{"id":"3bf612fa-daeb-4f28-927c-0ac9be50ea98","arxiv_id":"2608.02222","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CRIP improves one-shot federated learning by matching and injecting only the most compatible feature channels from other clients' models, without extra training or public data.","lead":"CRIP is a one-shot federated learning method that shares feature extractors, not parameters, and injects a few well-matched feature channels from other clients into each client's model. It reports large accuracy gains on domain-shifted image benchmarks, though it does not beat all baselines everywhere.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm selects argmax channels without enforcing the τ required by Eq. (17); theoretical guarantee may not cover actual injections.","rationale":"The core promise of CRIP is that CKA-based channel matching identifies semantically compatible source features, so residual injection adds signal rather than noise. The only formal support for this is Eq. (17), stated as a sufficient condition for the high-alignment region s≥τ. But the algorithm never defines τ; it simply takes the argmax over source channels for each target channel. Thus the selected set is not guaranteed to lie in the region where the assumption holds. The paper's own quantile analysis suggests many scores are moderate, and the validation in §V-C uses top/bottom percentiles rather than the actual selected pairs, so it does not close this gap. If low-score argmax pairs are injected, the denoising argument following Eq. (21) has no basis, and negative transfer is possible. This is a correctness risk in the central mechanism, distinct from the benchmark-overclaim issue. The reader's verdict is already conditional; this concern reinforces that conditionality without changing the verdict category, so I recommend UNCHANGED. Agreement is partial because the reader identified Eq. (17) but not the missing τ in the algorithm itself.","tokens_in":16255,"tokens_out":9129,"duration_ms":62630,"concrete_test":"Re-run CRIP with logging of every selected (target-channel, source-channel) CKA score on DomainNet, PACS, and Office-Home. Compute the fraction of selected channels whose score falls below candidate thresholds τ ∈ {0.3, 0.5, 0.7}; then run a variant that discards below-τ source channels and rescales α by the number of retained sources. If the thresholded variant matches or exceeds the reported no-threshold accuracy, the actual selection is not covered by Eq. (17), and the reported gains rely on channels outside the theoretically justified region. If thresholding degrades accuracy, then Eq. (17) is not the operative sufficient condition and the theoretical explanation needs revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CRIP's central justification is Eq. (17): for s ≥ τ, higher CKA alignment implies lower expected semantic discrepancy, so injection is denoising. But the method in §III-B selects, for each target channel, the single argmax channel from each source client, with no threshold on the CKA score. The theory in §IV-B explicitly restricts to the 'high-alignment region, where s≥τ,' yet the algorithm never applies τ; any channel that is the best match, however low its score, is injected. The CKA quantiles in §V-E (q0.5=0.314, q0.9=0.507, q0.99=0.685) suggest many argmax matches fall in a moderate range where Eq. (17) is not assumed to hold. The validation in §V-C compares top-1% and bottom-1% CKA regions, not the actual argmax-selected set, so it does not demonstrate that selected channels satisfy the sufficient condition. If low-scoring argmax channels carry domain-specific noise, residual injection can hurt—consistent with the Office-Home results, where CRIP underperforms FOL. Thus the claim that CRIP 'selectively fuses only the most compatible features' is not established: relative argmax compatibility is not the same as absolute compatibility above τ.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CRIP, a one-shot federated learning method for domain-heterogeneous clients. Each client uploads only its feature extractor; the server broadcasts all extractors back. For a target client, CRIP computes channel-wise CKA between target and source feature maps on a small local calibration batch, selects the highest-CKA channel from each source model for every target channel, and fuses the aligned source features into the target feature map via residual injection, leaving all parameters unchanged. Experiments on DomainNet, PACS, and Office-Home under a few-shot setting report strong average accuracies, and a theoretical section decomposes features into semantic and residual components to argue that the fusion is denoising under a CKA-based sufficient condition.","tokens_in":16591,"tokens_out":7422,"duration_ms":51903,"significance":"If the empirical claims could be substantiated, CRIP would be a valuable contribution: it is a representation-space, public-data-free approach to one-shot FL, and the channel-level CKA alignment idea is simple, training-free, and potentially generalizable. The paper also includes a useful ablation (CRIP-noA), sensitivity analyses, a scalability study, and a membership-inference evaluation, which are strengths. However, the