{"id":"30c6e004-7589-4939-8fa0-f27b4357546d","arxiv_id":"2507.12871","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GMC uses shared discrete semantic item IDs and a unified generative recommender with domain-specific LoRA to improve multi-target cross-domain recommendation.","lead":"This paper proposes GMC, a recommendation model that turns item descriptions into shared token codes and generates the next item with one unified model for multiple categories. It reports better accuracy than existing cross-domain and generative baselines on five Amazon datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 5.1.4's beam-size-20 generation cannot implement Section 5.1.3's 'rank over the entire item set'; no scoring rule for items outside the beam is specified, so GMC's headline NDCG gains are not reproducible.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the evaluation protocol for generative ranking is underspecified. The central claim is empirical; for it to hold, the reported Recall/NDCG numbers must reflect a well-defined ranking over all candidate items. Since §5.1.3 promises full-corpus ranking while §5.1.4 only provides beam search with beam width 20, this condition is not met in the text. No code release or error bars are provided, so the gap cannot be filled externally. The issue is concrete: any item not in the beam has no score, and the paper does not say whether misses are assigned a fixed rank, excluded, or scored by some hidden function. Because the reported gains over the strongest baselines are often small in absolute value (e.g., Pantry N@5: 0.0262 vs 0.0209), even a minor change in miss handling could alter the conclusion. This supports the CONDITIONAL verdict rather than ACCEPT; it does not justify REJECT, since the method architecture and ablations are coherent and the concern is addressable by specifying and/or releasing the scoring protocol. Hence I leave the reader's verdict unchanged.","tokens_in":20333,"tokens_out":7029,"duration_ms":83366,"concrete_test":"Run a controlled reproduction of Table 3 with two explicitly documented protocols: (A) truncated beam protocol: rank = beam position for generated items, and rank = beam_size + 1 (or ∞) for all other items; (B) full-corpus scoring protocol: for every item in the target domain, compute the sum (or product) of the model's token probabilities along the item's L=4 semantic identifier (or a prefix-tree completion score) and rank all items by this score. If GMC's NDCG@5/10 advantages over VQ-Rec, UniSRec, TIGER, and IDGenRec do not survive both protocols, then the reported 'entire item set' ranking is not established and the headline results may be an artifact of the unspecified fallback. Releasing the scoring code would settle the ambiguity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"At inference (§3.5, Eq. (10)) GMC generates tokens by argmax over the vocabulary and applies constrained generation on a prefix tree with beam size 20 (§5.1.4). Section 5.1.3 states that the target item is ranked over the entire item set, but the paper never defines a score for an item whose identifier is not emitted in the beam. It also does not state a fallback rule (e.g., miss rank = beam_size + 1 or miss rank = ∞ for truncated metrics), nor a full-corpus scoring function such as summing token log-probabilities over each item's identifier path. Table 3's headline numbers, in particular the NDCG advantages (Pantry N@5 +25.41%; AVERAGE N@10 +6.67%), therefore cannot be reproduced from the described algorithm. The gap between GMC and the strongest baseline is small in absolute terms (e.g., Pantry N@5 0.0262 vs 0.0209), so any unspecified handling of non-generated items can change which method wins. This is a reproducibility gap in the evaluation, not an attack on the method's internal soundness; however, the ablations in Table 4 depend on the same undefined metric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GMC, a generative framework for multi-target cross-domain recommendation (MTCDR) that does not rely on overlapping users or items. Items are first encoded by an LLM into semantic embeddings, then tokenized by a shared RQ-VAE into hierarchical discrete codes; a domain-aware contrastive loss encourages intra-domain code alignment. A unified T5-based sequence-to-sequence model is trained to generate next-item code sequences from all domains, and domain-specific LoRA adapters are fine-tuned. Experiments on five Amazon subsets compare GMC against ID-only, ID-text, pre-training-based, and generative baselines using Recall@5/10 and NDCG@5/10. Ablations isolate the shared codebook, unified recommender, contrastive loss, and LoRA fine-tuning. The central empirical claim is that GMC outperforms the strongest baselines on most metrics, particularly NDCG, with relative gains up to about 25% on smaller domains.","tokens_in":20627,"tokens_out":7220,"duration_ms":78213,"significance":"If the empirical claims hold, GMC is a useful instantiation of generative MTCDR: it shows that shared semantic codebooks plus a unified generative recommender can transfer knowledge across domains without overlapping entities and without auxiliary pre-training data. The method is clean and well motivated, and the paper has several strengths: evaluation on five public datasets, a component-level ablation (Table 4), a parameter-efficiency statement (LoRA 0.17M, 3.7% of the 4.3M model), and analyses of code-sharing behavior across domains (Figures 3 and 4). The main weaknesses are that the evaluation's full-corpus ranking protocol for beam-search generative models is underspecified, no variance or significance information is