{"id":"45e37998-6828-4c48-9db1-205e39708370","arxiv_id":"2607.09695","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A federated framework that adversarially separates causal vs. spurious features, reliability-weights class prototypes, and contrastively aligns them, reporting SOTA accuracy on Office-10, Digits, and PACS.","lead":"FedCausal-Dyn is a federated learning method that splits features into domain-invariant “causal” parts and client-specific noise, then aggregates class prototypes with reliability weights. It reports higher and more stable accuracy than prior FL methods on three domain-generalization image benchmarks used as proxies for feature drift.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The SOTA claim rests on static multi-domain benchmarks that never instantiate the paper’s own dynamic (temporal) feature-drift model.","rationale":"The Reader correctly flags that the causal-separation story (Eqs. 2–4) lacks identification guarantees and that the paper itself admits failure under entanglement (§7). That is a real soft spot. However the more load-bearing gap for the strongest claim is the mismatch between the dynamic formulation (§3.1) and the purely static experimental regime (§4). Without a temporal-drift test, one cannot tell whether the reliability-weighted aggregation and L_Reg actually address the problem the abstract advertises. The empirical ranking on static DG benchmarks remains plausible, so the verdict stays CONDITIONAL rather than REJECT; the concrete temporal-drift re-run would decide whether the claim can be upgraded or must be further qualified.","tokens_in":18088,"tokens_out":493,"duration_ms":6213,"concrete_test":"Re-run the three benchmarks under a controlled temporal-drift protocol: after every R rounds, apply a fixed domain shift (e.g., progressive style transfer or Gaussian feature noise whose variance grows with t) to each client’s local data while keeping labels fixed. Compare FedCausal-Dyn against FedPall and ADCOL under identical drift schedules. If the average accuracy gap shrinks below 1 % or the rank order changes, the dynamic-drift justification for the SOTA claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that FedCausal-Dyn solves dynamic feature drift—explicitly formulated in §3.1 as time-evolving domains D_n^t with non-stationary P_n^t(x|y)—and thereby attains SOTA averages (68.95/89.34/61.26) via causal separation + reliability-weighted prototypes. Yet every experiment (§4) uses the standard static leave-one-domain-out protocol on Office-10, Digits and PACS; no client’s distribution is altered across communication rounds. Consequently the “dynamic” prototype update (Eq. 6) and the reliability weights never face genuine temporal non-stationarity. If the reported gains disappear (or reverse) once true concept drift is injected, the causal-dynamic story no longer explains the numbers and the SOTA claim does not follow from the stated mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes FedCausal-Dyn, a federated learning framework for dynamic feature drift. It models client data as time-evolving domains D_n^t (§3.1) and introduces three components: (i) causal–domain feature separation via dual projection heads C(·), S(·) and adversarial losses L_Causal, L_Inv, L_Spur (Eqs. 2–4); (ii) reliability-weighted dynamic prototype aggregation (Eqs. 5–6); and (iii) causal-feature guided collaborative regularization L_Reg combining prototype contrastive alignment with domain invariance (Eq. 7), plus a privacy-enhanced causal mixup (Eqs. 8–9). Experiments on Office-10, Digits, and PACS report state-of-the-art average accuracies (68.95%, 89.34%, 61.26%) and lowest average standard deviation versus a broad baseline suite (Tables 1–4), with ablations (Tables 5–8) attributing the largest gains to L_Reg and supporting each module.","tokens_in":18419,"tokens_out":1531,"duration_ms":20769,"significance":"If the method genuinely improved federated learning under temporal non-stationarity, the contribution would be practically important for settings such as financial technology that the abstract highlights. The empirical package is carefully executed for static multi-domain federated domain generalization: three seeds, many baselines, component ablations, reliability-weight variants, and λ sensitivity. Those results would still be a useful incremental advance over prototype and adversarial FL methods (FedProto, ADCOL, MOON, FedBN). The significance of the stated “causal-dynamic” claim, however, is limited by the gap between the temporal formulation and the static evaluation protocol; closing that gap (or reframing) is what would make the paper’s central story load-bearing rather than aspirational.","major_comments":[{"comment":"§3.1 formulates the problem as dynamic feature drift: domains D_n^t with non-stationary P_n^t(x|y) over communication rounds t, and Eq. (1) explicitly takes an expectation over t. The abstract and introduction contrast this with prior work that “assume[s] static drift.” Yet §4 evaluates only the standard static multi-domain leave-one-domain-out protocol on Office-10, Digits, and PACS; no client’s P(x|y) is altered across rounds. Consequently the “dynamic” update of g_k^t (Eq. 6) and the reliability weights ω_n,k^t never face genuine temporal concept drift. The SOTA averages in Tables 1–4 therefore do not substantiate the paper’s central claim as stated. Either inject controlled temporal drift (e.g., rotating domain statistics or gradual style shifts across rounds) and re-evaluate, or reframe the contribution as federated domain generalization / static feature drift and revise the abstrac","section":null},{"comment":"§3.2.1 and Eqs. (2)–(4) treat the dual-head adversarial construction as separating domain-invariant causal features z_c from spurious z_s. No identification argument, intervention, or causal metric is provided; success is inferred from held-out accuracy and a qualitative t-SNE case study (§6.2). The paper itself notes in §7 that separation “may be less effective” when factors