{"id":"5798c7eb-6cf7-4546-b9db-9cd05ee26228","arxiv_id":"2411.10595","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FedAli adds label-free local and global prototype layers, matched by optimal transport, to personalized federated learning, improving cross-client generalization most clearly on the HHAR activity recognition benchmark.","lead":"Researchers propose FedAli, a personalized federated learning method that adds prototype memory layers to each client model to align features across users while still adapting to each user's local data. It targets mobile sensing tasks like activity recognition, where models need to work for new users and new environments without sharing raw data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generalization metric tests only participating clients' test sets, so the central claim about robustness to unseen mobile scenarios is not supported by the reported experiments.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: the generalization metric treats test sets from participating clients as a valid proxy for unseen mobile environments. I agree with that assessment. The concern is load-bearing because the abstract and introduction promise robustness to unseen scenarios, while the metric used in Tables 1-2 only measures transfer to other participating clients. This is not a minor scope nit: FedAli's mechanism is built from prototypes derived from participating clients' training distributions, so the reported generalization numbers are precisely the wrong evidence for novel users, devices, or body positions. A leave-one-domain-out experiment would settle whether the proposed mechanism actually confers the claimed robustness. I do not see an internal inconsistency in the algorithm itself; the ALP layer, the Sinkhorn matching, and the FedAli aggregation are described with enough specificity to be reproduced, and the released source code is independent supporting evidence. Secondary issues such as the lack of seed-level variance in the main tables and hyperparameter tuning on HHAR reinforce the need for caution but are less central than the metric-scope problem. The reader's CONDITIONAL verdict is therefore appropriate, and my read does not change it.","tokens_in":16672,"tokens_out":4093,"duration_ms":48742,"concrete_test":"Run a leave-one-body-position-out evaluation on RealWorld: train FedAli, FedAvg, FedProx, and MOON on data from six of the seven on-body sensor positions, hold out the seventh, and evaluate the personalized client models on the held-out position. Repeat for all seven positions and report the mean held-out accuracy with seed-level error bars. If FedAli's generalization advantage over FedAvg is not reproduced under this protocol, the claim that it enhances robustness to unseen mobile scenarios is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1.4 defines the Generalization score as 'the average performance of each client model on the test sets of all participating clients.' Every test set therefore comes from a client whose data were used during training; in HHAR and RealWorld this means the same subjects, device models, and on-body sensor positions seen during training. The t-SNE analysis in Figures 4-5 does show a model from one dataset acting on a client from another dataset in the Combined setup, but that 'unseen' client is still a participant in federated learning, and the quantitative Generalization score never excludes a client, device, or body position from training. FedAli's inference-time alignment uses local prototypes learned from participating clients' training data, so it can only adapt toward distributions already represented by those prototypes. The motivating scenario in the introduction—a new user, a new phone model, or a new body position—is therefore not tested by the metric. If the central claim is about robustness to unseen mobile environments, the experiments provide at most evidence of improved transfer among known client distributions, not evidence of generalization to genuinely novel domains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces FedAli, a personalized federated learning strategy that adds an ALignment with Prototypes (ALP) layer to transformer-based models. The layer matches input embeddings to local and global prototype banks via Sinkhorn optimal transport, aligns embeddings with the matched prototypes through a gated linear unit, and updates local prototypes by EMA during training; the server aggregates prototypes by K-means clustering. The method is evaluated on three HAR datasets and two vision datasets, reporting personalization, generalization, and global scores, plus a pre-trained initialization study and ablation and sensitivity analyses. The abstract claims that FedAli 'significantly enhances client generalization while preserving strong personalization in heterogeneous settings.'","tokens_in":16912,"tokens_out":6317,"duration_ms":67344,"significance":"The proposed label-free prototype layer is a plausible and relatively lightweight mechanism for reducing representation drift in federated learning, and the paper has tangible strengths: the source code is released, the ablation study isolates the contribution of each component, the communication and computation overhead is reported, and the pre-training study addresses a practically relevant question. If the central claim were established, FedAli would be a useful drop-in component for mobile FL. However, as reported, the evidence supports only modest and inconsistent gains, and the generalization metric does not measure the 'unseen scenarios' that motivate the work. The contribution is therefore promising but not yet convincingly validated.","major_comments":[{"comment":"The 'Generalization score' is defined as the average performance of each client model on the test sets of all