Lune

ICLR2025Top-tier venue

Fine-Tuning Attention Modules Only: Enhancing Weight Disentanglement in Task Arithmetic

Ruochen Jin, Bojian Hou, Jiancong Xiao, Weijie J. Su, Li Shen

Abstract

In recent years, task arithmetic has garnered increasing attention. This approach edits pre-trained models directly in weight space by combining the fine-tuned weights of various tasks into a unified model. Its efficiency and cost-effectiveness stem from its training-free combination, contrasting with traditional methods that require model training on large datasets for multiple tasks. However, applying such a unified model to individual tasks can lead to interference from other tasks (lack of weight disentanglement). To address this issue, Neural Tangent Kernel (NTK) linearization has been employed to leverage a "kernel behavior", facilitating weight disentanglement and mitigating adverse effects from unrelated tasks. Despite its benefits, NTK linearization presents drawbacks, including doubled training costs, as well as reduced performance of individual models. To tackle this problem, we propose a simple yet effective and efficient method that is to finetune the attention modules only in the Transformer. Our study reveals that the attention modules exhibit kernel behavior, and fine-tuning the attention modules only significantly improves weight disentanglement. To further understand how our method improves the weight disentanglement of task arithmetic, we present a comprehensive study of task arithmetic by differentiating the role of the representation module and task-specific module. In particular, we find that the representation module plays an important role in improving weight disentanglement whereas the task-specific modules such as the classification heads can degenerate the weight disentanglement performance. 1

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

17 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • Dataless Weight Disentanglement in Task Arithmetic via Kronecker-Factored Approximate Curvature cites Learning Transferable Visual Models From Natural Language Supervision
  • Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging cites Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
  • Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging cites Learning Transferable Visual Models From Natural Language Supervision
  • Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  • Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time cites Learning Transferable Visual Models From Natural Language Supervision
  • Understanding and Enforcing Weight Disentanglement in Task Arithmetic cites Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
  • Understanding and Enforcing Weight Disentanglement in Task Arithmetic cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  • Understanding and Enforcing Weight Disentanglement in Task Arithmetic cites Learning Transferable Visual Models From Natural Language Supervision
  • Distilling Linearized Behavior into Non-linear Fine-Tuning for Effective Task Arithmetic cites Dataless Weight Disentanglement in Task Arithmetic via Kronecker-Factored Approximate Curvature
  • Distilling Linearized Behavior into Non-linear Fine-Tuning for Effective Task Arithmetic cites Learning Transferable Visual Models From Natural Language Supervision
  • Navigating the Accuracy-Size Trade-Off with Flexible Model Merging cites Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time
  • Navigating the Accuracy-Size Trade-Off with Flexible Model Merging cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  • Navigating the Accuracy-Size Trade-Off with Flexible Model Merging cites Learning Transferable Visual Models From Natural Language Supervision
  • Learning Transferable Visual Models From Natural Language Supervision cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers17

  1. Merging on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model MergingAnke Tang, Enneng Yang, Li Shen, Yong Luo et al.NeurIPS 2025 · 8 citations
  2. Dataless Weight Disentanglement in Task Arithmetic via Kronecker-Factored Approximate CurvatureAngelo Porrello, Pietro Buzzega, Felix Dangel, Thomas Sommariva et al.ICLR 2026 · 6 citations
  3. Understanding and Enforcing Weight Disentanglement in Task ArithmeticShangge Liu, Yuehan Yin, Lei Wang, Qi Fan et al.CVPR 2026 · 3 citations
  4. Distilling Linearized Behavior into Non-linear Fine-Tuning for Effective Task ArithmeticThomas Sommariva, Francesca Morandi, Simone Calderara, Angelo PorrelloICML 2026 · 2 citations
  5. Navigating the Accuracy-Size Trade-Off with Flexible Model MergingAkash Balasaheb Dhasade, Divyansh Jhunjhunwala, Milos Vujasinovic, Gauri Joshi et al.ICLR 2026 · 2 citations

Builds on18

  1. Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
  2. Training language models to follow instructions with human feedbackLong Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida et al.NeurIPS 2022 · 24,707 citations
  3. An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn et al.ICLR 2021 · 21,477 citations
  4. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional DomainsMatthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil et al.NeurIPS 2020 · 4,036 citations
  5. Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference timeMitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs et al.ICML 2022 · 1,464 citations

Related papers

  1. Task Arithmetic in the Tangent Space: Improved Editing of Pre-Trained ModelsGuillermo Ortiz-Jiménez, Alessandro Favero, Pascal FrossardNeurIPS 2023 · 272 citations
  2. Efficient Model Editing with Task-Localized Sparse Fine-tuningLeonardo Iurada, Marco Ciccone, Tatiana TommasiICLR 2025
  3. Leveraging Submodule Linearity Enhances Task Arithmetic Performance in LLMsRui Dai, Sile Hu, Xu Shen, Yonggang Zhang et al.ICLR 2025
  4. When is Task Vector Provably Effective for Model Editing? A Generalization Analysis of Nonlinear TransformersHongkang Li, Yihua Zhang, Shuai Zhang, Pin-Yu Chen et al.ICLR 2025
  5. Editing models with task arithmeticGabriel Ilharco, Marco Túlio Ribeiro, Mitchell Wortsman, Ludwig Schmidt et al.ICLR 2023 · 31 citations
Dusk over the sea between two cliffs drawn in fine vertical lines