reported, and the non-overlap property of the evaluation data is not demonstrated. These are fixable but currently prevent the headline improvements from being fully reproducible.","major_comments":[{"comment":"The ranking protocol for generative models is underspecified. Section 5.1.3 states that the target item is ranked over the entire item set, but Section 5.1.4 describes only beam search with beam size 20 over a prefix tree and never defines a score for an item whose identifier is not emitted in the beam. No fallback rule is given (e.g., miss rank = beam_size + 1, or exclusion from truncated metrics), and no full-corpus scoring function is provided (e.g., summing token log-probabilities over each item's identifier path under constrained decoding). Without this, Table 3's NDCG improvements—e.g., Pantry N@5 +25.41% and AVERAGE N@10 +6.67%—cannot be reproduced, and the small absolute gaps (e.g., AVERAGE N@5 0.0780 vs 0.0742) mean the unspecified handling of non-generated items could change which method wins. Please specify the exact inference-time scoring rule, including how items outside the beam are ranked, and provide evaluation code or pseudocode. Note also that Eq. (10) describes greedy argmax decoding, which does not match the beam-search description in Section 5.1.4.","section":"§5.1.3, §5.1.4, §3.5"},{"comment":"No error bars, confidence intervals, or significance tests are reported, and the number of random seeds is not stated. Several headline advantages are small in absolute terms (e.g., Office N@5 0.0912 vs IDGenRec 0.0902, +1.11%; Instruments R@10 0.1224 vs VQ-Rec 0.1209, +1.24%), so it is unclear whether these differences are stable. Please report mean and standard deviation over at least three seeds, and state the significance test used, for both the main comparison and the ablations.","section":"Tables 3 and 4"},{"comment":"The non-overlap property of the evaluation data is not established. The paper emphasizes that GMC targets scenarios where overlapping users or items are unavailable, and Table 1 marks GMC as 'Non-OL'. However, Section 5.1.1 does not state whether users or items appearing in multiple Amazon subsets were removed, and the statistics in Table 2 report per-domain user counts without giving the number of distinct users across all five domains. Since Amazon users commonly purchase in multiple categories, the five subsets may contain overlapping users. Please report the overlap statistics, or filter the data to make the domains disjoint, so that the experiments genuinely instantiate the non-overlapped setting claimed in the paper.","section":"§5.1.1, Table 2"}],"minor_comments":[{"comment":"The RQ-VAE is described as 'trained for 10k epochs' with batch size 1024; given the dataset sizes, this is likely intended to mean training steps or iterations. Please clarify.","section":"§5.1.4"},{"comment":"Eq. (10) describes inference as greedy argmax decoding, while Section 5.1.4 says beam size 20 is used. Reconcile the two descriptions and state whether beam search replaces the argmax in Eq. (10).","section":"§3.5, §5.1.4"},{"comment":"There are several typos: 'generative recommendan paradigm' in Section 2, 'to to 1024' in Section 5.1.4, 'benericial' in Section 5.2, and 'A VERAGE' in Table 3. Please proofread.","section":"§2, §5.1.4, §5.2"},{"comment":"The third-best highlighting appears as '::::' in the manuscript text; please ensure the wave underline renders correctly so that the optimal, sub-optimal, and third-best values are all clearly visible.","section":"Table 3"},{"comment":"In Figure 5, the 'relative improvement' axes are ambiguous: please specify whether the relative improvements are computed against the single-domain IDGenRec baseline or against VQ-Rec for each data point, and define 'Gap' explicitly in the caption.","section":"§5.4.3, Figure 5"},{"comment":"The claim of being the 'first time' the generative paradigm is applied to MTCDR is strong; consider softening to 'to our knowledge' and discussing any concurrent or closely related work.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent and readable conference submission. The central contribution is plausible, and the reported ablations support the individual design choices. The main risk is the undefined full-corpus ranking protocol for beam-search generative models; if the authors provide a precise scoring rule and evaluation code, the empirical claims become testable. I would not require new experiments beyond clarifying the protocol, adding variance estimates, and reporting dataset overlap statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: GMC is a plausible, well-ablated engineering contribution—semantic item IDs from a shared RQ-VAE, a unified generative recommender, and per-domain LoRA adapters—and the empirical story mostly checks out. The one thing I'd want fixed before trusting the numbers is the evaluation protocol: Section 5.1.3 says the target item is ranked over the entire item set, but 5.1.4 gives only beam-size-20 constrained generation and never defines a score for items not emitted in the beam. In the paper as written, Table 3's NDCG gains, including the large relative wins on Pantry, aren't reproducible from the described algorithm. That's a reproducibility gap in the write-up, not a flaw in the method's logic; the solution is a sentence or two: state the fallback rank (e.g., beam size + 1) or give a full-corpus scoring rule.