AI Paper Atlas

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

AAAI 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Non-IID federated learning · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
use FedGRPO instead of the best federated baseline DPSDA-FL+SFTFedGRPO · Math-benchmark, Qwen2.5-Math-7Baverage accuracy: 0.253 -> 0.369Improved
use FedGRPO instead of DPSDA-FL+GRPOFedGRPO · Math-benchmark, Qwen2.5-Math-1.5Baverage accuracy: 0.275 -> 0.338Improved
use FedGRPO instead of centralized GRPO (upper bound)FedGRPO · OpenR1-Math, Qwen2.5-Math-7Baverage accuracy: 0.379 -> 0.388Improved
use FedGRPO instead of centralized GRPOFedGRPO · Math-benchmark, Qwen2.5-Math-7Baverage accuracy: 0.370 -> 0.369No clear effect
use FedGRPO instead of centralized GRPOFedGRPO · Math-benchmark, Qwen2.5-3Baverage accuracy: 0.168 -> 0.160Hurt
use reward-model rewards instead of ground-truth answersFedGRPO · Math-benchmark, Qwen2.5-Math-1.5Baverage accuracy: 0.338 -> 0.310Hurt
use FedGRPO without ground-truth answers instead of zero-shotFedGRPO · Math-benchmark, Qwen2.5-Math-1.5Baverage accuracy: 0.163 -> 0.310Improved
use FedGRPO instead of zero-shot on multi-hop QAFedGRPO · Qwen2.5-3B QAaverage accuracy: 0.080 -> 0.245Improved
use FedGRPO instead of DPSDA-FL+SFT on multi-hop QAFedGRPO · Qwen2.5-1.5B QAaverage accuracy: 0.163 -> 0.206Improved
increase number of clients from 4 to 20 (ablation)FedGRPO · Qwen2.5-Math-1.5Baverage accuracy: 0.29 -> 0.36Improved
increase number of clients from 4 to 20 (ablation)FedGRPO · Qwen2.5-Math-1.5B, AMCaccuracy: 0.37 -> 0.47Improved
increase number of clients from 4 to 20 (ablation)FedGRPO · Qwen2.5-Math-1.5B, Olympiadaccuracy: 0.32 -> 0.38Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy