{"id":"afcade00-4d9d-4438-a19a-b9acf5c365a4","arxiv_id":"2608.10929","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A federated recommender keeps a fixed semantic item vocabulary and adds local, reliability-gated collaborative evidence, avoiding privacy-risky alignment of private item spaces.","lead":"FedCGR is a federated learning method for cross-domain product recommendation that represents items as shared semantic codes trained from public item metadata, then lets each domain privately add its own collaborative signals. It reports consistent gains over federated generative baselines on six Amazon scenarios and competitive results against stronger non-federated baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The local CF extractor may be trained on full user sequences including validation/test items; if so, the reliability-aware residual (Eq. 8) leaks the next item and inflates FedCGR's reported gains.","rationale":"The reader's weakest assumption concerns the semantic SID vocabulary; I do not dispute that as a conceptual risk, but the immediate load-bearing risk for the central claim is experimental: the empirical headline depends on the local CF channel not seeing test-time transitions. The paper's wording creates a real ambiguity, and the public code makes it checkable in one pass. I therefore recommend keeping the verdict CONDITIONAL, with the additional condition that the CF extractor split be verified and, if necessary, the experiments re-run. If the leakage proves real, the verdict would need to become REJECT or UNVERDICTED; if the check passes, the semantic vocabulary concern remains as a secondary limitation for generalization.","tokens_in":19378,"tokens_out":13315,"duration_ms":125276,"concrete_test":"Check the public repository's data pipeline for the local SASRec CF extractor: is the SASRec training set the same as the training prefix (first T-2 items per user) or the full sequences? Independently, re-run FedCGR with SASRec trained only on the training prefix and recompute Table 3 and Table 4; if the reported FedCGR margins over LocalSASRec and TIGER+FedProx largely persist, the leakage concern is resolved, and if they shrink, the headline is an artifact. The check should also report whether the same split was used for confidence scores rho_v,i and for the dense-loss positives.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section 3.2 says each client trains a local SASRec model 'using only its private interaction sequences R_i' and extracts CF item embeddings (Eq. 5), which are injected into the generator's input via the reliability-aware residual (Eq. 8). Section 4.1.3 defines the evaluation split: last item test, second-to-last validation, remaining training. The paper never states that the local CF extractor is restricted to the training prefix. If SASRec is trained on the full 5-core-filtered sequences, the CF embedding of a history item v_t is fitted to predict the subsequent item in the same user sequence, including the validation/test item. At inference, the last history item's embedding then encodes the target, so the model's input representation can 'see' the answer through the local CF channel. That would inflate every FedCGR result, especially the large margins over LocalSASRec and TIGER variants, and would invalidate the central empirical claim. This is not an accusation; the ambiguity in the text makes the experimental protocol unverifiable, and the public code should settle it.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes FedCGR, a federated cross-domain generative recommendation framework. The core idea is to represent items as discrete semantic-ID (SID) sequences produced by a frozen RQ-VAE tokenizer trained on public item metadata, so that cross-domain alignment is induced by a shared token vocabulary rather than by aligning private embeddings. To overcome the resulting semantic-only bottleneck, each client injects local collaborative-filtering evidence through a reliability-aware residual (item frequency confidence, client-local gate, and a local adapter), and a dense auxiliary head is trained only locally. To avoid negative transfer under domain heterogeneity, shared generator parameters are aggregated per target domain with weights computed from EMA-updated domain prototypes (Eqs. 17-20), while private experts, domain embeddings, CF adapters, and local statistics remain on clients. Experiments on six Amazon scenarios report that FedCGR consistently beats federated generative baselines (TIGER+FedAvg/FedProx) under full-ranking and achieves the best results among federated discriminative CDR baselines under 999-negative sampling; ablations show a cross-over pattern in which CF enrichment matters most in high-affinity scenarios and personalized aggregation matters most in heterogeneous ones. The paper includes a public code repository, an algorithmic description, and a parameter