\n\nWhat I like: the task choice is genuinely useful—MTCDR without overlapped users/items and without expensive auxiliary pretraining is an important gap, and GMC is the first generative recipe aimed at it. The design is coherent: shared codebook, unified seq2seq training, LoRA fine-tuning, and a domain-aware contrastive loss that actually shows up in the analysis (first-level codes are domain-specific, deeper levels shared). The ablations are thorough and directionally consistent; the similarity analysis in Figure 4 supports the identifier quality story. They also did the right thing aligning pretraining-based baselines to the same target data.\n\nSoft spots beyond the ranking protocol: no error bars or significance tests, and several headline gains are small in absolute terms (Pantry N@5 goes from 0.0209 to 0.0262), so a single protocol decision could flip the ordering in some domains. No code or data release is mentioned. None of these are fatal, but they're the difference between 'says works' and 'shown to work.'\n\nBottom line: worth a serious referee. I'd send it out, with the clear expectation that the authors fix the evaluation description, add variance, and ideally release code.","headline":"GMC is a sensible and well-ablated combination of shared semantic item IDs and a unified generative recommender; the main obstacle is an underspecified ranking protocol that makes the headline NDCG numbers not reproducible as written.","tokens_in":21140,"tokens_out":2486,"would_cite":true,"duration_ms":27304,"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":"This paper claims that multi-target cross-domain recommendation can be reframed as a generation task over shared discrete semantic codes, and that this reframing lets a single recommender improve all domains at once, with the largest…","keywords":["cross-domain recommendation","multi-target cross-domain recommendation","generative recommendation","semantic item identifiers","residual quantization","domain-aware contrastive learning","low-rank adaptation","sequential recommendation"],"falsifier":"Hold the ranking protocol fixed by either scoring every catalogue item or restricting evaluation to beam outputs, then rerun GMC against VQ-Rec and IDGenRec on the same five domains; if the NDCG advantage over the strongest baseline collapses under the fixed protocol, the claim that shared semantic codes drive the gain is not supported.","tokens_in":20163,"feed_emoji":"🎯","tokens_out":6068,"duration_ms":60871,"temperature":0.7,"pith_summary":"This paper argues that multi-target cross-domain recommendation can be solved without overlapping users, items, or auxiliary pre-training data by making the recommender generative. The proposed method, GMC, encodes each item's text into a shared set of discrete semantic codes using one residual-quantization tokenizer across all domains, then trains a single sequence-to-sequence model to generate the next item's code sequence from all domains' interaction logs. Per-domain low-rank adapters are added at fine-tuning time to preserve domain-specific behaviour without overwriting shared knowledge. On five Amazon Review subsets, GMC reports the best or near-best Recall@K and NDCG@K in almost all settings, with its largest relative gains on the smallest domains. The author's claim is that the shared code vocabulary is what makes cross-domain transfer possible, because codes, not domain-specific IDs, are the medium the recommender reasons over.","feed_headline":"Shared semantic codes let one model recommend across domains","feed_subtitle":"GMC turns item texts into shared codes and a unified generator, beating single-domain baselines on five Amazon subsets.","key_machinery":"The load-bearing object is the shared discrete semantic identifier, produced by a residual quantization variational autoencoder (RQ-VAE) with four levels and 256 codewords per level. Each item's text is first embedded by a large language model; the RQ-VAE recursively quantizes the residual error to give a coarse-to-fine code sequence, and the same codebooks are used for every domain so a codeword can be shared across domains. A domain-aware contrastive loss on quantized embeddings sharpens intra-domain similarity, which the paper shows makes first-level codes domain-specific while second-level codes mix across domains. The recommender is a unified text-to-text transformer trained to generate the next item's code sequence autoregressively from all domains' histories, with constrained generation over a prefix tree of valid item codes and beam search of size 20 at inference. Per-domain LoRA matrices are the only parameters updated during fine-tuning, adding roughly 3.7% of the model's parameters per domain.","core_discovery":"GMC's central claim is that discrete semantic item identifiers can serve as a transfer medium for multi-target cross-domain recommendation. A single RQ-VAE tokenizer maps the text-derived embeddings of items from all domains into the same L-level codebooks, so a code such as (3,2,4,1) can be shared by items in different domains; a unified transformer then treats recommendation as next-token generation over these shared codes using interaction sequences from every domain. A domain-aware contrastive loss pulls same-domain quantized embeddings together during tokenizer training, and domain-specific LoRA modules adapt the frozen unified recommender per domain. Experiments on five Amazon product categories report that GMC outperforms the strongest baseline on almost all metric-dataset combinations, with relative NDCG@5 gains up to about 25% on the smallest domain, and ablations show that removing the shared codebook, the unified recommender, the contrastive loss, or the LoRA fine-tuning each lowers performance.","pith_inferences":["Editorial inference: if the ranking procedure is made explicit and fair, the same architecture could