{"id":"f751b54e-4dfe-4f3d-9dbe-f366db0e4819","arxiv_id":"2608.06997","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"HD-REC combines hierarchical quantization, domain-adaptive sparse MoE, and routing consistency learning to improve cross-domain generative recommendation.","lead":"A recommendation system encodes items into hierarchical codes with shared coarse and routed fine codebooks, then uses a transformer with a shared expert plus one selected specialist per token. The authors report consistent gains over existing sequential and generative cross-domain models on three public data pairs, especially in sparse domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's reported improvements are internally inconsistent (Sports N@10: 1.7% vs 11.4% from the raw values), so the 'best on all metrics' claim is not currently supported by the displayed evidence.","rationale":"The paper is a serious engineering contribution: the HDQ/DAS-MoE/CRCL design is coherent, the ablations in Tables 3 and 4 point in the expected direction, and Section 7 explicitly acknowledges the content-embedding dependency. I therefore do not object to the method's conceptual validity. The central claim, however, is 'best performance across all domains and metrics' (Section 4.2), and that claim is only as strong as Table 2. The table's own Impr row is not reproducible: Sports H@10 and N@10 percentages do not follow from any single baseline in the table, and the Sports N@10 figure matches TIGER rather than the strongest CDSR baseline. The paper asserts t-tests at p<=0.05 without reporting seeds, standard deviations, or any paired-test details, and no code is released. These are not stylistic quibbles; they are the evidence for the headline. The reader's weakest-assumption framing (content-embedding alignment) is a real scope limitation, but it is acknowledged in Section 7 and affects all content-based baselines equally; it is not the reason the headline claim is currently unverifiable. The conditional verdict is already appropriate, so I recommend no change; if the arithmetic is corrected and code or per-seed results are supplied, the concern dissolves.","tokens_in":12735,"tokens_out":8774,"duration_ms":83298,"concrete_test":"Recompute every entry in the Impr row of Table 2 from the raw scores and list which baseline was used for each metric; specifically resolve Sports H@10 (0.0489 vs 0.0403 = 21.3%, not 17.6%) and Sports N@10 (0.0292 vs 0.0262 = 11.4%, not 1.7%). If the corrected values still show HD-REC best on all 12 metrics, obtain per-seed results and rerun the paired t-test; if the table cannot be reconstructed, the claimed 'best on all metrics' is unverified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: HD-REC 'achieves the best performance across all domains and metrics.' The weakest load-bearing point is that the displayed evidence for this claim is internally inconsistent. In Table 2, the Sports row reports Impr = 17.6% for H@10 and 1.7% for N@10, but from the raw values, HD-REC vs GenCDR gives 0.0489/0.0403 = 21.3% and 0.0292/0.0262 = 11.4%. The N@10 figure only matches TIGER's 0.0287 (1.7%), meaning the 'strongest baseline' is not consistently GenCDR. The paper asserts t-tests at p<=0.05 without reporting per-seed variance, error bars, or any paired-test details, and no code is released. The method itself is coherent and Section 7 candidly scopes the content-embedding dependency, but the headline comparison rests on a table whose arithmetic and baseline choices are not reproducible.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes HD-REC, a generative framework for cross-domain sequential recommendation. The method has three components: a Hierarchical Domain-Aware Quantizer (HDQ) that builds semantic IDs from shared coarse-level codebooks and an adaptively routed fine-level codebook; a Domain-Adaptive Sparse Mixture-of-Experts (DAS MoE) that combines an always-active shared expert with a dynamically selected specialized expert; and a Cross-Granularity Routing Consistency Learning (CRCL) objective that regularizes token-level routing decisions toward item-level consensus. Experiments on three cross-domain benchmark pairs (Clothing-Sports, Electronics-Phones, Books-Movies) compare HD-REC against single-domain, generative, and cross-domain baselines, with ablations, hyperparameter sensitivity, efficiency, and routing-consistency analyses.","tokens_in":13050,"tokens_out":4300,"duration_ms":38284,"significance":"The proposed architecture is coherent and well motivated: hierarchical quantization with shared coarse and adaptively routed fine codebooks is a plausible way to balance common semantics and domain-specific variation, and the routing-consistency regularizer directly targets the multi-token semantic-ID problem in generative recommendation. The paper also includes an honest limitations section that acknowledges the dependency on content-derived item embeddings and the current restriction to two-domain settings. If the empirical results are correct, the contribution is meaningful for the generative cross-domain recommendation line of work. However, the central claim of consistent state-of-the-art performance currently rests on tables with arithmetic