partition table.","tokens_in":19633,"tokens_out":14325,"duration_ms":115947,"significance":"If the empirical claims hold, the paper makes a useful conceptual contribution: it shows that federated CDR can be staged over a stable discrete item language, with adaptation handled explicitly through reliability-gated local CF residuals and prototype-personalized aggregation, rather than through private-space alignment. The framework is described precisely (Algorithm 1, Table 1), the tokenizer pipeline is reproducible from public metadata, and the code is released, so the main results are checkable. The ablation cross-over between CF enrichment and personalized aggregation is a falsifiable prediction that organizes the design space usefully. The main risk to significance is the unresolved training-split ambiguity for the local CF extractor discussed in the major comments; if the extractor saw the full sequences, the reported gains would be artifacts of target leakage. The shared-vocabulary assumption also deserves a caveat: the method's transfer value is bounded by how well metadata semantics track behavioral equivalence across domains, though the low-affinity GS results provide some positive evidence.","major_comments":[{"comment":"The experimental protocol for the local CF extractor is unverifiable as written. Section 3.2 trains SASRec \"using only its private interaction sequences R_i\" and Eq. (8) injects the resulting CF embeddings of every history item, including the last one, into the generator input; Section 4.1.3 defines the evaluation split (last item test, second-to-last validation, remaining training) but never states that the SASRec extractor is restricted to the training prefix. If SASRec is trained on the full 5-core-filtered sequences, the CF embedding of the last history item (the validation item) was fitted to predict the test item, so the residual channel in Eq. (8) leaks the target at inference, and the dense loss in Eq. (12) reinforces that leak during training; this would inflate every FedCGR result, especially the large margins over TIGER+FedAvg and TIGER+FedProx. The authors must clarify in the text, and verify with the public code, which data split the local CF extractor was trained on; if it was the full sequence, the experiments must be re-run with the extractor trained on the training prefix only and the tables re-reported.","section":"§3.2 / §4.1.3 / Eq. (8)"},{"comment":"There is an internal inconsistency between the stated data processing and the single-domain baseline results. Section 4.1.1 and the Table 2 footnote describe 5-core filtering as independent per domain, in which case SASRec and GRU4Rec results for the same domain must be identical across scenarios; however, Table 3 reports SASRec Grocery H@10 0.164 in GS and GBS but 0.111 in GKBS, and SASRec Sports H@10 0.105 in GS but 0.177 in GBS. The table footnote attributes this to \"scenario-specific data filtering,\" which is not consistent with independent per-domain filtering. In particular, the anomalous drop of SASRec on GS Sports (H@10 0.105 vs. GRU4Rec 0.198) means that the claim of FedCGR being best in GS rests on a possibly collapsed baseline. Please reconcile the two statements, report the exact filtering and splitting pipeline per scenario, and confirm that the GKBS single-domain numbers are not a filtering or logging artifact.","section":"Table 3 / §4.1.1"}],"minor_comments":[{"comment":"The ablation table reports the full FedCGR model at FK with N@10 = 0.048 as a scenario-level macro-average, but the corresponding macro-average from Table 3 is (0.077 + 0.029) / 2 = 0.053; please reconcile the two values or state explicitly that the ablations were computed on the validation split rather than the test split.","section":"Table 5 vs. Table 3"},{"comment":"The paper states that standard deviations are below 0.003 and omits them; because several reported margins are large but some comparisons are close, reporting the per-seed values or error bars in an appendix would make the significance of the results verifiable.","section":"§4.1.4"},{"comment":"The 999-negative sampled evaluation table contains only discriminative baselines; adding TIGER+FedAvg and TIGER+FedProx under the same sampled protocol would directly support the abstract's claim of consistent gains over federated generative baselines under both evaluation protocols.","section":"Table 4"},{"comment":"Only the FedProx bars in Figure 3 are annotated with their NDCG values; annotating the SID-only and Full bars would allow the reader to verify the claimed widening gap between FedCGR and its SID-only ablation as user activity increases.","section":"Figure 3"},{"comment":"Please state whether hyperparameters other than the aggregation temperature (notably λ_d, μ, and τ) were tuned on validation data or fixed by convention; the sensitivity analysis in Figure 4 shows a clear optimum for λ_d, so a