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..
Layer-wise Update Aggregation with Recycling for Communication-Efficient Federated Learning
Authors: Jisoo Kim, Sungmin Kang, Sunwoo Lee
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our extensive empirical study demonstrates that the update recycling scheme significantly reduces the communication cost while maintaining model accuracy. For example, our method achieves nearly the same AG News accuracy as Fed Avg, while reducing the communication cost to just 17%. We evaluate the performance of Fed LUAR 2 using representative benchmark datasets: CIFAR-10 [19], CIFAR-100, FEMNIST [4], and AG News [42]. We first compare Fed LUAR to several state-of-the-art communication-efficient FL methods: Look-back Gradient Multiplier [2], Fed PAQ [31], Fed Para [12], Prune FL [13], Fed Dropout Avg [8], and Fed BAT [23]. |
| Researcher Affiliation | Academia | 1Inha University Incheon, Republic of Korea 2University of Southern California Los Angeles, CA, USA EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Layer-wise Update Aggregation with Recycling (LUAR) Algorithm 2 Federated Learning with Layer-wise Update Aggregation with Recycling (Fed LUAR) |
| Open Source Code | Yes | The code is publicly available at https://github.com/swblaster/ Fed LUAR, and the repository link is also provided at the bottom of page 2 of the paper for reference. |
| Open Datasets | Yes | Datasets We evaluate the performance of our proposed method on representative benchmarks: CIFAR-10 [19], CIFAR-100, FEMNIST [4], and AG News [42] |
| Dataset Splits | Yes | To generate non-IID datasets, we use label-based Dirichlet distributions with α = 0.1, which indicates highly non-IID conditions. Data Heterogeneity For IID datasets, we simulate non-IID settings using Dirichlet distributions. The concentration coefficient α is set to 0.1 for CIFAR-10/100 and 0.5 for AG News. |
| Hardware Specification | Yes | All experiments are conducted on a GPU cluster which has 2 NVIDIA A6000 GPUs per machine. |
| Software Dependencies | Yes | We use Tensor Flow 2.15.0 for training and MPI for model aggregations. |
| Experiment Setup | Yes | All our experiments are conducted on a GPU cluster that contains 2 NVIDIA A6000 GPUs per machine. We use Tensor Flow 2.15.0 for training and MPI for model aggregations. All individual experiments are performed at least three times, and the average accuracies are reported. The total number of clients is 128 and randomly chosen 32 clients participate in every communication round. We use mini-batch SGD with momentum (0.9) as the local optimizer. Table 6 shows the hyper-parameter settings for all our experiments, used not only for our method but also for other SOTA methods. |