inconsistencies and conflicting ablation numbers, so the reported results are not yet reliable as published.","major_comments":[{"comment":"The reported improvements in the Sports row are internally inconsistent. For H@10, HD-REC (0.0489) versus GenCDR (0.0403) yields a relative improvement of 21.3%, not the reported 17.6%. For N@10, HD-REC (0.0292) versus GenCDR (0.0262) yields 11.4%, not the reported 1.7%; the 1.7% value matches the comparison against TIGER (0.0287). Thus the 'strongest baseline' is not applied consistently across metrics, and the headline claim that HD-REC 'achieves the best performance across all domains and metrics' is not supported by the displayed arithmetic.","section":"Section 4.2, Table 2"},{"comment":"The 'Full HDQ' row in Table 4 (Clothing H@10 0.0256, N@10 0.0191; Sports H@10 0.0418, N@10 0.0259) does not match the full HD-REC row in Table 3 (Clothing H@10 0.0278, N@10 0.0209; Sports H@10 0.0489, N@10 0.0292). The Table 4 numbers instead match the 'w/o DAS MoE' row in Table 3. This suggests that the tokenizer design ablations in Table 4 were run on a variant lacking DAS MoE and CRCL, so 'Full HDQ' is not the full proposed model. As a result, the tokenizer ablation study is not a controlled comparison of HDQ within the complete HD-REC model.","section":"Section 4.3 and Section 4.4, Tables 3 and 4"},{"comment":"The statement that t-tests showed significant performance improvements (p≤0.05) is not substantiated. No standard deviations, number of seeds, per-seed results, or details of the paired test procedure are reported. This is load-bearing because several improvements are very small (e.g., Movies H@10 +0.4%, Electronics N@10 +0.7%), and without variance information these differences may not be statistically meaningful.","section":"Section 4.2, Table 2 caption"}],"minor_comments":[{"comment":"The method name is written 'HD-Rec' in the abstract but 'HD-REC' throughout the rest of the paper; please unify the notation.","section":"Abstract"},{"comment":"The phrase 'Same to the HDQ router' should be 'As with the HDQ router' or 'Similar to the HDQ router'.","section":"Section 3.2"},{"comment":"There is a typo in 'we propose a softCross-Granularity Routing Consistency regularization'; a space is missing after 'soft'.","section":"Section 3.3"},{"comment":"The symbol Lbalance is used both for the quantizer load-balancing loss in Eq. (15) and for the MoE load-balancing loss in Eq. (24), but these are different losses. Please use distinct notation or clarify which loss is intended in each equation.","section":"Equations (15) and (24)"},{"comment":"No code or detailed hyperparameter settings are provided for the baseline methods; to support reproducibility, please release code or provide precise configuration details.","section":"Section 4.1.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is sound and the limitations are candidly acknowledged, but the empirical evidence as presented contains load-bearing arithmetic and ablation inconsistencies. These are fixable within the scope of the manuscript, so I do not recommend rejection. However, the authors should be asked to correct Table 2's improvement calculations, align Table 4's 'Full HDQ' with the full model or clearly describe the ablated setting, and provide proper statistical significance details before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, the architecture is a real and sensible step forward: hierarchical quantization that keeps coarse semantics shared across domains while routing fine-level codebooks, combined with a shared-expert MoE and a cross-granularity routing consistency loss. That combination is new in the cross-domain generative recommendation literature, and the paper makes a solid case for why it should help. Second, the empirical evidence as presented is not trustworthy enough to back the headline claim. The percentages in Table 2 don't always line up with the raw values, and the ablation tables give conflicting numbers for the same configuration.\n\nWhat the paper does well: the motivation is clear, the design choices are tested with targeted ablations, the inference overhead is measured, and the routing-consistency metric gives a direct check on whether the regularizer does what it claims. The limitations section is unusually candid—it names the content-embedding dependency and the two-domain scope without hedging.\n\nThe soft spots are real but fixable. In Table 2, Sports N@10 is reported as a 1.7% improvement but the raw values give 11.4%; H@10 is 17.6% but raw gives 21.3%. Either the strongest baseline is not consistently GenCDR, or the arithmetic is wrong. The t-test claim (p≤0.05) is not backed by per-seed numbers, error bars, or any test details. And Table 3 vs Table 4: HD-REC scores 0.0278 H@10 on Clothing in Table 3, but 'Full HDQ' scores 0.0256 in Table 4. That is a conflict that needs an explanation—without it, the tokenizer ablation story is muddled. The gains on the Douban pairs are also quite small (0.4–0.8%), which is fine, but it tempers the 'consistent improvements' rhetoric.