tuning statement is important for reproducibility.","section":"§4.1.4"},{"comment":"A one-sentence clarification that the local CF extractor is trained on the same training prefix used for the federated generator would resolve the protocol ambiguity raised in the major comments and should be added regardless of the code check outcome.","section":"§3.2"}],"recommendation":"major_revision","confidential_remarks":"The verdict hinges on whether the local CF extractor respects the evaluation split (major comment 1). I recommend that the editor or a reviewer first run the public repository and inspect the SASRec training script before and during the revision; if the extractor trains on full sequences, the entire empirical section must be re-run, and the central claims would likely change materially. The Table 3 inconsistency (major comment 2) also suggests that an audit of the data filtering pipeline is warranted. If the protocol turns out to be clean, the paper is a solid and well-scoped contribution with an innovative design; if not, the empirical section cannot be trusted as reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: FedCGR is a reasonably new combination—federated generative CDR over Semantic IDs built from item metadata, with reliability-gated local CF residuals and prototype-personalized aggregation. Not a breakthrough, but a genuine new design. Experiments are extensive, ablations are informative, and the code is public. I would send it to a serious referee, with one important caveat.\\n\\nThe caveat is the local CF extractor. The paper never says the SASRec local model is trained only on the training prefix of each user sequence. Section 3.2 says it trains on \"private interaction sequences R_i\" and Section 4.1.3 splits last item test / second-to-last validation / remaining training. If SASRec sees full sequences, the CF embedding of the last history item is trained to predict the test item, so the residual injection in Eq. 8 leaks the answer. That would explain some of the larger margins. I am not accusing anyone—the protocol is under-specified, and the public code should settle it. This is potentially load-bearing and must be clarified before the numbers can be trusted.\\n\\nOther issues are minor by comparison. All main tables omit error bars; the paper says standard deviations are below 0.003 and omits them for readability, but some gaps are that small. The full-ranking table includes no federated discriminative baselines even though the abstract says FedCGR is competitive with federated CDR methods under full-ranking; read closely, that claim is only supported under 999-neg sampling. A few numbers look off, e.g., SASRec at 0.105 H@10 on GS Sports while GRU4Rec gets 0.198; that deserves a sentence. The privacy framing is honest about not claiming DP, which is fine.\\n\\nThe strongest parts: the design constraints (frozen tokenizer, semantic-only bottleneck, negative transfer) are well identified, and the two remedies map to them. The ablation cross-over pattern is a nice piece of evidence that the components matter differently by relatedness. Related-work coverage is good, and the code link is a plus.\\n\\nWho this is for: people working on federated recommendation or cross-domain transfer will want to read it once the leakage question is resolved. I would not cite it yet—I would check the code first. Send it to peer review, but demand that the authors state the CF training split explicitly and ideally add error bars.","headline":"A genuinely new federated CDR design built on shared semantic IDs, but the local CF extractor's training split is under-specified and could leak test items; the numbers should not be trusted until that is resolved.","tokens_in":20163,"tokens_out":2787,"would_cite":false,"duration_ms":25295,"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":"A shared discrete vocabulary for items lets federated recommender systems transfer knowledge across domains without sharing user behavior.","keywords":["federated cross-domain recommendation","generative recommendation","semantic IDs","residual quantization","federated learning","personalized aggregation","cold-start recommendation","reliability-aware fusion"],"falsifier":"Run FedCGR on a domain pair whose metadata vocabularies are deliberately disjoint—for instance, one domain described with a category taxonomy the other never uses—and compare against training each domain locally: if the federated method no longer beats the local models on cold-start users, the claim that the shared SID vocabulary induces alignment is refuted.","tokens_in":19164,"feed_emoji":"🧩","tokens_out":7273,"duration_ms":61527,"temperature":0.7,"pith_summary":"The paper argues that the hard part of federated cross-domain recommendation—aligning item spaces across clients without sharing interaction data—can be sidestepped by treating items as discrete semantic ID (SID) sequences