central claim of consistent state-of-the-art performance is contradicted by the paper's own Table I; the theoretical guarantee in Section IV does not cover the implemented selection rule; and the main hyperparameter is tuned on the same benchmarks used for evaluation. The contribution is promising but requires substantive revision before its claims can be accepted.","major_comments":[{"comment":"Eq. (17) assumes a sufficient condition only for s ≥ τ, and the text in §IV-B explicitly says 'Channels selected only from the high-alignment region, where s≥τ.' Yet the algorithm in §III-B selects, for each target channel, the single argmax-CKA channel from each source client, with no threshold τ. The CKA quantiles reported in §V-E (q0.5=0.315, q0.9=0.507) show that many pairwise scores fall in a range where Eq. (17) is not assumed to hold, so the theoretical denoising guarantee does not cover the actual selection rule. Table III validates top-1% vs bottom-1% CKA regions, not the argmax-selected set. The paper should either enforce a threshold τ, validate Eq. (17) on the actual selected channels, or substantially weaken the theoretical claims. This is load-bearing because low-scoring argmax matches could inject domain-specific noise, consistent with CRIP's underperformance on Office-Hom","section":"§III-B, §IV-B, §V-C, §V-E"},{"comment":"The implementation uses ⁱˆz = z + α·̄z in Eq. (11), while the theory analyzes ̃Z = (1−α)Z + (α/m)ΣZ_i in Eq. (18). These are not equivalent transformations because Eq. (11) leaves the target coefficient at 1 and adds α times the source average, whereas Eq. (18) renormalizes target and source weights to sum to 1. Since the prediction head was trained on z, the global scaling factor (1+α) in Eq. (11) changes the operating point of the head. Equations (19)–(21) rely on the convex weights in Eq. (18), so the variance-reduction derivation does not apply to the implemented injection. The authors must either add a normalization step and report it, or re-derive the bounds for the actual formula used.","section":"§III-B Eq. (11) vs §IV-B Eq. (18)"},{"comment":"Eq. (21) claims the cross term in the residual decomposition vanishes 'since E[R_i|Y] = 0.' Zero conditional mean does not imply zero conditional covariance; E[⟨R_k, R_i⟩|Y] = Cov(Z_k, Z_i|Y), which need not vanish without an additional independence or decorrelation assumption. Consequently, the claimed denoising effect is not established by the derivation as written. The following sentence — 'if residuals across clients are not perfectly correlated given Y, then averaging strictly reduces the residual energy' — is also incomplete: positive but imperfect correlation can still leave the averaged residual energy above the individual level. The authors should state the additional assumption or provide the full quadratic form with the covariance terms.","section":"§IV-B Eq. (21)"},{"comment":"The main hyperparameter α is selected on the evaluation benchmarks. Section V-E states that 'in all experiments, we set α=0.8' after examining sensitivity on DomainNet (Fig. 5) and PACS-Cartoon (Table II), and the calibration batch size is also studied on DomainNet (Fig. 6). No held-out validation split is described for these choices. This constitutes selection on the test data and inflates the reported accuracies. Because α controls the core contribution (α=0 corresponds to the local model), a fair evaluation must set α and B on a validation split or via cross-validation, and should report results for a range of α rather than a single value tuned on the same benchmarks.","section":"§V-E, Table I"},{"comment":"The abstract and introduction claim that CRIP 'consistently outperforms' state-of-the-art baselines. The paper's own Table I contradicts this. On Office-Home, FOL achieves 76.23±0.51 versus CRIP's 70.77±1.14; on PACS Photo, FOL achieves 97.15±0.88 versus CRIP's 93.38±0.43; on DomainNet Real, FDRL achieves 96.99 versus CRIP's 96.41. The discussion acknowledges FOL's Office-Home advantage only as 'considerably less stable' on DomainNet, which does not address the direct comparison. The claim must be softened to 'competitive' or supported with aggregate statistical tests across benchmarks. As written, the central empirical claim is not established.","section":"Abstract, §I, Table I"}],"minor_comments":[{"comment":"The CRIP-K variant is introduced but not defined in Section III, and no table or figure reports its results. The sentence 'To justify the use of CRIP-K' refers to an unexplained variant; please clarify the method or remove the reference.","section":"§V-E"},{"comment":"Typo: 'Muti-round' should be 'Multi-round.' Also, Table II uses 'PACS-Cartoon' while Table I uses 'C' for the Cartoon domain; consistent naming would help.","section":"§V-B"},{"comment":"In the paragraph beginning 'Overall, these results show that CLIP is sensitive...', 'CLIP' appears to be a typo for 'CRIP.' This should be corrected.","section":"§V-E"},{"comment":"Typo: 'Tabel II' should be 'Table II' in the reference to the supplementary material.","section":"§V-G"},{"comment":"The notation in Eq. (16) uses X_k for target-domain inputs, while the decomposition in Eqs. (12)–(14) writes S_k and R_k for a random representation