Research map / Federated learning and privacy / Personalized and heterogeneous FL
Federated Sketching LoRA: A Flexible Framework for Heterogeneous Collaborative Fine-Tuning of LLMs
Wenzhi Fang, Dong-Jun Han, Liangqi Yuan, Seyyedali Hosseinalipour, Christopher Brinton
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use Random-k sketching instead of spectral-norm importance-based sketching (ablation) | FSLoRA · Commonsense reasoning | average accuracy: 75.4 -> 79.1 | Improved |
| use Random-k sketching instead of magnitude-sum importance-based sketching (ablation) | FSLoRA · Commonsense reasoning | average accuracy: 73.6 -> 79.1 | Improved |
| add sketching under a fixed upload budget (vs plain Federated LoRA) (ablation) | Federated LoRA · GLUE, RoBERTa, ranks 2 to 64 | testing accuracy | Improved |
| increase global LoRA rank while keeping client submatrix rank fixed at 8 (ablation) | FSLoRA · Commonsense reasoning, LLaMA-3.2-3B | testing accuracy | Improved |
| lower the sketching ratio from 1 toward 0.125 at global rank 64 (ablation) | FSLoRA · Commonsense reasoning, LLaMA-3.2-3B | testing accuracy | Hurt |
| move from normal to heavy-tail client rank distribution (ablation) | all methods · Commonsense reasoning, 50 clients | average accuracy: 79.20 -> 79.10 (FSLoRA) | Hurt |
| increase data heterogeneity from Dir(1) to Dir(0.1) (ablation) | all methods · Commonsense reasoning, 50 clients | average accuracy: 79.35 -> 79.10 (FSLoRA) | Hurt |
| use FSLoRA instead of FlexLoRA (truncated SVD aggregation) | heterogeneous federated LoRA · Commonsense reasoning, heavy-tail ranks, 50 clients | average accuracy: 75.59 -> 79.10 | Improved |
| use FSLoRA instead of HeteroLoRA on RoBERTa GLUE | heterogeneous federated LoRA · GLUE average | average accuracy: 79.6 -> 83.3 | Improved |
| use FSLoRA instead of FlexLoRA on LLaMA-3.2-3B | heterogeneous federated LoRA · Commonsense reasoning average | GPU hours: 68.3h -> 44.3h | Improved |
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.
- RAVAN — Ravan: Multi-Head Low-Rank Adaptation for Federated Fine-Tuning (NeurIPS 2025) · better on 2 of 2 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
- Ravan: Multi-Head Low-Rank Adaptation for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.92)
- Robust Federated Finetuning of LLMs via Alternating Optimization of LoRA (NeurIPS 2025 · similarity 0.91)
- Adaptive LoRA Experts Allocation and Selection for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.90)
- Beyond Parameter Aggregation: Semantic Consensus for Federated Fine-Tuning of LLMs (NeurIPS 2026 · similarity 0.90)
- FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations (NeurIPS 2024 · similarity 0.90)
- Sketch to Adapt: Fine-Tunable Sketches for Efficient LLM Adaptation (ICML 2025 · similarity 0.89)
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.