be tested as an industrial multi-scenario recommender where text metadata is the only common signal across channels.","Editorial inference: a testable extension is to make identifier learning and recommender training end-to-end; the paper notes this direction itself, and coupling could remove the current dependence on a separately trained tokenizer.","Editorial inference: the domain-aware contrastive loss assumes intra-domain similarity helps, but in a domain containing heterogeneous categories, forcing intra-domain closeness could blur category boundaries, so one could test whether code overlap hurts when domain diversity is high.","Editorial inference: replacing the large language model text encoder with cheaper embeddings would show how much of the gain comes from text semantics versus the quantization and unified training machinery."],"forward_implications":["In non-overlapped scenarios, where no user or item is shared between domains, GMC still has a transfer channel: item text quantized into a common codebook.","Small domains benefit most: ablations show removing the shared codebook or unified recommender cuts Recall@5 by 25 to 49 percent on Scientific and Pantry.","The unified model can be adapted per domain with small parameter overhead, adding only about 0.17 million LoRA parameters per domain against a 4.3 million parameter model.","Adding more domains' data consistently improves GMC, whereas the VQ-Rec baseline does not always improve as data grows.","First-level codes encode domain membership while deeper codes carry domain-agnostic semantics, so cross-domain transfer happens at the finer granularities of the identifier."],"supporting_citations":[{"why":"Supplies the generative retrieval formulation of recommendation as next-token generation over RQ-VAE semantic identifiers, which GMC generalizes to multiple domains.","marker":"[44]"},{"why":"Provides the residual quantization algorithm used to build hierarchical discrete item identifiers.","marker":"[25]"},{"why":"Introduces VQ-Rec, a vector-quantized sequential recommender that serves as a main pre-training-dependent baseline and comparison point.","marker":"[13]"},{"why":"Introduces UniSRec, a universal sequence representation baseline using item texts, representative of pre-training-based approaches GMC claims to avoid.","marker":"[14]"},{"why":"Provides SASRec, the standard self-attention sequential baseline and the leave-one-out evaluation setting.","marker":"[23]"},{"why":"Introduces P5-CID, a generative recommender with collaborative item identifiers, contrasted with GMC's semantic identifiers.","marker":"[17]"},{"why":"Supplies the LoRA low-rank adaptation method used for domain-specific fine-tuning with a frozen unified backbone.","marker":"[15]"},{"why":"Provides the T5 text-to-text transformer used as the backbone of the unified generative recommender.","marker":"[43]"},{"why":"Introduces LC-Rec, a single-domain RQ-VAE-based generative recommender whose identifier construction and training setup GMC extends.","marker":"[75]"}],"fun_headline_variants":["Discrete codes unify multi-domain recommendation in one model","One generative model, shared item codes, many domains","Unified recommender via semantically quantized item identifiers","Cross-domain recommendation with shared semantic tokens","Generative paradigm: shared codes for multi-target recommendation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported evaluation assumes that holding out the last item and ranking it over the whole catalogue is well-defined for a model that only performs beam-search generation, and the paper does not specify a scoring rule for items outside the beam.","fun_headline_variants_meta":{"raw":{"variants":["Discrete codes unify multi-domain recommendation in one model","One generative model, shared item codes, many domains","Unified recommender via semantically quantized item identifiers","Cross-domain recommendation with shared semantic tokens","Generative paradigm: shared codes for multi-target recommendation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000672,"raw_usage":{"total_tokens":3083,"prompt_tokens":988,"completion_tokens":2095,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":2021}},"tokens_in":604,"tokens_out":2095,"duration_ms":17225,"temperature":1.0,"reasoning_tokens":2021,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:36:25.179422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold the ranking protocol fixed by either scoring every catalogue item or restricting evaluation to beam outputs, then rerun GMC against VQ-Rec and IDGenRec on the same five domains; if the NDCG advantage over the strongest baseline collapses under the fixed protocol, the claim that shared semantic codes drive the gain is not supported.","supporting_citations":[{"cited_title":"Generative Multi-Target Cross-Domain Recommendation Conference acronym ’XX, June 03–05, 2018, Woodstock, NY","cited_arxiv_id":null,"evidence_quote":"Supplies the generative retrieval formulation of recommendation as next-token generation over RQ-VAE semantic identifiers, which GMC generalizes to multiple domains."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces P5-CID, a generative recommender with collaborative item identifiers, contrasted with GMC's semantic identifiers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LoRA low-rank adaptation method used for domain-specific fine-tuning with a frozen unified backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces LC-Rec, a single-domain RQ-VAE-based generative recommender whose identifier construction and training setup GMC extends."}],"review_version":1}