are highly entangled. If z_c is only partially purified, reliability-weighted prototypes and L_Reg (Eq. 7) align to the wrong representation, so the mechanistic explanation of the gains in Tables 1–5 does not follow. At minimum, add quantitative checks (domain predictability of z_c vs z_s, label predictability of each stream, or controlled spurious-correlation synthetic data) so that the “causal” label is empirically justified rather than terminological.","section":null},{"comment":"Table 4 and §4.2.4 claim overall rank 1.00 and “most stable results,” but the experimental suite never includes a true non-stationary baseline or a dynamic-drift ablation of the reliability weights. Without that, the attribution of stability to “reliable and dynamic prototype aggregation” under dynamic drift (§4.2.5, §6.1) remains untested. A minimal load-bearing addition is a controlled non-stationarity experiment (or an explicit statement that all reported gains are under static multi-domain heterogeneity only).","section":null}],"minor_comments":[{"comment":"Related Work and Introduction cite a large number of concurrent/arXiv works (including many from overlapping author groups) that are only loosely connected to feature drift; tighten to the methods that actually compete on the same problem (FedProto, ADCOL, MOON, FedBN, FRAug, FedCiR).","section":null},{"comment":"§3.2.2: reliability weight ω is defined as either validation accuracy or inverse intra-class variance, but the main tables do not state which was used for the reported SOTA numbers; Table 6 shows Acc is best—make the default explicit in §4.1.","section":null},{"comment":"Eq. (9) is a constrained min–max over a binary mask m; implementation details (how m is optimized, whether it is continuous relaxation, cost per round) are missing and should be added for reproducibility.","section":null},{"comment":"Figure 1 is described but the manuscript text does not specify encoder backbone, local epochs, client sampling rate, or communication rounds used for Tables 1–3; add a hyperparameter table.","section":null},{"comment":"Notation: D_n^t is used for both the domain and the local dataset; distinguish distribution vs. sample set. Also L_Inv appears both as a standalone adversarial loss (Eq. 3) and inside L_Reg (Eq. 7)—clarify whether the same discriminator is shared.","section":null},{"comment":"§6.3 lists “potential extended analyses” (partial participation, label noise) without results; either run them or move to future work to avoid implying unfinished experiments.","section":null},{"comment":"Typos/style: “FedCausal-Dyn” sometimes loses bold/spacing; “PerfedAvg” vs “Per-FedAvg”; duplicate FedProto citations (Tan et al. 2022a/b/c).","section":null}],"recommendation":"major_revision","confidential_remarks":"The citation graph is unusually dense with recent arXiv preprints from a small set of overlapping groups, which can read as citation padding and may raise novelty questions relative to FedProto/ADCOL-style prototype and adversarial FL. The technical core is a reasonable combination of known pieces; the main editorial risk is overselling “dynamic” and “causal” without matching experiments. If the authors reframe cleanly to static federated DG and add the separation diagnostics, the paper is within scope for a solid systems/methods venue; if they insist on the dynamic claim without new experiments, I would lean reject on mismatch of claim and evidence."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: this is a competent systems paper that stacks known pieces—dual projection heads + GRL-style domain adversarial training, reliability-weighted class prototypes, contrastive alignment to those prototypes, and a privacy mixup—and reports consistent first-place averages on Office-10, Digits, and PACS (68.95 / 89.34 / 61.26) with the lowest variance among the listed baselines. The ablations (Tables 5–8) are clean: collaborative regularization G is the biggest lever, removing the causal/spurious split hurts, and accuracy-based reliability beats uniform averaging. That part is real and useful for people who actually ship multi-site vision models.\n\nWhat is new is the packaging and the reliability-weighted server step, not a new theory of causality or of temporal drift. The math is standard (cross-entropy, GRL, InfoNCE-style, weighted mean of prototypes). The citation pattern is noisy—lots of loosely related graph/diagnosability and group-adjacent papers—but the core FL baselines are present and the ranking is plausible.\n\nThe soft spot that matters is the gap between claim and protocol. Section 3.1 formulates time-evolving domains D_n^t and non-stationary P_n^t(x|y), yet every experiment is ordinary static leave-one-domain-out. The “dynamic” update (Eq. 6) and the reliability weights never see genuine concept drift across rounds. The paper itself flags that the causal/spurious split can fail when factors are entangled. So the SOTA numbers are real under static multi-domain FL; they do not yet demonstrate the causal-dynamic story that the title and abstract sell. Free parameters (λ, γ, η, τ, ω definition) are ordinary and sensitivity is shown for λ.