participating clients. This metric does not measure robustness to the unseen scenarios that the introduction motivates: a new user, a new phone model, or a new body position. No client, device, or on-body position is held out from training, and the t-SNE 'unseen' client in Section 5.3 is still a participant in the Combined federation. The central claim therefore needs either a leave-one-client-out or leave-one-device/position-out evaluation, or a revised claim that is explicitly limited to inter-client transfer among participating clients.","section":"Section 5.1.4 and Section 1"},{"comment":"The claimed significant generalization advantage is not consistently present across the six experimental settings. Concretely, the generalization gain over FedAvg is about +6.0 points on HHAR, but only +0.9 on RealWorld and +1.0 on Combined; on CIFAR10 with 20 clients FedAli is slightly behind FedAvg (43.14 vs. 43.29), and on CIFAR100 with 50 clients the gain is +0.9 points. The global score on RealWorld is below FedAvg (71.50 vs. 72.47). Only standard deviations over clients are reported, with no statistical significance tests or confidence intervals over repeated runs, so the abstract's 'significantly enhances' is not supported by the tables as presented.","section":"Tables 1 and 2"},{"comment":"The hyperparameters beta = 0.2 and the prototype counts were selected on the HHAR dataset using the same generalization metric that is later reported as the headline result in Table 1. This selection on the evaluation dataset can inflate the apparent HHAR advantage, and no validation protocol is described for the other datasets. Additionally, Table 5 shows that beta = 0.5 yields a higher generalization score (70.30 vs. 70.04) and the same global score (83.00) as beta = 0.2, so the claim that beta = 0.2 gives the 'best balance' is not clearly supported by the table.","section":"Section 6.1, Tables 5 and 6"},{"comment":"The pre-training section claims large gains ('more than 20%' in generalization and 'more than 11%' in global performance) and faster convergence, but these claims are made without a quantitative table or error bars, relying on learning curves that are not numerically summarized. For a load-bearing performance claim about convergence and pre-training benefits, the authors should report final numbers with run-level variability for all compared strategies.","section":"Section 5.4 and Figures 6 and 7"}],"minor_comments":[{"comment":"The text says 20 and 100 clients are generated for the vision datasets, but Table 2 reports experiments with 20 and 50 clients; the text and the table need to be reconciled.","section":"Section 5.1.2"},{"comment":"The personalization score is described as a macro F1 score, while the generalization and global scores are described as 'average performance' or 'accuracy'; the manuscript should state explicitly whether all three metrics use the same measure and how the macro averaging is performed.","section":"Section 5.1.4"},{"comment":"The displayed PFL objective is not mathematically well-formed: it defines F as a sum of local losses and then introduces '∀c min w_c f_c(w_c; D_c)' with a constraint on H(w_c; D), but the relationship between F, the inner minimization, and the constraint is unclear.","section":"Section 2, Equation (2)"},{"comment":"The ablation and sensitivity tables report plus-minus values from 'three random runs', but it is not stated whether the spread is over runs or over clients; this should be clarified because client-level heterogeneity is large in these datasets.","section":"Section 6"},{"comment":"The phrase 'unseen RealWorld client' is misleading: the client participates in the federated training process and is only unseen to the particular HHAR local model. The authors should qualify this as 'not locally seen' to avoid overstating the out-of-distribution nature of the experiment.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The paper would benefit from a clearly stated target venue; if the target is a machine learning venue, the lack of statistically significant gains on the vision benchmarks is a serious concern. The 'generalization' framing is likely to be challenged by reviewers: either leave-one-client-out experiments on HHAR and RealWorld should be added, or the claims should be explicitly restricted to transfer among participating clients. The method itself is interesting and the released code is a strength, so I see a viable revision path rather than a rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FedAli is a real contribution on the method side—the ALP layer with Sinkhorn optimal transport assignment to dual local and global prototype banks is new, label-free, and well-described. The code is public, the ablations are useful, and the pre-training experiments are a nice bonus. Keep it on your radar.\n\nThe soft spot is the evaluation of the central claim. The paper says FedAli 'significantly enhances client generalization' to unseen mobile scenarios, but the Generalization score (Section 5.1.4) averages each client model's accuracy on the test sets of all participating clients. Those test sets come from the same subjects, devices, and body positions that were used in training. So the metric measures transfer to known clients, not robustness to a new user, phone model, or sensor position. The stress-test note is right about this. The t-SNE analysis does show a model from one dataset acting on a client from another dataset, but that client is still a participant. For a claim about unseen environments, you need a true leave-one-client-out or leave-one-domain-out evaluation.