Empirically, LP-FT outperforms both fine-tuning and linear probing on the above datasets (1% better ID, 10% better OOD than full fine-tuning).

Key facts. Cited by 722 papers in the map (data through 7 Oct 2026). Builds on 63 of its 87 references that are in the map. First posted on arXiv 21 Feb 2022.

Builds on

63 of 87 references.

References · 63

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2020A Simple Framework for Contrastive Learning of Visual Representationscited by 11,979
  3. 2014ImageNet Large Scale Visual Recognition Challengecited by 8,441
  4. 2019Momentum Contrast for Unsupervised Visual Representation Learningcited by 7,482
  5. 2019Parameter-Efficient Transfer Learning for NLPcited by 4,195
  6. 2021Prefix-Tuning: Optimizing Continuous Prompts for Generationcited by 3,878
  7. 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,769
  8. 2018Deep contextualized word representationscited by 2,285
  9. 2020Improved Baselines with Momentum Contrastive Learningcited by 2,282
  10. 2021Learning to Prompt for Vision-Language Modelscited by 2,269
  11. 2020The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalizationcited by 1,842
  12. 2021An Empirical Study of Training Self-Supervised Vision Transformerscited by 1,610
38 more
  1. 2019Do ImageNet Classifiers Generalize to ImageNet?cited by 1,507
  2. 2019Natural Adversarial Examplescited by 1,401
  3. 2020WILDS: A Benchmark of in-the-Wild Distribution Shiftscited by 1,366
  4. 2018Moment Matching for Multi-Source Domain Adaptationcited by 1,327
  5. 2018BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learningcited by 1,282
  6. 2019Learning Robust Global Representations by Penalizing Local Predictive Powercited by 1,010
  7. 2015End-to-End Training of Deep Visuomotor Policiescited by 847
  8. 2021Robust fine-tuning of zero-shot modelscited by 785
  9. 2013Exact solutions to the nonlinear dynamics of learning in deep linear neural networkscited by 748
  10. 2018Do Better ImageNet Models Transfer Better?cited by 560
  11. 2019Benign Overfitting in Linear Regressioncited by 544
  12. 2019Surprises in High-Dimensional Ridgeless Least Squares Interpolationcited by 508
  13. 2020Measuring Robustness to Natural Distribution Shifts in Image Classificationcited by 452
  14. 2019The generalization error of random features regression: Precise asymptotics and double descent curvecited by 410
  15. 2015An Introduction to Matrix Concentration Inequalitiescited by 391
  16. 2019Using Pre-Training Can Improve Model Robustness and Uncertaintycited by 373
  17. 2014In Search of the Real Inductive Bias: On the Role of Implicit Regularization in Deep Learningcited by 339
  18. 2017Functional Map of the Worldcited by 338
  19. 2017Implicit Regularization in Matrix Factorizationcited by 318
  20. 2018On the Optimization of Deep Networks: Implicit Acceleration by Overparameterizationcited by 280
  21. 2020The Risks of Invariant Risk Minimizationcited by 262
  22. 2019FreeLB: Enhanced Adversarial Training for Natural Language Understandingcited by 261
  23. 2019SMART: Robust and Efficient Fine-Tuning for Pre-trained Natural Language Models through Principled Regularized Optimizationcited by 253
  24. 2021Accuracy on the Line: On the Strong Correlation Between Out-of-Distribution and In-Distribution Generalizationcited by 248
  25. 2018Explicit Inductive Bias for Transfer Learning with Convolutional Networkscited by 226
  26. 2020Few-Shot Learning via Learning the Representation, Provablycited by 213
  27. 2018Do CIFAR-10 Classifiers Generalize to CIFAR-10?cited by 197
  28. 2019Two models of double descent for weak featurescited by 194
  29. 2020Geography-Aware Self-Supervised Learningcited by 190
  30. 2020On the Theory of Transfer Learning: The Importance of Task Diversitycited by 186
  31. 2018Algorithmic Regularization in Learning Deep Homogeneous Models: Layers are Automatically Balancedcited by 172
  32. 2019To Tune or Not to Tune? Adapting Pretrained Representations to Diverse Taskscited by 167
  33. 2018SpotTune: Transfer Learning through Adaptive Fine-tuningcited by 166
  34. 2017What do Neural Machine Translation Models Learn about Morphology?cited by 165
  35. 2020Better Fine-Tuning by Reducing Representational Collapsecited by 156
  36. 2020Understanding and Mitigating the Tradeoff Between Robustness and Accuracycited by 148
  37. 2019Implicit Regularization of Discrete Gradient Dynamics in Linear Neural Networkscited by 122
  38. 2019Harmless interpolation of noisy data in regressioncited by 121

50 most cited of 63 All 63 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 87; 2023: 196; 2024: 175; 2025: 132; 2026: 132.

