mlsanityData to 14 Oct 2026

Fine-tuning can cripple your foundation model; preserving features may be the solution

arXiv 2308.13320 ↗First posted 25 Aug 2023cs.LG cs.CVcited by 96

Not on the map. Why?

Cited by 34 in LLM safety and security, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

Pre-trained foundation models, due to their enormous capacity and exposure to vast amounts of data during pre-training, are known to have learned plenty of real-world concepts. An important step in making these pre-trained models effective on downstream tasks is to fine-tune them on related datasets. While various fine-tuning methods have been devised and have been shown to be highly effective, we observe that a fine-tuned model's ability to recognize concepts on tasks textitdifferent from the downstream one is reduced significantly compared to its pre-trained counterpart. This is an undesirable effect of fine-tuning as a substantial amount of resources was used to learn these pre-trained concepts in the first place. We call this phenomenon ''concept forgetting'' and via experiments show that most end-to-end fine-tuning approaches suffer heavily from this side effect. To this end, we propose a simple fix to this problem by designing a new fine-tuning method called textitLDIFS (short for ℓ2 distance in feature space) that, while learning new concepts related to the downstream task, allows a model to preserve its pre-trained knowledge as well. Through extensive experiments on 10 fine-tuning tasks we show that textitLDIFS significantly reduces concept forgetting. Additionally, we show that LDIFS is highly effective in performing continual fine-tuning on a sequence of tasks as well, in comparison with both fine-tuning as well as continual learning baselines.

Key facts. Cited by 96 papers in the map (data through 14 Oct 2026). Builds on 51 of its 79 references that are in the map. First posted on arXiv 25 Aug 2023.

Builds on

51 of 79 references.

References · 51

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,744
  2. 2018The Unreasonable Effectiveness of Deep Features as a Perceptual Metriccited by 11,177
  3. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  4. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,403
  5. 2016Overcoming catastrophic forgetting in neural networkscited by 4,787
  6. 2021On the Opportunities and Risks of Foundation Modelscited by 4,033
  7. 2021Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervisioncited by 3,864
  8. 2016Learning without Forgettingcited by 2,496
  9. 2016iCaRL: Incremental Classifier and Representation Learningcited by 2,333
  10. 2021Learning to Prompt for Vision-Language Modelscited by 2,281
  11. 2013Describing Textures in the Wildcited by 2,125
  12. 2017Gradient Episodic Memory for Continual Learningcited by 1,898
38 more
  1. 2022Large Language Models Encode Clinical Knowledgecited by 1,830
  2. 2017EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classificationcited by 1,496
  3. 2022Conditional Prompt Learning for Vision-Language Modelscited by 1,433
  4. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,396
  5. 2022CoCa: Contrastive Captioners are Image-Text Foundation Modelscited by 1,274
  6. 2022Editing Models with Task Arithmeticcited by 1,120
  7. 2016Paying More Attention to Attention: Improving the Performance of Convolutional Neural Networks via Attention Transfercited by 1,077
  8. 2021CLIP-Adapter: Better Vision-Language Models with Feature Adapterscited by 1,026
  9. 2021Learning to Prompt for Continual Learningcited by 944
  10. 2017Remote Sensing Image Scene Classification: Benchmark and State of the Artcited by 797
  11. 2021Robust fine-tuning of zero-shot modelscited by 797
  12. 2018Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigencecited by 742
  13. 2022Fine-Tuning can Distort Pretrained Features and Underperform Out-of-Distributioncited by 739
  14. 2021FLAVA: A Foundational Language And Vision Alignment Modelcited by 649
  15. 2022DualPrompt: Complementary Prompting for Rehearsal-free Continual Learningcited by 642
  16. 2022MedCLIP: Contrastive Learning from Unpaired Medical Images and Textcited by 570
  17. 2022CODA-Prompt: COntinual Decomposed Attention-based Prompting for Rehearsal-Free Continual Learningcited by 450
  18. 2021Tip-Adapter: Training-free CLIP-Adapter for Better Vision-Language Modelingcited by 431
  19. 2022Test-Time Prompt Tuning for Zero-Shot Generalization in Vision-Language Modelscited by 382
  20. 2017Measuring Catastrophic Forgetting in Neural Networkscited by 360
  21. 2018Knowledge Transfer via Distillation of Activation Boundaries Formed by Hidden Neuronscited by 238
  22. 2018Explicit Inductive Bias for Transfer Learning with Convolutional Networkscited by 227
  23. 2018Learning Deep Representations with Probabilistic Knowledge Transfercited by 213
  24. 2022Patching open-vocabulary models by interpolating weightscited by 190
  25. 2023SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Modelcited by 180
  26. 2022Fine-tuned Language Models are Continual Learnerscited by 156
  27. 2022Finetune like you pretrain: Improved finetuning of zero-shot vision modelscited by 150
  28. 2021Lifelong Pretraining: Continually Adapting Language Models to Emerging Corporacited by 115
  29. 2023Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Modelscited by 112
  30. 2016Less-forgetting Learning in Deep Neural Networkscited by 97
  31. 2020ALP-KD: Attention-Based Layer Projection for Knowledge Distillationcited by 66
  32. 2022CLIP model is an Efficient Continual Learnercited by 65
  33. 2022SVL-Adapter: Self-Supervised Adapter for Vision-Language Pretrained Modelscited by 34
  34. 2023Chain of Thought Prompt Tuning in Vision Language Modelscited by 34
  35. 2023Preserving Pre-trained Features Helps Calibrate Fine-tuned Language Modelscited by 29
  36. 2022Context-Aware Robust Fine-Tuningcited by 28
  37. 2023Prompt Tuning based Adapter for Vision-Language Model Adaptioncited by 10
  38. 2022Self-Distillation for Further Pre-training of Transformerscited by 9

50 most cited of 51 All 51 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 2; 2024: 22; 2025: 44; 2026: 28.

96 papers.

Sort:

By area:

  1. 2024Safety Alignment Should Be Made More Than Just a Few Tokens Deepcited by 472
  2. 2024Safety Layers in Aligned Large Language Models: The Key to LLM Securitycited by 134
  3. 2024Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attackcited by 115
  4. 2024Keeping LLMs Aligned After Fine-tuning: The Crucial Role of Prompt Templatescited by 96
  5. 2024Representation Noising: A Defence Mechanism Against Harmful Finetuningcited by 96
  6. 2024Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbationcited by 92
  7. 2024Harmful Fine-tuning Attacks and Defenses for Large Language Models: A Surveycited by 90
  8. 2024Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuningcited by 72
  9. 2024SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selectioncited by 64
  10. 2023When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitationscited by 43
  11. 2025CAD-Coder: An Open-Source Vision-Language Model for Computer-Aided Design Code Generationcited by 43
  12. 2025SaLoRA: Safety-Alignment Preserved Low-Rank Adaptationcited by 42
38 more
  1. 2024Targeted Vaccine: Safety Alignment for Large Language Models against Harmful Fine-Tuning via Layer-wise Perturbationcited by 37
  2. 2025SafeMERGE: Preserving Safety Alignment in Fine-Tuned Large Language Models via Selective Layer-Wise Model Mergingcited by 25
  3. 2024Fine-Tuning is Fine, if Calibratedcited by 24
  4. 2024A Practitioner's Guide to Continual Multimodal Pretrainingcited by 23
  5. 2025Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safetycited by 22
  6. 2024kNN-CLIP: Retrieval Enables Training-Free Segmentation on Continually Expanding Large Vocabulariescited by 20
  7. 2025Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbationcited by 20
  8. 2025Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasetscited by 20
  9. 2025AsFT: Anchoring Safety During LLM Fine-Tuning Within Narrow Safety Basincited by 19
  10. 2024BendVLM: Test-Time Debiasing of Vision-Language Embeddingscited by 17
  11. 2025Compromising Honesty and Harmlessness in Language Models via Deception Attackscited by 17
  12. 2025Vulnerability-Aware Alignment: Mitigating Uneven Forgetting in Harmful Fine-Tuningcited by 17
  13. 2025In-Training Defenses against Emergent Misalignment in Language Modelscited by 15
  14. 2025Multi-Scale Finetuning for Encoder-based Time Series Foundation Modelscited by 14
  15. 2025Unveiling the Basin-Like Loss Landscape in Large Language Modelscited by 13
  16. 2025Video Unlearning via Low-Rank Refusal Vectorcited by 13
  17. 2023FD-Align: Feature Discrimination Alignment for Fine-tuning Pre-Trained Models in Few-Shot Learningcited by 12
  18. 2025ResMoE: Space-efficient Compression of Mixture of Experts LLMs via Residual Restorationcited by 12
  19. 2025LoX: Low-Rank Extrapolation Robustifies LLM Safety Against Fine-tuningcited by 12
  20. 2025Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Schedulercited by 12
  21. 2025CTRAP: Embedding Collapse Trap to Safeguard Large Language Models from Harmful Fine-Tuningcited by 11
  22. 2025Shape it Up! Restoring LLM Safety during Finetuningcited by 11
  23. 2025Gradient Surgery for Safe LLM Fine-Tuningcited by 11
  24. 2025Self-Destructive Language Modelcited by 10
  25. 2025Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learningcited by 10
  26. 2025Subspace-Boosted Model Mergingcited by 9
  27. 2025A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Spacecited by 9
  28. 2024Toward Secure Tuning: Mitigating Security Risks from Instruction Fine-Tuningcited by 8
  29. 2024FEET: A Framework for Evaluating Embedding Techniquescited by 8
  30. 2025Token Buncher: Shielding LLMs from Harmful Reinforcement Learning Fine-Tuningcited by 8
  31. 2026Why Supervised Fine-Tuning Fails to Learn: A Systematic Study of Incomplete Learning in Large Language Modelscited by 8
  32. 2025LookAhead Tuning: Safer Language Models via Partial Answer Previewscited by 7
  33. 2025Calibrated Language Models and How to Find Them with Label Smoothingcited by 7
  34. 2025No, of Course I Can! Deeper Fine-Tuning Attacks That Bypass Token-Level Safety Mechanismscited by 6
  35. 2025LoRA is All You Need for Safety Alignment of Reasoning LLMscited by 6
  36. 2026Narrow Fine-Tuning Erodes Safety Alignment in Vision-Language Agentscited by 6
  37. 2026Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influencecited by 6
  38. 2026Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignmentcited by 6

50 most cited of 96 All 96 (JSON)