produced once from public item-side metadata. Because every client uses the same frozen tokenizer, cross-domain alignment is inherited from the shared vocabulary instead of being learned from private behavior. Two problems then remain, and the paper's method FedCGR targets each: the frozen tokenizer blocks local collaborative-filtering knowledge, so FedCGR injects that knowledge through reliability-gated residual signals; and naive federated averaging causes negative transfer among heterogeneous domains, so FedCGR aggregates shared generator parameters by domain relatedness. On six cross-domain scenarios built from public review data, FedCGR consistently outperforms federated generative baselines and is competitive with strong non-federated sequential and federated cross-domain baselines. If correct, this shows federated CDR can be staged over a shared discrete item language with local adaptation, rather than by aligning private embedding spaces.","feed_headline":"Shared item vocabulary lifts federated cross-domain recommendation","feed_subtitle":"Items become stable semantic IDs, aligning domains without sharing interactions, while local signals pass a reliability gate.","key_machinery":"The load-bearing object is the fixed RQ-VAE SID vocabulary: a discrete tokenizer trained on public item metadata, such as titles, categories, and descriptions, that assigns every item a short sequence of codes and gives all clients the same item language without any interaction data. On top of it, a reliability-aware semantic interface fuses SID token embeddings with a client-local collaborative-filtering residual: item frequency gives a confidence weight, a learned gate calibrates each domain's overall collaborative contribution, and a local adapter maps the local embedding coordinates into the generator's hidden space. The second mechanism is a prototype-personalized federated generator: shared layers use a mixture-of-experts block with shared experts plus a client-private expert, clients upload only shared parameters and a domain prototype, and the server forms each domain's aggregate by softmax weights over cosine similarity between prototypes, falling back to data-volume federated averaging as the temperature diverges. These two mechanisms answer the paper's two constraints: information loss from frozen tokenization and negative transfer under heterogeneity.","core_discovery":"The central claim is that federated cross-domain recommendation can be reformulated as autoregressive generation over a stable semantic item language, and that this reformulation converts privacy-sensitive alignment into a property of a fixed vocabulary. Items are mapped by a residual-quantization tokenizer over metadata embeddings into short discrete codes; the tokenizer is trained once and never updated, preserving token consistency across clients. To compensate for the resulting semantic-only bottleneck, each client learns a local collaborative-filtering embedding from its own interaction sequences and inserts it as a residual whose contribution is scaled by item reliability and a client-local gate. To survive domain heterogeneity, the generator's shared parameters are aggregated per domain using weights derived from cosine similarity between client prototypes, while domain-specific parameters stay local. The paper reports that this design outperforms federated SID-generation baselines on all reported full-ranking cells, beats federated discriminative cross-domain baselines under sampled evaluation, and yields a cross-over in ablations: local collaborative evidence matters most when domains are closely related, while personalized aggregation is the binding constraint when they are diverse.","pith_inferences":["Beyond the paper, the shared-vocabulary argument implies a zero-shot entry test: a domain that never participated in federation but whose items can be encoded by the same tokenizer should inherit cross-domain knowledge immediately, a claim the experiments do not directly verify.","The cross-over ablation also suggests an adaptive control rule that the paper leaves implicit: a federation could inspect prototype divergences each round and adjust the strength of collaborative residuals and the sharpness of aggregation accordingly.","A natural stress test not reported here is a domain pair with disjoint metadata vocabularies; the framework predicts graceful degradation toward local baselines, whereas an alignment-based method would be expected to fail loudly, so this comparison would separate the two families of approaches."],"forward_implications":["The fixed item language removes reliance on overlapping users or shared interaction signals for alignment, since the shared tokenizer defines the alignment before federated training begins.","Because only shared generator parameters and domain prototypes leave the client, raw interaction sequences and all domain-specific