\n\nSecond issue: the headline numbers are not consistent across benchmarks. On HHAR, FedAli beats FedAvg by 6 points in generalization (81.84 vs 75.85), which is impressive. On RealWorld and Combined, the gains are under 1.5 points, and FedAli's global score is actually lower than FedAvg's on RealWorld. On CIFAR-10/100, FedAli's generalization is essentially tied with or below FedAvg. So the 'significant' claim in the abstract is only really supported on HHAR. And the hyperparameters (β, prototype counts) were tuned on HHAR itself, which further weakens the headline result.\n\nWhat the paper does well: the method is clearly explained, the ablations decompose the contribution, and the authors are honest about some limitations (e.g., Walking vs Upstair in t-SNE). The communication/computation overhead analysis is a good addition. The novelty is solid: FedProto, FedPAC, and FedProc use per-class prototypes with auxiliary losses; FedAli's label-free, OT-based layer with k-means aggregation is genuinely different.\n\nBottom line: this deserves a serious referee. The evaluation needs a held-out domain split and seed-level variance before the generalization claim can be taken at face value, but the method and the write-up are worth engaging with.","headline":"Novel prototype layer and clear write-up, but the headline generalization claim overreaches the metric, which never holds out a client or domain.","tokens_in":17446,"tokens_out":2469,"would_cite":true,"duration_ms":30234,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FedAli claims that an alignment layer matching embeddings to dual prototype banks lets personalized federated learning generalize across mobile clients without sacrificing local accuracy.","keywords":["personalized federated learning","prototype alignment","optimal transport","Sinkhorn-Knopp","non-IID data","human activity recognition","mobile sensing","domain generalization"],"falsifier":"Train FedAli with an entire subject, device model, or body position held out from all clients, then evaluate on that held-out domain; if the cross-client generalization score does not improve over FedAvg when the test distribution was never present in training, the central claim fails.","tokens_in":16459,"feed_emoji":"📱","tokens_out":9322,"duration_ms":79067,"temperature":0.7,"pith_summary":"FedAli is a personalized federated learning strategy aimed at a property most PFL methods do not evaluate: a client model should stay accurate when it meets data from other clients' environments, not just on its own local data. The paper argues that existing prototype-based methods fail on this because they tie a single global prototype to each class label and never adapt the model's embeddings at inference time. FedAli instead inserts an alignment layer into each transformer block that matches embeddings to two prototype banks—local ones updated per client, global ones formed by clustering across clients—using a label-free optimal transport plan. The reported experiments on wearable human-activity-recognition and CIFAR benchmarks claim that this dual prototype alignment raises cross-client generalization substantially while keeping personalization at or near the best baseline. If that is right, FedAli is a drop-in layer that improves generalization in personalized mobile sensing without sacrificing the personal fit.","feed_headline":"Prototype alignment lifts federated learning generalization","feed_subtitle":"FedAli's ALP layer preserves per-client accuracy while improving performance on other clients' data.","key_machinery":"The central object is the ALignment with Prototypes (ALP) layer, placed before the multi-head attention in every encoder block of a small transformer. The layer receives a sequence of embeddings and computes a Sinkhorn-Knopp optimal transport plan between those embeddings and two prototype banks: local prototypes, updated by exponential moving average and acting as the client's memory, and global prototypes, produced on the server by K-means clustering of all clients' local prototypes each round. During training the transport plan assigns embeddings to global prototypes, the matched prototypes pass through a gated linear unit, and are blended into the embeddings with coefficient $\\beta$. During inference only local prototypes are matched, pulling unseen inputs toward the client's learned geometry. This label-free matching is what lets the layer redirect embeddings toward a shared representation without a per-class prototype.","core_discovery":"FedAli's central claim is that a personalized federated model can be made more transferable by aligning its embeddings to prototypes at two different moments: during training, shared global prototypes pull all clients' representations into a common coordinate system, and during inference, client-specific local prototypes pull novel inputs toward the geometry the client learned. The paper packages this as the ALignment with Prototypes (ALP) layer, inserted at the start of each transformer encoder block. The layer computes a Sinkhorn-Knopp optimal transport plan between the input embeddings and the two prototype banks, applies a gated linear unit to the matched prototypes, and blends them into the embeddings with a weight $\\beta$. Because the prototypes are feature-based rather than class-based, no labels are required, and the whole system can be pre-trained with a masked autoencoder. Empirically, the paper reports that on HHAR the generalization score rises from 75.85 percent for FedAvg to 81.84 percent while personalization rises from 96.88 to 98.15 percent, and on CIFAR-100 with 20 clients it reports the best personalization and generalization among the compared strategies.","pith_inferences":["A true test of the generalization claim would hold out an entire subject, device, or body position from all clients and measure accuracy on that never-seen domain; the paper's benchmark averages over test sets of participating clients, which may overstate transfer to genuinely novel environments.","The label-free prototype matching