Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedLoVA: Value-Only Aggregation for Federated LoRA Fine-Tuning of Large Language Models
Ensieh Khazaei, Baturalp Buyukates, Dimitrios Hatzinakos
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FedLoVA, a federated LoRA fine-tuning method that aggregates only the value-projection LoRA components, motivated by analysis of the distinct roles of query and value projections.
Problem
LoRA is applied to both query and value projections in transformers in federated learning without considering their distinct roles, which affects federated LoRA fine-tuning of LLMs.
Approach
Analyzes the distinct roles of LoRA on query versus value projections and proposes value-only aggregation in federated LoRA, so that only value-projection adapters are aggregated across clients.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| share only value-projection LoRA instead of both Q and V | FedLoRA · GLUE avg (RoBERTa-large) | accuracy: 87.69 -> 88.13 | Improved |
| share V rather than Q with server (ablation) | FedLoVA vs FedLoQA · GLUE avg | accuracy: 86.89 -> 88.13 | Improved |
| share only V (halving communication) | FedLoRA · communication cost to target accuracy | total comm.: 215.3 M -> 68.3 M (MNLI-m) | Improved |
| apply LoRA to V instead of Q (ablation) | LoRA · SST2/MNLI-mm convergence (RoBERTa-large) | rounds to target: 3.3x / 2.4x fewer rounds | Improved |
| FedLoVA vs FedLoRA on multilingual sentiment | FedLoRA · Multilingual Sentiment | accuracy: 68.90 -> 70.18 | Improved |
| FedLoVA vs FedLoRA on GSM8K | FedLoRA · GSM8K (Llama3-8B) | accuracy: 43.59 -> 45.79 | Improved |
| share Q only instead of V (ablation) | FedLoVA · GSM8K (Llama3-8B) | accuracy: 45.79 -> 38.74 | Hurt |
| add FedLoVA over FFA-LoRA on RTE | FFA-LoRA · RTE | accuracy: 80.99 -> 84.13 | Improved |
| add FedLoVA over FedSA-LoRA on QQP | FedSA-LoRA · QQP | accuracy: 82.82 -> 86.23 | Improved |
| apply LoRA to Q+K+V instead of Q+V (ablation) | LoRA · MNLI-mm | accuracy: 87.80 -> 88.03 | No clear effect |
| adapt MLP instead of Q/V (ablation) | LoRA · SST2 | accuracy: 95.49 -> 95.70 | Mixed |
| add FedLoVA vs FedLoRA with 50 clients | FedLoRA baselines · MNLI-m, 50 clients | accuracy: 88.60 -> 89.18 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LoRA-FAIR — LoRA-FAIR: Federated LoRA Fine-Tuning with Aggregation and Initialization Refinement (ICCV 2025) · better on 5 of 5 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.
Experimental setup
Builds on: LoRA
Closest papers by meaning
- Robust Federated Finetuning of LLMs via Alternating Optimization of LoRA (NeurIPS 2025 · similarity 0.92)
- FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations (NeurIPS 2024 · similarity 0.92)
- Adaptive LoRA Experts Allocation and Selection for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- Ravan: Multi-Head Low-Rank Adaptation for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.91)
- Federated Fine-tuning of Large Language Models under Heterogeneous Tasks and Client Resources (NeurIPS 2024 · similarity 0.90)
- FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models (NeurIPS 2025 · similarity 0.90)
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.