\n\nThis paper deserves a serious referee. It is not ready to accept as-is; it needs a corrected Table 2, consistency between tables, error bars, and ideally a code release. The architecture is worth engaging with, and the routing-consistency idea is worth borrowing even if the numbers get revised. I would bring it to a reading group focused on generative recommendation.","headline":"A sensible and novel architecture for generative cross-domain recommendation, but the headline performance claim is undermined by inconsistent arithmetic in Table 2 and conflicting ablation numbers.","tokens_in":13556,"tokens_out":3603,"would_cite":true,"duration_ms":33653,"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":"HD-REC shows that a hierarchical semantic-ID tokenizer with shared coarse codebooks and routed fine codebooks, combined with sparse shared-plus-specialized experts, improves generative cross-domain recommendation across all tested domains.","keywords":["generative recommendation","cross-domain recommendation","semantic IDs","hierarchical quantization","residual quantization","mixture-of-experts","sparse routing","sequential recommendation"],"falsifier":"Retrain HD-REC after replacing each item's content feature with random noise of the same dimension, or with shuffled metadata, while keeping interaction sequences unchanged; if Hit@10 stays close to an ID-only baseline, the semantic-ID premise is not what drives the result.","tokens_in":12577,"feed_emoji":"📈","tokens_out":8896,"duration_ms":80920,"temperature":0.7,"pith_summary":"The paper tries to establish that a single generative recommender can handle cross-domain recommendation better than existing sequential, generative, and cross-domain baselines by building item identifiers hierarchically: coarse semantic codes come from codebooks shared across all domains, while the final fine code is chosen from one of several codebooks routed per item. On the backbone side, the same philosophy appears as a sparse mixture-of-experts layer with an always-active shared expert and one token-selected specialist. The authors argue that this combination matches the geometry of real cross-domain item embeddings, which show both shared coarse structure and domain-specific fine structure. If they are right, cross-domain transfer in generative recommendation does not need domain-specific modules or manually aligned features, and a small consistency regularizer that keeps the semantic tokens of one item routing to the same expert is enough to stabilize the learned representations. The reported result is consistent wins on all three benchmark pairs, with the largest gains on the sparser Sports and Electronics domains; the authors also note that noisy or weakly aligned item metadata would hurt the approach.","feed_headline":"One routed codebook layer lifts cross-domain recommendations","feed_subtitle":"HD-REC shares coarse semantics and routes fine details to specialized experts, beating all baselines on three benchmark pairs.","key_machinery":"The load-bearing machinery is a hierarchical residual quantizer (HDQ) together with a sparse expert layer (DAS MoE) and a consistency regularizer (CRCL). HDQ runs $L-1$ quantization steps against codebooks shared by all domains, then routes the final residual through a Gumbel-Softmax selection to one of $K$ specialized codebooks, so coarse semantics stay common and fine residual semantics become input-dependent. DAS MoE applies the same shared/routed split to the Transformer feed-forward computation, activating one shared expert for every token and one specialized expert chosen from the token's hidden state. CRCL averages the token embeddings of each item, computes an item-level routing distribution, and penalizes the KL divergence between each token's distribution and that item-level consensus; this is what forces the tokens that make up one semantic ID to route coherently. Together the three components are claimed to provide hierarchical representation capacity and sparse conditional computation without domain-specific parameters.","core_discovery":"The central claim is that HD-REC, a unified generative framework, outperforms competitive methods on pairwise cross-domain sequential recommendation by allocating adaptive capacity in two places: the tokenizer routes the residual fine-level code to one of $K$ codebooks after $L-1$ layers of globally shared quantization, and the backbone replaces dense feed-forward layers with a shared expert plus a top-1 routed specialist. The paper reports best Hit@10 and NDCG@10 across all six domains of the three benchmark pairs; relative to the strongest baseline GenCDR, Hit@10 improves by 17.6% on Sports (0.0403 to 0.0489), 16.3% on Electronics (0.0342 to 0.0398), and 9.9% on Phones, with smaller consistent gains elsewhere. Ablations on Clothing-Sports attribute the gains to all three components: removing the hierarchical quantizer, the sparse MoE, or the routing consistency loss each lowers performance, and replacing learned routing with random routing or with a single expanded codebook also hurts. The authors caution that the performance comparison supports the hierarchical design but does not by itself establish representation disentanglement. The paper further claims the consistency regularizer cuts routing variance across an item's tokens by roughly $6\\times$ while adding only about 2 ms per test