Z_k. The paper would benefit from explicitly stating that S_i and S_k in Eq. (17) are evaluated on the same target-domain input distribution.","section":"§IV-B"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a CS/LG venue and the core idea is interesting, but the reported results are affected by tuning on the test benchmarks, the theoretical section is not aligned with the implemented algorithm, and the main empirical claim is contradicted in one of the three datasets. These are fixable in revision, provided the authors re-run with a proper validation split, reconcile the injection formula with the theory, and moderate the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: CRIP is a genuinely new combination—channel-level CKA alignment plus residual injection for one-shot personalized FL—and it delivers large gains on DomainNet and PACS. But the paper overstates its case: on Office-Home it loses to FOL, and the theory in Section IV does not actually cover the algorithm as implemented.\n\nWhat's good: the method is simple, training-free, and doesn't need public data or generative models. Each client keeps its parameters; you only mix activations at inference. The empirical wins on DomainNet (84.46 avg vs 76.33 for the best baseline) and PACS (94.24 vs 83.74) are substantial. The ablation (CRIP-noA) and the CKA-vs-semantic-correspondence validation in V-C are useful. The privacy risk evaluation is a nice addition.\n\nThe soft spots are real. First, the abstract claims 'consistently outperforms' state-of-the-art baselines, but Table I shows FOL beats CRIP on Office-Home (76.23 vs 70.77) and on PACS Photo (97.15 vs 93.38). That's not consistent outperformance; it's strong gains on two benchmarks and a clear loss on another. Second, the theory-assumption gap. Eq (17) is a sufficient condition for denoising, but it's restricted to pairs with CKA score s≥τ. The algorithm selects the argmax channel per source without any threshold, so many injected channels likely fall below τ. The validation in V-C compares top-1% vs bottom-1% CKA pairs, not the actual argmax-selected set, so it doesn't establish that the condition holds for what's actually fused. Third, Eq (21) claims the cross term vanishes because E[R_i|Y]=0; zero conditional mean doesn't imply zero conditional covariance between residuals, so the denoising argument needs an additional independence assumption that isn't stated or tested. Fourth, α=0.8 is chosen after running sensitivity on the same benchmarks, with no separate validation split; that's tuning on test, albeit mild. And there's no code or data release, so the numbers aren't independently checkable.\n\nNone of this is fatal to the core idea. The method is precisely specified and the positive results on two datasets are worth taking seriously. But the claims need to be reined in, the threshold issue needs to be addressed (either apply τ or weaken the theory), and code should be released. This deserves a serious referee—send it out, but expect major revision.","headline":"CRIP's channel-level CKA matching with residual injection is a genuinely new and promising idea for one-shot personalized FL, with large wins on DomainNet and PACS, but the 'consistently outperforms' claim is contradicted by Office-Home, and the theory doesn't cover the actual algorithm as implemented.","tokens_in":17019,"tokens_out":2857,"would_cite":false,"duration_ms":18831,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that in one-shot federated learning, sharing knowledge in representation space—by aligning channels with CKA and injecting only the most compatible features as residuals—outperforms parameter-space aggregation under domain","keywords":["one-shot federated learning","domain heterogeneity","representation-space personalization","channel alignment","CKA","residual injection","feature fusion","few-shot learning"],"falsifier":"On a held-out domain pair, rank all channel pairs by CKA and compute their class-profile correlation Sim(c,c') as in Section V-C. If the top-1% CKA pairs have average Sim at or below zero, Eq. (17) fails for that pair, and CRIP should reduce accuracy below the target client's local model. Comparing CRIP with and without the alignment module on such a pair would directly test whether the claimed semantic-alignment mechanism carries the gain.","tokens_in":16154,"feed_emoji":"🧩","tokens_out":4004,"duration_ms":29870,"temperature":0.7,"pith_summary":"The paper argues that in one-shot federated learning, where clients communicate only once, the right place to share knowledge is not the weight space but the representation space. It proposes CRIP, which has each client forward a small local mini-batch through every other client's feature extractor, measures channel-wise similarity with CKA, and injects only the best-matched foreign channels as residual features. This gives each client a personalized model that keeps its own parameters untouched while borrowing semantically compatible features from others. On three domain-heterogeneous benchmarks, CRIP reports gains over local models and strong baselines, especially on heavily shifted domains. If correct, the result means training-free representation fusion can substitute for parameter aggregation