AI Paper Atlas

Research map / Federated learning and privacy / Personalized and heterogeneous FL

Heterogeneous Customizable Personalized Federated Fine-Tuning Approach for Large Language Models

xin tong, Baojiang cui

ICML 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Personalization with global-local trade-off

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
add our scaling factor (personalization rank proportion) instead of none (ablation)Het-CPFLoRA · Dataset 2 (heterogeneous)Avg-Gen: 44.38 -> 47.49Improved
add our scaling factor instead of FedDPA's (ablation)Het-CPFLoRA · Dataset 2 (heterogeneous)Avg-Gen: 46.15 -> 47.49Improved
add our scaling factor instead of none (ablation)Het-CPFLoRA · Dataset 2 (heterogeneous)Avg-Per: 57.01 -> 58.91Improved
rear vs front position decoupling of the personalized subspace (ablation)Het-CPFLoRA · Dataset 2 (rp=4, r=4, homogeneous)Avg-Per / Avg-Gen: 52.27 -> 52.88No clear effect
increase personalized subspace rank share from 1:7 to 4:4 (ablation)Het-CPFLoRA · Dataset 1Avg-Per: 54.22 -> 57.64Improved
increase personalized subspace rank share from 1:7 to 4:4 (ablation)Het-CPFLoRA · Dataset 1Avg-Gen: 43.13 -> 38.63Hurt
add dynamic weighted inference (Our-auto vs Our) (ablation)Het-CPFLoRA · Dataset 1 (All Homogeneous)Avg-Gen: 38.48 -> 54.72Improved
add dynamic weighted inference (Our-auto vs Our) (ablation)Het-CPFLoRA · Dataset 2 (All Homogeneous)Avg-Per: 42.12 -> 52.27Improved
replace FedDPA's embedding-similarity weighting with our dynamic weighting (ablation)Het-CPFLoRA · Dataset 1 (homogeneous)Avg-Per: 56.77 -> 59.12Improved
use single-adapter decoupled LoRA (Our-auto) vs FedDPA-Auto dual adapterHet-CPFLoRA · Dataset 2 (All Homogeneous)Avg-Per: 45.86 -> 52.27Improved
scale to 48 clients with Het-CPFLoRA vs pFedSeqHet-CPFLoRA · Dataset 1 (48 clients)Avg-Per: 53.77 -> 59.36Improved
use Het-CPFLoRA on Qwen2.5-1.5BHet-CPFLoRA · Dataset 2 (Qwen2.5-1.5B, heterogeneous)Avg-Per: +6.64%Improved

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