mlsanityData to 7 Oct 2026

Editing Models with Task Arithmetic

arXiv 2212.04089 ↗First posted 8 Dec 2022cs.LG cs.CL cs.CVcited by 1,076

Not on the map. Why?

Cited by 154 in LLM safety and security, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

Changing how pre-trained models behave -- e.g., improving their performance on a downstream task or mitigating biases learned during pre-training -- is a common practice when developing machine learning systems. In this work, we propose a new paradigm for steering the behavior of neural networks, centered around textittask vectors. A task vector specifies a direction in the weight space of a pre-trained model, such that movement in that direction improves performance on the task. We build task vectors by subtracting the weights of a pre-trained model from the weights of the same model after fine-tuning on a task. We show that these task vectors can be modified and combined together through arithmetic operations such as negation and addition, and the behavior of the resulting model is steered accordingly. Negating a task vector decreases performance on the target task, with little change in model behavior on control tasks. Moreover, adding task vectors together can improve performance on multiple tasks at once. Finally, when tasks are linked by an analogy relationship of the form ``A is to B as C is to D", combining task vectors from three of the tasks can improve performance on the fourth, even when no data from the fourth task is used for training. Overall, our experiments with several models, modalities and tasks show that task arithmetic is a simple, efficient and effective way of editing models.

Key facts. Cited by 1,076 papers in the map (data through 7 Oct 2026). Builds on 81 of its 111 references that are in the map. First posted on arXiv 8 Dec 2022.

Builds on

81 of 111 references.

References · 81

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  3. 2020Language Models are Few-Shot Learnerscited by 32,199
  4. 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,160
  5. 2022Training language models to follow instructions with human feedbackcited by 15,712
  6. 2019Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformercited by 14,491
  7. 2018GLUE: A Multi-Task Benchmark and Analysis Platform for Natural Language Understandingcited by 4,854
  8. 2022Flamingo: a Visual Language Model for Few-Shot Learningcited by 4,337
  9. 2016SQuAD: 100,000+ Questions for Machine Comprehension of Textcited by 4,115
  10. 2021Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervisioncited by 3,848
  11. 2021The Power of Scale for Parameter-Efficient Prompt Tuningcited by 3,769
  12. 2021Finetuned Language Models Are Zero-Shot Learnerscited by 3,747
38 more
  1. 2016Pointer Sentinel Mixture Modelscited by 2,992
  2. 2022Training Compute-Optimal Large Language Modelscited by 2,927
  3. 2018Neural Tangent Kernel: Convergence and Generalization in Neural Networkscited by 2,597
  4. 2015Character-level Convolutional Networks for Text Classificationcited by 2,293
  5. 2019HuggingFace's Transformers: State-of-the-art Natural Language Processingcited by 2,201
  6. 2013Describing Textures in the Wildcited by 2,110
  7. 2017EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classificationcited by 1,488
  8. 2021Multitask Prompted Training Enables Zero-Shot Task Generalizationcited by 1,462
  9. 2019Natural Adversarial Examplescited by 1,401
  10. 2022Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timecited by 1,366
  11. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  12. 2018Averaging Weights Leads to Wider Optima and Better Generalizationcited by 1,194
  13. 2019A Comprehensive Survey on Transfer Learningcited by 1,154
  14. 2017Visualizing the Loss Landscape of Neural Netscited by 1,126
  15. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  16. 2018Neural Network Acceptability Judgmentscited by 952
  17. 2017RACE: Large-scale ReAding Comprehension Dataset From Examinationscited by 851
  18. 2017Remote Sensing Image Scene Classification: Benchmark and State of the Artcited by 791
  19. 2021Robust fine-tuning of zero-shot modelscited by 785
  20. 2021Cross-Task Generalization via Natural Language Crowdsourcing Instructionscited by 696
  21. 2019Linear Mode Connectivity and the Lottery Ticket Hypothesiscited by 627
  22. 2021Merging Models with Fisher-Weighted Averagingcited by 613
  23. 2019Eternal Sunshine of the Spotless Net: Selective Forgetting in Deep Networkscited by 599
  24. 2018Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNscited by 585
  25. 2021Editing Factual Knowledge in Language Modelscited by 548
  26. 2022Improving alignment of dialogue agents via targeted human judgementscited by 527
  27. 2021Datasets: A Community Library for Natural Language Processingcited by 526
  28. 2021Fast Model Editing at Scalecited by 491
  29. 2020UnifiedQA: Crossing Format Boundaries With a Single QA Systemcited by 475
  30. 2022Git Re-Basin: Merging Models modulo Permutation Symmetriescited by 460
  31. 2021MetaICL: Learning to Learn In Contextcited by 450
  32. 2022Memory-Based Model Editing at Scalecited by 443
  33. 2019Deep Ensembles: A Loss Landscape Perspectivecited by 438
  34. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 434
  35. 2019Multi-News: a Large-Scale Multi-Document Summarization Dataset and Abstractive Hierarchical Modelcited by 431
  36. 2019Nuanced Metrics for Measuring Unintended Bias with Real Data for Text Classificationcited by 396
  37. 2020Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stoppingcited by 377
  38. 2020What is being transferred in transfer learning?cited by 365

50 most cited of 81 All 81 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 3; 2023: 65; 2024: 238; 2025: 337; 2026: 433.

