Research map / Federated learning and privacy / Personalized and heterogeneous FL
FedGRPO: Privately Optimizing Foundation Models with Group-Relative Rewards from Domain Clients
Gongxi Zhu, Hanlin Gu, Lixin Fan, Qiang Yang, Yuxing Han
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 |
|---|---|---|---|
| use FedGRPO instead of the best federated baseline DPSDA-FL+SFT | FedGRPO · Math-benchmark, Qwen2.5-Math-7B | average accuracy: 0.253 -> 0.369 | Improved |
| use FedGRPO instead of DPSDA-FL+GRPO | FedGRPO · Math-benchmark, Qwen2.5-Math-1.5B | average accuracy: 0.275 -> 0.338 | Improved |
| use FedGRPO instead of centralized GRPO (upper bound) | FedGRPO · OpenR1-Math, Qwen2.5-Math-7B | average accuracy: 0.379 -> 0.388 | Improved |
| use FedGRPO instead of centralized GRPO | FedGRPO · Math-benchmark, Qwen2.5-Math-7B | average accuracy: 0.370 -> 0.369 | No clear effect |
| use FedGRPO instead of centralized GRPO | FedGRPO · Math-benchmark, Qwen2.5-3B | average accuracy: 0.168 -> 0.160 | Hurt |
| use reward-model rewards instead of ground-truth answers | FedGRPO · Math-benchmark, Qwen2.5-Math-1.5B | average accuracy: 0.338 -> 0.310 | Hurt |
| use FedGRPO without ground-truth answers instead of zero-shot | FedGRPO · Math-benchmark, Qwen2.5-Math-1.5B | average accuracy: 0.163 -> 0.310 | Improved |
| use FedGRPO instead of zero-shot on multi-hop QA | FedGRPO · Qwen2.5-3B QA | average accuracy: 0.080 -> 0.245 | Improved |
| use FedGRPO instead of DPSDA-FL+SFT on multi-hop QA | FedGRPO · Qwen2.5-1.5B QA | average accuracy: 0.163 -> 0.206 | Improved |
| increase number of clients from 4 to 20 (ablation) | FedGRPO · Qwen2.5-Math-1.5B | average accuracy: 0.29 -> 0.36 | Improved |
| increase number of clients from 4 to 20 (ablation) | FedGRPO · Qwen2.5-Math-1.5B, AMC | accuracy: 0.37 -> 0.47 | Improved |
| increase number of clients from 4 to 20 (ablation) | FedGRPO · Qwen2.5-Math-1.5B, Olympiad | accuracy: 0.32 -> 0.38 | Improved |
Closest papers by meaning
- FedP3: Federated Personalized and Privacy-friendly Network Pruning under Model Heterogeneity (ICLR 2024 · similarity 0.86)
- FedLMT: Tackling System Heterogeneity of Federated Learning via Low-Rank Model Training with Theoretical Guarantees (ICML 2024 · similarity 0.86)
- Ranking-based Client Imitation Selection for Efficient Federated Learning (ICML 2024 · similarity 0.86)
- Clients Collaborate: Flexible Differentially Private Federated Learning with Guaranteed Improvement of Utility-Privacy Trade-off (ICML 2025 · similarity 0.85)
- Dual-Personalizing Adapter for Federated Foundation Models (NeurIPS 2024 · similarity 0.85)
- Beyond the Federation: Topology-aware Federated Learning for Generalization to Unseen Clients (ICML 2024 · similarity 0.85)
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.