\n\nWho it is for: FL practitioners who need a stronger prototype + invariance recipe and who will re-run under true non-stationarity. It deserves a serious referee—empirical ranking and ablations are enough for that—but the causal-dynamic language should be dialed back or the temporal protocol added. I would engage with the engineering, not the theory claim as written.","headline":"Solid FL engineering package with real ablations and modest SOTA gains, but the “dynamic causal” framing is not what the experiments test.","tokens_in":19029,"tokens_out":547,"would_cite":false,"duration_ms":5268,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Separating causal from spurious features and weighting class prototypes by reliability lets federated models keep learning when feature distributions change across clients and over time, reaching the highest average accuracy on three domain","keywords":["federated learning","dynamic feature drift","causal feature separation","prototype aggregation","domain generalization","adversarial training","non-stationary data"],"falsifier":"Re-run the three benchmarks with the causal and spurious heads forcibly collapsed into a single adversarial representation (or with deliberately entangled synthetic factors); if the reported accuracy and stability gaps over the strongest baselines disappear, the central claim that the separation is what drives the gains is falsified.","tokens_in":18943,"feed_emoji":"📈","tokens_out":933,"duration_ms":13235,"temperature":0.7,"pith_summary":"Federated learning usually keeps raw data on each client, but when the same class looks different across clients and those differences keep changing over time, the shared model degrades. This paper claims that treating the problem as dynamic feature drift—not static non-IID data—and solving it with a causal-dynamic pipeline restores accuracy and stability. The method first peels domain-invariant causal features away from client-specific spurious ones with two projection heads and adversarial training, then builds global class prototypes only from the causal stream and weights each client’s prototype by how reliable that client looks. A single collaborative regularizer then pulls local causal features toward those reliable global prototypes while still enforcing domain invariance. On Office-10, Digits, and PACS the resulting system reports the best average accuracies and the lowest run-to-run variance among a large set of federated and domain-generalization baselines, with ablations showing each piece contributes.","feed_headline":"Causal feature split tops federated accuracy under drift","feed_subtitle":"Reliability-weighted prototypes of domain-invariant features beat static-alignment baselines on three benchmarks","key_machinery":"Causal-domain feature separation: two lightweight projection heads on a shared encoder, trained with causal prediction, domain-invariance (via gradient reversal), and domain-capturing losses so that only the causal stream is used for reliability-weighted prototype aggregation and contrastive alignment.","core_discovery":"The paper establishes that dynamic feature drift in federated learning can be handled by explicitly disentangling domain-invariant causal features from domain-specific spurious features, then aggregating only the causal class prototypes with reliability-aware weights and regularizing local models against those prototypes. When this causal-dynamic loop is closed, the global model achieves state-of-the-art average top-1 accuracy (68.95 % Office-10, 89.34 % Digits, 61.26 % PACS) and the most stable results across three standard federated domain-generalization benchmarks.","pith_inferences":["The reliability weight itself could be treated as a first-class federated signal and reused for client selection or early stopping, even outside the causal-separation setting.","If the separation works under true concept drift rather than only style/domain shift, the same pipeline may transfer to streaming recommendation or fraud models where both features and labels slowly change.","The privacy-enhanced mixup step is largely orthogonal; it could be swapped for stronger cryptographic or differential-privacy mechanisms without rewriting the causal core."],"forward_implications":["Federated systems facing non-stationary streams (e.g., changing market or sensor conditions) can keep a usable global model without sharing raw features or synthetic data.","Reliability-weighted prototype aggregation supplies a practical filter against low-quality or temporarily unreliable client updates.","The same causal-stream prototypes can serve both inter-client alignment and domain-invariance, reducing the need for separate ad-hoc regularizers.","Ablation patterns imply that removing the collaborative regularizer hurts more than removing any other single component, so future designs should treat prototype contrastive alignment as primary."],"fun_headline_variants":["Causal disentangling lifts federated accuracy under dynamic drift","Reliability-weighted causal prototypes top three federated benchmarks","Separating invariant from spurious features stabilizes federated drift","Causal-dynamic aggregation yields highest stable federated accuracy","Adversarial domain split enables robust prototypes under feature drift"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that a pair of projection heads plus domain-adversarial training can cleanly separate the features that truly predict the label from the features that only mark which client the data came from; if that separation fails, the prototypes and regularizer target the wrong representation.","fun_headline_variants_meta":{"raw":{"variants":["Causal disentangling lifts federated accuracy under dynamic drift","Reliability-weighted causal prototypes top three federated benchmarks","Separating invariant from spurious features stabilizes federated drift","Causal-dynamic aggregation yields highest stable federated accuracy","Adversarial domain split enables robust prototypes under feature drift"]},"model":"grok-4.5","effort":"low","cost_usd":0.004592,"raw_usage":{"total_tokens":1360,"prompt_tokens":796,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":45920000,"prompt_tokens_details":{"text_tokens":796,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":484,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":796,"tokens_out":80,"duration_ms":4390,"temperature":1.0,"reasoning_tokens":484,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T17:35:14.162339+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the three benchmarks with the causal and spurious heads forcibly collapsed into a single adversarial representation (or with deliberately entangled synthetic factors); if the reported accuracy and stability gaps over the strongest baselines disappear, the central claim that the separation is what drives the gains is falsified.","supporting_citations":[],"review_version":1}