722 papers.

Sort:

By area:

  1. 2022An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversioncited by 2,231
  2. 2022A Time Series is Worth 64 Words: Long-term Forecasting with Transformerscited by 1,981
  3. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,366
  4. 2024Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Headscited by 821
  5. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 445
  6. 2022Last Layer Re-Training is Sufficient for Robustness to Spurious Correlationscited by 379
  7. 2022Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Modelscited by 379
  8. 2023Fine-Tuning Language Models with Just Forward Passescited by 337
  9. 2023Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Biascited by 230
  10. 2023Where are we in the search for an Artificial Visual Cortex for Embodied Intelligence?cited by 215
  11. 2022Surgical Fine-Tuning Improves Adaptation to Distribution Shiftscited by 205
  12. 2022Patching open-vocabulary models by interpolating weightscited by 187
38 more
  1. 2024On the Effects of Data Scale on UI Control Agentscited by 187
  2. 2023Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Needcited by 183
  3. 2023Mitigating the Alignment Tax of RLHFcited by 181
  4. 2023LIV: Language-Image Representations and Rewards for Robotic Controlcited by 180
  5. 2024Probing the 3D Awareness of Visual Foundation Modelscited by 175
  6. 2024Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AIcited by 175
  7. 2022Finetune like you pretrain: Improved finetuning of zero-shot vision modelscited by 149
  8. 2022Diverse Weight Averaging for Out-of-Distribution Generalizationcited by 148
  9. 2023Debiasing Vision-Language Models via Biased Promptscited by 147
  10. 2024WARM: On the Benefits of Weight Averaged Reward Modelscited by 145
  11. 2024Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMscited by 142
  12. 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Trackingcited by 140
  13. 2022Domain Generalization by Mutual-Information Regularization with Pre-trained Modelscited by 133
  14. 2023LLM4TS: Aligning Pre-Trained LLMs as Data-Efficient Time-Series Forecasterscited by 127
  15. 2023A Unified Continual Learning Framework with General Parameter-Efficient Tuningcited by 124
  16. 2022Vision-Language Pre-training: Basics, Recent Advances, and Future Trendscited by 117
  17. 2023SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understandingcited by 111
  18. 2023Mechanistically analyzing the effects of fine-tuning on procedurally defined taskscited by 110
  19. 2025Sonata: Self-Supervised Learning of Reliable Point Representationscited by 108
  20. 2025Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoningcited by 101
  21. 2022Universal Prompt Tuning for Graph Neural Networkscited by 100
  22. 2023Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalizationcited by 100
  23. 2023Toward General-Purpose Robots via Foundation Models: A Survey and Meta-Analysiscited by 99
  24. 2022Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalizationcited by 97
  25. 2023CLIPood: Generalizing CLIP to Out-of-Distributionscited by 96
  26. 2022"This is my unicorn, Fluffy": Personalizing frozen vision-language representationscited by 95
  27. 2024Learning Dynamics of LLM Finetuningcited by 93
  28. 2023Fine-tuning can cripple your foundation model; preserving features may be the solutioncited by 92
  29. 2023Multimodality Helps Unimodality: Cross-Modal Few-Shot Learning with Multimodal Modelscited by 91
  30. 2022Picking on the Same Person: Does Algorithmic Monoculture lead to Outcome Homogenization?cited by 89
  31. 2024Model Stock: All we need is just a few fine-tuned modelscited by 87
  32. 2023Stronger, Fewer, & Superior: Harnessing Vision Foundation Models for Domain Generalized Semantic Segmentationcited by 82
  33. 2024Asymmetry in Low-Rank Adapters of Foundation Modelscited by 81
  34. 2023VPGTrans: Transfer Visual Prompt Generator across LLMscited by 79
  35. 2023A Foundation Language-Image Model of the Retina (FLAIR): Encoding Expert Knowledge in Text Supervisioncited by 78
  36. 2023Sensitivity-Aware Visual Parameter-Efficient Fine-Tuningcited by 74
  37. 2023Few-shot In-context Learning for Knowledge Base Question Answeringcited by 74
  38. 2022GLUE-X: Evaluating Natural Language Understanding Models from an Out-of-distribution Generalization Perspectivecited by 72

50 most cited of 722 All 722 (JSON)