modules remain private, and secure aggregation or differential privacy can be added without altering the design.","Ablations show a cross-over: local collaborative evidence is what matters most when domains are closely related, while personalized aggregation becomes the binding constraint under high heterogeneity.","On the reported scenarios, the faster convergence of personalized aggregation cuts total upload communication by roughly a third to a half relative to data-volume averaging, despite an 18 percent larger per-round payload."],"supporting_citations":[{"why":"Defines the SID-based generative retrieval paradigm, representing items as discrete code sequences, that FedCGR federates.","marker":"[37]"},{"why":"Supplies the local self-attentive sequential model used to extract per-client collaborative-filtering item embeddings.","marker":"[21]"},{"why":"Provides the frozen text encoder that maps public item-side metadata into the embedding space the tokenizer quantizes.","marker":"[3]"},{"why":"Introduces residual quantization, the codebook mechanism that converts metadata embeddings into discrete semantic ID codes.","marker":"[22]"},{"why":"Defines federated averaging, the aggregation rule that FedCGR replaces with prototype-personalized weighting.","marker":"[35]"},{"why":"Supplies the review collection from which the six cross-domain evaluation scenarios are constructed.","marker":"[36]"},{"why":"Centralized generative cross-domain recommendation with semantic tokenization, the non-federated precursor whose assumptions FedCGR drops.","marker":"[17]"},{"why":"Federated cross-domain sequential recommendation baseline compared under the sampled 999-negative evaluation protocol.","marker":"[53]"}],"fun_headline_variants":["Semantic IDs align federated domains with no shared interactions","Generative cross-domain beats baselines: fix vocab, personalize rest","Shared tokenizer + local CF makes federated CDR competitive","Federated generative recommendation: stable SIDs, selective aggregation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole cross-domain gain rests on a tokenizer trained from public item metadata producing a shared item vocabulary whose semantic grouping actually tracks how users behave in each domain; if metadata categories do not align with behavioral similarity, the shared language carries no real transfer signal and the local collaborative signals cannot make up for it.","fun_headline_variants_meta":{"raw":{"variants":["Semantic IDs align federated domains with no shared interactions","Generative cross-domain beats baselines: fix vocab, personalize rest","Shared tokenizer + local CF makes federated CDR competitive","Federated generative recommendation: stable SIDs, selective aggregation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1445,"prompt_tokens":1011,"completion_tokens":434,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":363}},"tokens_in":627,"tokens_out":434,"duration_ms":4728,"temperature":1.0,"reasoning_tokens":363,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:03:49.512730+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FedCGR on a domain pair whose metadata vocabularies are deliberately disjoint—for instance, one domain described with a category taxonomy the other never uses—and compare against training each domain locally: if the federated method no longer beats the local models on cold-start users, the claim that the shared SID vocabulary induces alignment is refuted.","supporting_citations":[{"cited_title":"Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Q","cited_arxiv_id":null,"evidence_quote":"Defines the SID-based generative retrieval paradigm, representing items as discrete code sequences, that FedCGR federates."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the local self-attentive sequential model used to extract per-client collaborative-filtering item embeddings."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the frozen text encoder that maps public item-side metadata into the embedding space the tokenizer quantizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces residual quantization, the codebook mechanism that converts metadata embeddings into discrete semantic ID codes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines federated averaging, the aggregation rule that FedCGR replaces with prototype-personalized weighting."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the review collection from which the six cross-domain evaluation scenarios are constructed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Centralized generative cross-domain recommendation with semantic tokenization, the non-federated precursor whose assumptions FedCGR drops."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Federated cross-domain sequential recommendation baseline compared under the sampled 999-negative evaluation protocol."}],"review_version":1}