Research map / Federated learning and privacy / Personalized and heterogeneous FL
Decoupled Training with Local Reinforcement Fine-Tuning in Federated Learning
Yuting Ma, Lechao Cheng, Xiaohua 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 |
|---|---|---|---|
| add decoupled encoder training (server text encoder, client image encoders) (ablation) | FedLoRA · Full Non-IID, seven datasets | base accuracy: 58.11 -> 86.68 | Improved |
| add decoupled encoder training (ablation) | FedLoRA · Few Non-IID, seven datasets | base accuracy: 78.32 -> 86.42 | Improved |
| add decoupled encoder training (ablation) | FedLoRA · IID settings, seven datasets | novel accuracy: 77.44 -> 75.35 (Few IID) | Hurt |
| add two-stage local fine-tuning (SFT then RL) (ablation) | FedLoRA · Few IID, seven datasets | novel accuracy: 77.44 -> 83.11 | Improved |
| add two-stage local fine-tuning alone (ablation) | FedLoRA · Full Non-IID, seven datasets | base accuracy: 58.11 -> 47.91 | Hurt |
| combine decoupled encoder and two-stage fine-tuning (ablation) | FedLoRA · Full Non-IID, seven datasets | harmonic mean: 63.12 -> 85.03 | Improved |
| use RL-only local fine-tuning instead of SFT-only (ablation) | local fine-tuning · base-to-novel datasets under communication rounds | novel accuracy | Improved |
| use SFT-only local fine-tuning (ablation) | local fine-tuning · base-to-novel datasets under communication rounds | base and novel accuracy | Hurt |
| use mixed reference model instead of latest or final-SFT only (ablation) | RL stage · few-shot vs full-data | accuracy | Mixed |
| vary image embedding upload ratio from 0.2 to 1.0 (ablation) | FedDTL · Full IID and Full Non-IID | accuracy | No clear effect |
| add noise perturbation to uploaded embeddings (ablation) | FedDTL · Full Non-IID, nine datasets | harmonic mean: 83.12 -> 84.15 | Improved |
| add embedding aggregation to uploaded embeddings (ablation) | FedDTL · Full Non-IID, nine datasets | harmonic mean: 83.12 -> 76.65 | Hurt |
Closest papers by meaning
- F^3OCUS - Federated Finetuning of Vision-Language Foundation Models with Optimal Client Layer Updating Strategy via Multi-objective Meta-Heuristics (CVPR 2025 · similarity 0.91)
- FedPHA: Federated Prompt Learning for Heterogeneous Client Adaptation (ICML 2025 · similarity 0.90)
- FedDAT: An Approach for Foundation Model Finetuning in Multi-Modal Heterogeneous Federated Learning (AAAI 2024 · similarity 0.88)
- Harmonizing Generalization and Personalization in Federated Prompt Learning (ICML 2024 · similarity 0.88)
- Federated Disentangled Tuning with Textual Prior Decoupling and Visual Dynamic Adaptation (ICML 2025 · similarity 0.88)
- Enhancing Foundation Models with Federated Domain Knowledge Infusion (ICML 2025 · similarity 0.87)
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.