could be adapted to federated regression, forecasting, or anomaly detection, where class labels are not well defined; whether the Sinkhorn-Knopp plan remains stable with continuous targets is an open question.","Since global prototypes are formed by K-means over clients' local prototypes each round, the method's behavior under client churn and partial participation has not been characterized; a natural extension is to study convergence when only a subset of clients uploads prototypes.","The ALP layer's position before attention suggests a design space: placing it at different depths or using different prototype counts per layer may trade off personalization and generalization in ways the sensitivity analysis only begins to explore."],"forward_implications":["FedAli can be added to an existing transformer model and trained under standard FedAvg-style aggregation, with prototypes transmitted alongside weights, so the method is a layer-level upgrade rather than a new optimization framework.","Because prototype alignment needs no class labels, the same ALP layer can be pre-trained in a self-supervised way on unlabeled sensor streams; the paper reports that this pre-training cuts convergence time and yields more than a 20 percent generalization gain on HHAR.","Client models trained with FedAli should transfer better across users, devices, and on-body positions, since the t-SNE evidence shows same activities from different datasets mapped into a shared region of embedding space.","The personalization-generalization trade-off is relaxed: global prototypes anchor training embeddings across clients while local prototypes anchor inference to the client's own distribution, so a client need not sacrifice local accuracy to gain cross-client generalization.","The extra cost is bounded: prototypes are non-trainable, inference uses only local prototypes, and local training time is 1.09 times FedAvg, faster than contrastive baselines such as MOON and FedPAC, though communication overhead is 1.84 times FedAvg."],"supporting_citations":[{"why":"Defines FedAvg, the aggregation baseline FedAli extends and against which generalization gains are measured.","marker":"[38]"},{"why":"Class-prototype federated learning baseline; FedAli's main point of departure because it removes label dependence and uses multiple prototypes per layer.","marker":"[54]"},{"why":"Feature-alignment and classifier-collaboration PFL baseline that FedAli compares against on HAR and vision.","marker":"[61]"},{"why":"Contrastive FL baseline that the empirical comparison uses as a strong prior art.","marker":"[26]"},{"why":"Sinkhorn-Knopp algorithm that supplies the fast optimal transport plan for matching embeddings to prototypes.","marker":"[12]"},{"why":"Source of the sharpness parameter and iteration count used to stabilize the Sinkhorn-Knopp alignment.","marker":"[7]"},{"why":"HHAR dataset used to create heterogeneous clients across smartphone models for the HAR experiments.","marker":"[51]"},{"why":"RealWorld dataset used for the multi-position wearable sensor experiments.","marker":"[52]"},{"why":"HART, the lightweight transformer backbone into which the ALP layer is inserted.","marker":"[15]"},{"why":"Masked autoencoder pre-training scheme used to initialize HART and prototypes without labels.","marker":"[18]"}],"fun_headline_variants":["FedAli's dual prototypes cut drift and boost generalization","Label-free prototype alignment improves federated models","Shared and local prototypes enhance federated learning","Prototype layers personalize and generalize in FL","Prototype alignment lifts FL generalization with no labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each client model's averaged accuracy on the test sets of all participating clients is a valid proxy for performance in genuinely unseen mobile environments; if that proxy is weak, the claimed generalization gains are not established.","fun_headline_variants_meta":{"raw":{"variants":["FedAli's dual prototypes cut drift and boost generalization","Label-free prototype alignment improves federated models","Shared and local prototypes enhance federated learning","Prototype layers personalize and generalize in FL","Prototype alignment lifts FL generalization with no labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1442,"prompt_tokens":959,"completion_tokens":483,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":414}},"tokens_in":575,"tokens_out":483,"duration_ms":5423,"temperature":1.0,"reasoning_tokens":414,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:31:36.504824+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FedAli with an entire subject, device model, or body position held out from all clients, then evaluate on that held-out domain; if the cross-client generalization score does not improve over FedAvg when the test distribution was never present in training, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Class-prototype federated learning baseline; FedAli's main point of departure because it removes label dependence and uses multiple prototypes per layer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Feature-alignment and classifier-collaboration PFL baseline that FedAli compares against on HAR and vision."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source of the sharpness parameter and iteration count used to stabilize the Sinkhorn-Knopp alignment."},{"cited_title":"Sztyler and H","cited_arxiv_id":null,"evidence_quote":"RealWorld dataset used for the multi-position wearable sensor experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HART, the lightweight transformer backbone into which the ALP layer is inserted."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Masked autoencoder pre-training scheme used to initialize HART and prototypes without labels."}],"review_version":1}