AI Paper Atlas

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

Beyond Description: Federated Adaptation via Semantic-Visual Prototype Alignment

Jiarong Yang, Yuan Liu

ICML 2026 · Federated learning and privacy › Personalized and heterogeneous FL · approach: Federated prompt and LoRA tuning

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 contrastive semantic alignment on the server instead of FedAvg prototype averaging (ablation)FedSPA-G · 10-dataset mean, Dir(0.1)mean accuracy: 59.18 -> 69.68Improved
add contrastive semantic alignment on the server instead of FedAvg prototype averaging (ablation)FedSPA-P · 10-dataset mean, Dir(0.1)mean accuracy: 64.55 -> 77.47Improved
add global semantic prototypes with alignment to local-only visual prototypes (ablation)FedSPA-P · 10-dataset mean, Dir(0.1)mean accuracy: 72.82 -> 77.47Improved
naive FedAvg aggregation of prototypes versus local-only adaptation (ablation)FedAvg-based personalization · 10-dataset mean, Dir(0.1)mean accuracy: 72.82 -> 64.55Hurt
set stability regularizer weight lambda to 0 (no regularization) (ablation)FedSPA · 10-dataset average, RN50, 8-shot, Dir(0.1)average accuracyHurt
use very large lambda (over-regularize toward initialization) (ablation)FedSPA · 10-dataset average, RN50, 8-shot, Dir(0.1)average accuracyHurt
lambda in intermediate range 0.5 to 2.0 (ablation)FedSPA · 10-dataset average, RN50, 8-shot, Dir(0.1)average accuracyImproved
extend local training to 100 iterations in the final roundFedSPA-P · 10-dataset mean, ViT-B/16accuracy: 83.02 -> 84.91Improved
replace prompt tuning (pFedMoAP) with visual prototype adaptation for efficiencyFedSPA-P · efficiency, ViT-B/16local execution time: 83.60s -> 1.35sImproved
increase shots per class from 1 to 16FedSPA · 10-dataset average, RN50, Dir(0.1)accuracyImproved
add feature-level Gaussian DP noise sigma=0.10 (T=5)FedSPA-P · 10-dataset mean, RN50accuracy: 79.89 -> 78.80Hurt
add feature-level DP noise sigma=0.48 (T=5)FedSPA-P · 10-dataset mean, RN50accuracy: 79.89 -> 72.96Hurt

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.

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.

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