in one-shot settings.","feed_headline":"Channel-level injection boosts one-shot FL by 10+ points","feed_subtitle":"Matching channels across clients lets each client borrow only compatible features—no parameter changes, no public data.","key_machinery":"The central object is the channel-level CKA alignment score s_{k,i}(c,c') computed on a calibration batch from the target domain. CKA (centered kernel alignment) is a similarity measure between two sets of activation vectors. For each target channel, CRIP selects the single best-matched channel per source client, averages their activations, and injects the result as a residual: z + α·z̄. The mechanism works because fusion is restricted to the high-CKA region, where the paper assumes (Eq. 17) that higher scores imply smaller expected semantic discrepancy; the averaging then suppresses uncorrelated residuals while the untouched target parameters preserve the model's own discriminative structur","core_discovery":"CRIP claims that under severe domain heterogeneity, one-shot federated learning can be solved in representation space: each client uploads only its feature extractor, the server broadcasts all extractors, and at inference each client runs a small local mini-batch through the other extractors, computes channel-wise CKA scores to find the best-matching channel per source, and adds the averaged aligned activations as a residual to its own feature map. The paper reports that this training-free procedure beats local models and existing one-shot methods on DomainNet, PACS, and Office-Home, with the largest gains on the most shifted domains. The accompanying theoretical decomposition splits each re","pith_inferences":["The method's success depends on clients having a small calibration batch from their own domain; in practice this implies a small labeled or unlabeled local set, which many FL deployments already possess but the paper does not discuss as a requirement.","CKA-based channel matching could serve as a general tool for cross-model transfer beyond FL, such as model merging or zero-shot knowledge distillation, wherever feature extractors see the same inputs.","Because inference cost grows linearly with the number of source clients, selecting a subset of high-CKA clients or parallelizing forward passes is a natural practical extension the paper does not develop.","If the alignment assumption holds only for channels with high CKA, a domain with unusual statistics (e.g., medical images) might need per-domain validation of Eq. (17) before injection."],"forward_implications":["One-shot federated learning can improve over local models even with no public dataset and no parameter averaging, as long as clients share feature extractors.","Under domain heterogeneity, representation-space alignment preserves transferable semantics better than parameter-space methods.","CKA-based channel selection is what makes fusion work: without alignment, averaging all channels degrades performance across all tested domains.","Fusing more source domains yields larger gains, consistent with a denoising effect from averaging uncorrelated residuals.","The method remains effective as the number of clients per domain grows, showing no obvious negative transfer at the tested scales."],"fun_headline_variants":["Channel matching turns one-shot FL up to 10 pts better","Selective channel fusion boosts personalized one-shot FL","CRIP: borrow only the channels that match your domain","One-shot FL sharpened by channel-level feature alignment","No public data, no params: just channel-wise CKA fusion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That the CKA similarity between a target channel and a source channel, measured on a small local batch, reliably tracks how semantically close those channels are on the test distribution—so that injecting only the top-matched channels adds transferable knowledge rather than noise; this is Eq. (17), an assumption the paper validates empirically on the three benchmarks rather than proves.","fun_headline_variants_meta":{"raw":{"variants":["Channel matching turns one-shot FL up to 10 pts better","Selective channel fusion boosts personalized one-shot FL","CRIP: borrow only the channels that match your domain","One-shot FL sharpened by channel-level feature alignment","No public data, no params: just channel-wise CKA fusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000659,"raw_usage":{"total_tokens":2854,"prompt_tokens":750,"completion_tokens":2104,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":2037}},"tokens_in":494,"tokens_out":2104,"duration_ms":13461,"temperature":1.0,"reasoning_tokens":2037,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:12:05.516008+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a held-out domain pair, rank all channel pairs by CKA and compute their class-profile correlation Sim(c,c') as in Section V-C. If the top-1% CKA pairs have average Sim at or below zero, Eq. (17) fails for that pair, and CRIP should reduce accuracy below the target client's local model. Comparing CRIP with and without the alignment module on such a pair would directly test whether the claimed semantic-alignment mechanism carries the gain.","supporting_citations":[],"review_version":1}