instance, so the extra capacity is nearly free at inference time.","pith_inferences":["Because the paper only evaluates pairwise two-domain settings, the clearest untested extension is adding domains one at a time: since coarse codebooks are shared and only fine codebooks and expert slots are routed, the design suggests new domains could be absorbed by adding routed capacity without retraining shared layers; that scalability intuition is mine, not the paper's.","The routing consistency idea could be applied symmetrically to the quantizer's codebook router, not just the expert router, and might further stabilize fine-level code assignment for items with very similar embeddings; the paper leaves that direction open.","The largest gains on sparser domains hint that the benefit of adaptive routing grows with domain divergence, so a controlled experiment varying the semantic overlap between source and target domains would test whether HD-REC's advantage tracks sparsity or domain distance."],"forward_implications":["Unified training across domains can work without explicit entity overlap or hand-built feature alignment: the shared coarse codebooks transfer common semantics, and the routed fine codebooks absorb domain-dependent residuals.","Sparse conditional capacity at the backbone level is better than a dense feed-forward network of similar size under the same cross-domain generative objective, because the ablation that replaces DAS MoE with a dense FFN loses accuracy.","Consistency among the tokens of an item is a useful inductive bias: the CRCL variant with $\\lambda_1=0.01$ outperforms the no-consistency variant by roughly 9--15% on Clothing and Sports and reduces routing variance by about $6\\times$.","The added computation is small, about 2% inference overhead under beam search, so the accuracy gains do not come at a deployment-prohibitive cost.","The gains concentrate in the sparser domains of each pair, suggesting adaptive route capacity matters most where in-domain signal is weak."],"supporting_citations":[{"why":"GenCDR is the strongest cross-domain generative baseline that HD-REC is directly compared against; it supplies the domain-adaptive tokenizer-and-recommender setup this work extends.","marker":"Hu et al. [2026]"},{"why":"Residual quantization with shared codebooks (RQ-VAE) is the tokenizer design that HDQ modifies by routing the final layer.","marker":"Lee et al. [2022]"},{"why":"Gumbel-Softmax provides the differentiable discrete routing used for both the codebook router and the expert router.","marker":"Jang et al. [2016]"},{"why":"TIGER establishes Semantic ID generative retrieval and is the efficiency baseline in the inference-overhead experiment.","marker":"Rajput et al. [2023]"},{"why":"T5 is the Transformer backbone that HD-REC trains for next-token prediction over semantic IDs.","marker":"Raffel et al. [2020]"},{"why":"VQ-Rec is the vector-quantized generative recommendation baseline whose tokenization approach HD-REC builds on and compares with.","marker":"Hou et al. [2023]"},{"why":"HSTU is a generative sequential transducer baseline that frames recommendation as autoregressive transduction; HD-REC compares against it.","marker":"Zhai et al. [2024]"}],"fun_headline_variants":["Routing fine codes to specialists lifts cross-domain Hit@10","Shared coarse codes, routed fine codes: cross-domain gain","Routing variance down 6x, cross-domain hit rate up 17.6%","HD-REC: hierarchical codes and sparse routing beat baselines","One shared codebook, one routed codebook: cross-domain win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline depends on content-derived item embeddings being available and aligning with what users actually prefer; if item metadata is noisy, incomplete, or weakly related to preference, the semantic IDs built from it lose meaning and performance suffers, as the paper itself concedes.","fun_headline_variants_meta":{"raw":{"variants":["Routing fine codes to specialists lifts cross-domain Hit@10","Shared coarse codes, routed fine codes: cross-domain gain","Routing variance down 6x, cross-domain hit rate up 17.6%","HD-REC: hierarchical codes and sparse routing beat baselines","One shared codebook, one routed codebook: cross-domain win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001268,"raw_usage":{"total_tokens":5237,"prompt_tokens":1037,"completion_tokens":4200,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":4110}},"tokens_in":653,"tokens_out":4200,"duration_ms":26649,"temperature":1.0,"reasoning_tokens":4110,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:53:33.344311+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain HD-REC after replacing each item's content feature with random noise of the same dimension, or with shuffled metadata, while keeping interaction sequences unchanged; if Hit@10 stays close to an ID-only baseline, the semantic-ID premise is not what drives the result.","supporting_citations":[{"cited_title":"Learning vector-quantized item representation for transferable sequential recommenders","cited_arxiv_id":null,"evidence_quote":"VQ-Rec is the vector-quantized generative recommendation baseline whose tokenization approach HD-REC builds on and compares with."}],"review_version":1}