Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FedQS: Optimizing Gradient and Model Aggregation for Semi-Asynchronous Federated Learning
Authors: Yunbo Li, Jiaping Gui, Zhihang Deng, Fanchao Meng, Yue Wu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on computer vision, natural language processing, and real-world tasks demonstrate that Fed QS achieves the highest accuracy, attains the lowest loss, and ranks among the fastest in convergence speed, outperforming state-of-the-art baselines. Our work bridges the gap between aggregation strategies in SAFL, offering a unified solution for stable, accurate, and efficient federated learning. The code and datasets are available at https://github.com/bkjod/Fed QS_. |
| Researcher Affiliation | Academia | Yunbo Li , Jiaping Gui , Zhihang Deng, Fanchao Meng, Yue Wu , School of Computer Science, Shanghai Jiao Tong University, Shanghai, China EMAIL |
| Pseudocode | No | The paper describes the |
| Open Source Code | Yes | The code and datasets are available at https://github.com/bkjod/Fed QS_. |
| Open Datasets | Yes | We evaluate Fed QS on three task types in SAFL: Computer Vision (CV) using Res Net-18 [38] on CIFAR-10 [39], Natural Language Processing (NLP) with LSTM [40] on Shakespeare [1], and Real-World Data (RWD) using FCN on UCI Adult [41]. |
| Dataset Splits | Yes | Each participant will split their local dataset into training and validation sets with an 8:2 ratio. (CIFAR-10, UCI Adult)... Each participant will split their local dataset into training and validation sets with a 9:1 ratio. (Shakespeare) |
| Hardware Specification | Yes | All experiments were conducted on a Linux system (Ubuntu 22.04 LTS) using an Intel Xeon Platinum 8468 Processor and an NVIDIA H100 80GB HBM3 GPU, with system memory capped at 20 GB per run. |
| Software Dependencies | Yes | The software stack included Python 3.8.0, Py Torch 2.1.0, and Torchvision 0.16.0, all within a Conda environment. |
| Experiment Setup | Yes | The default settings in our experiments are as follows: number of participants N = 100, initial local learning rate η0 = 0.1, learning rate bounds α = 0.001 and β = 0.2, learning rate adaptation rate a = 0.002, initial momentum m0 = 0.1, momentum adaptation parameter k = 0.2, global epochs T = 400, local epochs E = 2, minimum number of updates for aggregation K = 10, gradient clipping threshold Gc = 20, and momentum clipping threshold θ = 0.9. |