Research map / Federated learning and privacy / Communication-efficient FL
Communication-efficient heterogeneous FL
Reduce communication via dynamic regularization, sparse training and heterogeneity-aware client sampling.
65 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Communication-efficient FL, keyword cluster “devices · edge · dynamic”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
19 papers in 2023–24 and 24 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 12.8% to 20.7%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2017 · Federated Multi-Task Learning (NeurIPS 2017 · 113 citations)
- 2020 · Federated Learning with Matched Averaging (ICLR 2020 · 97 citations)
- 2021 · Federated Learning Based on Dynamic Regularization (ICLR 2021 · 114 citations)
- 2022 · Federated Dynamic Sparse Training: Computing Less, Communicating Less, Yet Learning Better (AAAI 2022 · 97 citations)
- 2023 · FedGCN: Convergence-Communication Tradeoffs in Federated Training of Graph Convolutional Networks (NeurIPS 2023 · 16 citations)
- 2024 · Mixed-Precision Quantization for Federated Learning on Resource-Constrained Heterogeneous Devices (CVPR 2024 · 23 citations)
- 2025 · DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game (AAAI 2025 · 3 citations)
- 2026 · Tackling Resource-Constrained and Data-Heterogeneity in Federated Learning with Double-Weight Sparse Pack (AAAI 2026 · 1 citation)
Earliest papers here
- Federated Multi-Task Learning (NeurIPS 2017 · 113 citations)
- Federated Learning with Matched Averaging (ICLR 2020 · 97 citations)
- FedBoost: A Communication-Efficient Algorithm for Federated Learning (ICML 2020 · 69 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
CIFAR-10 · accuracy (higher is better)
4 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 89.3 | Delayed Gradient Averaging: Tolerate the Communication Latency for Federated Learning (NeurIPS 2021 · 22 citations) |
What has been tried
50 papers here report 496 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 9%, hurt 18%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| increase the MSB-pruning threshold ϵ | CIFAR10, α=0.5 · test accuracy | ϵ=0.01: 66.4, 68.0, 67.1, 64.7, 64.3; ϵ=0.02: 67.1, 70.0, 65.5, 64.7, 64.4; ϵ=0.03: 65.9, 67.1, 65.5, 64.7, 64.4 | Hurt | Mixed-Precision Quantization for Federated Learning on Resource-Constrained Heterogeneous Devices (CVPR 2024 · 23 citations) |
| Use modular updates instead of full federated updates | KITTI City · D1-all (%) | 1.42 -> 1.48 | Hurt | Federated Online Adaptation for Deep Stereo (CVPR 2024 · 16 citations) |
| add parameter updates using importance extracted from global thresholds | FMNIST · accuracy | 88.20±1.10 -> 89.21±0.25 | Helped | SpaFL: Communication-Efficient Federated Learning With Sparse Models And Low Computational Overhead (NeurIPS 2024 · 10 citations) |
| omit p_t-auxiliary communication | Synthetic, J=10 · regret | 10% reduction | Hurt | Federated Contextual Cascading Bandits with Asynchronous Communication and Heterogeneous Users (AAAI 2024 · 6 citations) |
| Replace random winning-coalition selection with score-based selection | FMNIST (0.1) · Total Score | 331.71 -> 516.29 | Helped | DualGFL: Federated Learning with a Dual-Level Coalition-Auction Game (AAAI 2025 · 3 citations) |
| add server-side momentum | MNIST and CIFAR-10 partitions across client counts and heterogeneity settings · relative wall-clock time to target accuracy | 0.72×; 1.19× | Mixed | FedCompass: Efficient Cross-Silo Federated Learning on Heterogeneous Client Devices Using a Computing Power-Aware Scheduler (ICLR 2024 · 2 citations) |
| use training-efficiency utility only instead of the joint utility | ResNet18@CIFAR10 · time to target accuracy | Hurt | WHALE-FL: Wireless and Heterogeneity Aware Latency Efficient Federated Learning over Mobile Devices via Adaptive Subnetwork Scheduling (AAAI 2025 · 1 citation) | |
| Stop transmitting server preconditioners and initialize client preconditioners from zero | CIFAR-100, GLD-23K, and FEMNIST · test accuracy | No clear effect | Efficient Adaptive Federated Optimization (NeurIPS 2025 · 1 citation) | |
| add BKD to FedMUD | FMNIST, SVHN, CIFAR-10, and CIFAR-100 (Non-IID-1 and Non-IID-2) · accuracy | 87.9(± 0.2) -> 88.2(± 0.1); 85.8(± 0.4) -> 86.7(± 0.6); 81.3(± 0.8) -> 84.8(± 0.5); 78.5(± 0 | Helped | The Panaceas for Improving Low-Rank Decomposition in Communication-Efficient Federated Learning (ICML 2025 · 0 citations) |
| Use initial construction without row-wise encoding or shared retrieval path | ML100K · upload communication cost | 0.17 -> 4.56 MB | Hurt | SecEmb: Sparsity-Aware Secure Federated Learning of On-Device Recommender System with Large Embedding (ICML 2025 · 0 citations) |
| Remove between-class scatter from classifier initialization | CIFAR-100 · accuracy | 57.1 -> 57.3 | Helped | Covariances for Free: Exploiting Mean Distributions for Training-free Federated Learning (NeurIPS 2025 · 0 citations) |
| increase server update ratio τ from 2 to 3 or 4 | CIFAR-10, Fashion-MNIST, CINIC-10, and CIFAR-100 test · accuracy | CIFAR-10: 77.86 -> 73.20, 69.40; Fashion-MNIST: 85.45 -> 85.28, 84.47; CINIC-10: 59.50 -> 55.75, 52.43; CIFAR-100: 32.16 | Hurt | Towards Straggler-Resilient Split Federated Learning: An Unbalanced Update Approach (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 14 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 8 of 14 papers
- CIFAR-100 · 7 of 14 papers
- MNIST · 3 of 14 papers
- FMNIST · 3 of 14 papers
- Fashion-MNIST · 2 of 14 papers
- FEMNIST · 2 of 14 papers
Baselines
- FedAvg · 8 of 14 papers
- FedPara · 2 of 14 papers
- FedBAT · 2 of 14 papers
- FedAdam · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 43 papers; from OpenAlex and the papers' affiliations).
- University of Chicago · 3 papers
- National University of Singapore · 2 papers
- Beijing University of Posts and Telecommunications · 2 papers
- Imperial College London · 2 papers
- Sony · 2 papers
- Argonne National Laboratory · 2 papers
- University of Virginia · 2 papers
- Yale University · 2 papers
- Seoul National University · 2 papers
- Tencent · 2 papers
Use it
Download BibTeX of 37 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Efficient local training in FL (77 papers)
- Decentralized peer-to-peer training (41 papers)
Elsewhere in the topic
- Federated optimization algorithms (70 papers)
- Client participation analysis (68 papers)
- Local drift correction (61 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).