1,076 papers.

Sort:

By area:

  1. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  2. 2023TIES-Merging: Resolving Interference When Merging Modelscited by 795
  3. 2023Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formattingcited by 778
  4. 2023Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunchcited by 679
  5. 2024The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearningcited by 488
  6. 2023Editing Large Language Models: Problems, Methods, and Opportunitiescited by 367
  7. 2023LoraHub: Efficient Cross-Task Generalization via Dynamic LoRA Compositioncited by 359
  8. 2023SVDiff: Compact Parameter Space for Diffusion Fine-Tuningcited by 338
  9. 2023Emergent Linear Representations in World Models of Self-Supervised Sequence Modelscited by 313
  10. 2024Improving Alignment and Robustness with Circuit Breakerscited by 302
  11. 2023Function Vectors in Large Language Modelscited by 292
  12. 2023Personalized Soups: Personalized Large Language Model Alignment via Post-hoc Parameter Mergingcited by 271
38 more
  1. 2023AdaMerging: Adaptive Model Merging for Multi-Task Learningcited by 248
  2. 2024Rethinking Machine Unlearning for Large Language Modelscited by 247
  3. 2024Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunitiescited by 247
  4. 2023In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steeringcited by 240
  5. 2022OPT-IML: Scaling Language Model Instruction Meta Learning through the Lens of Generalizationcited by 235
  6. 2023Rewarded soups: towards Pareto-optimal alignment by interpolating weights fine-tuned on diverse rewardscited by 232
  7. 2023Parameter-Efficient Fine-Tuning Methods for Pretrained Language Models: A Critical Review and Assessmentcited by 227
  8. 2023Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained Modelscited by 218
  9. 2023LCM-LoRA: A Universal Stable-Diffusion Acceleration Modulecited by 211
  10. 2023LIV: Language-Image Representations and Rewards for Robotic Controlcited by 180
  11. 2024A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicitycited by 179
  12. 2023ZipIt! Merging Models from Different Tasks without Trainingcited by 178
  13. 2024Simplicity Prevails: Rethinking Negative Preference Optimization for LLM Unlearningcited by 160
  14. 2024A Comprehensive Study of Knowledge Editing for Large Language Modelscited by 157
  15. 2023Can Sensitive Information Be Deleted From LLMs? Objectives for Defending Against Extraction Attackscited by 156
  16. 2023Composing Parameter-Efficient Modules with Arithmetic Operationscited by 155
  17. 2024Eight Methods to Evaluate Robust Unlearning in LLMscited by 153
  18. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 149
  19. 2024WARM: On the Benefits of Weight Averaged Reward Modelscited by 145
  20. 2024Task Singular Vectors: Reducing Task Interference in Model Mergingcited by 143
  21. 2024Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Trackingcited by 140
  22. 2023Model Breadcrumbs: Scaling Multi-Task Model Merging with Sparse Maskscited by 130
  23. 2024Localizing Task Information for Improved Model Merging and Compressioncited by 130
  24. 2024Safe LoRA: the Silver Lining of Reducing Safety Risks when Fine-tuning Large Language Modelscited by 122
  25. 2024Model merging with SVD to tie the Knotscited by 118
  26. 2024Twin-Merging: Dynamic Integration of Modular Expertise in Model Mergingcited by 114
  27. 2024Large Language Model Unlearning via Embedding-Corrupted Promptscited by 112
  28. 2023SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understandingcited by 111
  29. 2025A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deploymentcited by 110
  30. 2024EMR-Merging: Tuning-Free High-Performance Model Mergingcited by 108
  31. 2023Merge, Then Compress: Demystify Efficient SMoE with Hints from Its Routing Policycited by 104
  32. 2022Model Ratatouille: Recycling Diverse Models for Out-of-Distribution Generalizationcited by 97
  33. 2023Language Models Implement Simple Word2Vec-style Vector Arithmeticcited by 97
  34. 2024Representation Surgery for Multi-Task Model Mergingcited by 95
  35. 2023Fine-tuning can cripple your foundation model; preserving features may be the solutioncited by 92
  36. 2025No Task Left Behind: Isotropic Model Merging with Common and Task-Specific Subspacescited by 90
  37. 2023Deep Model Fusion: A Surveycited by 89
  38. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 89

50 most cited of 1,076 All 1,076 (JSON)