Research map / Federated learning and privacy / Personalized and heterogeneous FL
LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement
Jieming Bian, Lei Wang, Letian Zhang, Jie Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply residual correction to B instead of to A (ablation) | LoRA-FAIR · DomainNet ViT | average accuracy: 76.42 -> 77.07 | Improved |
| apply residual correction to B instead of to both A and B (ablation) | LoRA-FAIR · DomainNet ViT | average accuracy: 75.55 -> 77.07 | Improved |
| add regularization term (lambda = 0.01 instead of 0) (ablation) | LoRA-FAIR · DomainNet ViT | average accuracy: 73.22 -> 77.07 | Improved |
| vary regularization weight among small positive values (0.01, 0.02) (ablation) | LoRA-FAIR · DomainNet ViT | average accuracy | No clear effect |
| use LoRA-FAIR instead of FLoRA with 2 local epochs | FLoRA · DomainNet ViT | average accuracy: 75.53 -> 77.07 | Improved |
| use LoRA-FAIR instead of FLoRA with 10 local epochs | FLoRA · DomainNet ViT | average accuracy: 73.60 -> 74.31 | Improved |
| add LoRA-FAIR on top of HETLoRA for heterogeneous ranks | HETLoRA · DomainNet ViT | average accuracy: 60.89 -> 74.22 | Improved |
| add LoRA-FAIR on top of FlexLoRA in heterogeneous ranks | FlexLoRA · DomainNet ViT | average accuracy: 73.50 -> 74.22 | Improved |
| use LoRA-FAIR instead of FedIT (feature non-IID) | FedIT · DomainNet MLP-Mixer | average accuracy: 64.37 -> 65.87 | Improved |
| use LoRA-FAIR instead of FlexLoRA (label non-IID) | FlexLoRA · DomainNet MLP-Mixer | average accuracy: 61.20 -> 62.28 | Improved |
| freeze A and train only B (FFA-LoRA) instead of FedIT | FedIT · DomainNet MLP-Mixer (label non-IID) | average accuracy: 60.77 -> 52.88 | Hurt |
| use LoRA-FAIR instead of FLoRA on NICO++ (feature non-IID) | FLoRA · NICO++ ViT | average accuracy: 90.93 -> 91.24 | Improved |
Later papers that compared with this method
2 later accepted papers report LoRA-FAIR as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- FedLoVA: Value-Only Aggregation for Federated LoRA Fine-Tuning of Large Language Models (NeurIPS 2026 · better on 5 of 5 shared results)
- OS-Fed: One Snapshot Is All You Need (CVPR 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adaptive LoRA Experts Allocation and Selection for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- Improving LoRA in Privacy-preserving Federated Learning (ICLR 2024 · similarity 0.90)
- LoRA-GA: Low-Rank Adaptation with Gradient Approximation (NeurIPS 2024 · similarity 0.90)
- Robust Federated Finetuning of LLMs via Alternating Optimization of LoRA (NeurIPS 2025 · similarity 0.90)
- Selective Aggregation for Low-Rank Adaptation in Federated Learning (ICLR 2025 · similarity 0.89)
- FedSVD: Adaptive Orthogonalization for Private Federated Learning with LoRA (NeurIPS 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
