- Rating
- 1
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper studies federated LoRA fine-tuning with partial participation, non-IID data and per-client rank budgets. Its method, BC-FedLoRA, has each selected client train two branches. The federated branch starts from the global adapter, and the proximal branch continues the client's own adapter with a penalty toward the global adapter. Their validation-accuracy gap is treated as a Gaussian observation of the client's benefit. A deployment rule thresholds the posterior at 0.85 to deploy the federated branch (share), the proximal branch (personalize) or the previous model (defer). Clients are selected by fixed priorities: capacity coverage, then participation counts, then label divergence. Experiments use Swin-T on CIFAR-100 and Tiny-ImageNet and DistilBERT on AG News and IMDB, with 12 clients, 5 seeds and seven baselines, including Ditto-LoRA. It reports the best mean client accuracy in 14 of 16 settings.
Critical strengths
None
Critical weaknesses
-
Limited novelty. The method is a published algorithm plus a deployment rule, and the paper's ablation finds no measurable effect from it. Specifically, the two branches and the upload are Ditto's. Eqs. 6–7 and the upload of every federated branch are Ditto's Algorithm 2 (Li et al., 2021). The paper cites Ditto only for the proximal term (l.245) and presents the paired branches as its own (l.229–230). MaxFL (Cho et al., TMLR 2024), not cited, already has each client pick a global-derived or a local model on held-out data.
-
The deployment rule is a fixed threshold. With the paper's prior and noise settings (, ), Eqs. 13–17 reduce to one test: a client shares when and personalizes when . Here is an observed benefit and the number of rounds the client has joined. The posterior variance depends only on , so the reported uncertainty ignores how much the observations vary.
Furthermore, replacing the rule changes accuracy by less than the variation across seeds. In Table 4, picking the branch at random costs 0.13–0.18 pp, and using only the latest observation costs 0.05–0.21. BC-FedLoRA's standard deviation over seeds is 0.45–1.81 (Tables 1–2).
-
The paper states no objective beyond the aim at l.176. "Long-term benefit" (Abstract; l.74, 134, 156) is never defined or measured. The method joins three separate problems: personalization (Eq. 17), client selection (Eqs. 2–5) and LoRA with a different rank per client (Eq. 18). The abstract says the decision controls what is shared and uses the resource budget (l.17–24, 87–89). In the method, every selected client uploads (l.311–312), and Eq. 17 has no resource term.
-
The metrics and baselines do not fit the method. The method measures benefit against the proximal branch (Eq. 8), but the negative-transfer rate (NTR) is measured against Local-only training. The paper itself says this reference drifts on small, skewed data. The four non-personalized aggregation baselines fall points below Local-only on the most skewed split (Table 1). FedAvg-LoRA with local fine-tuning, the method's own federated branch, is missing.
-
The paper is poorly presented and extremely difficult to read.
Questions for authors
-
Apart from the deployment rule, does any step differ from Ditto's Algorithm 2 run on LoRA?
-
Please report Table 4 with standard deviations over five seeds, and paired tests of the rule against random and latest-observation deployment.
-
Is a fraction or in percentage points, and how often is a defer decision caused by an exact tie? In percentage points, one disagreeing sample among about 38 validation samples crosses the threshold; as a fraction, a client needs a mean benefit above 0.30 after 13 observations.
-
Please add FedAvg-LoRA with local fine-tuning, and Ditto deploying the better of its two models on validation, with paired per-seed results to show the efficacy of the method.
-
What objective does the deployment rule optimize? Please also define NTR, and say why its reference is Local-only while the method's benefit uses the proximal branch.
-
Which model is evaluated for each baseline, and is the classification head aggregated? FedFomo-LoRA stays very close to Local-only in all vision settings: does its weighting collapse onto each client's own model?