{"id":"3a44514b-3540-41dd-8f86-34c9b910b000","arxiv_id":"2505.05035","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"MoDiffE applies per-view diffusion experts and a cold-aware hierarchical mixture-of-experts gate to generate cold-start bundle representations and reports large Recall@20 gains on three datasets.","lead":"This paper presents MoDiffE, a recommender that creates representations for brand-new bundles (playlists, outfits, book lists) using diffusion models and mixture-of-experts gates, and reports large gains over prior methods on three datasets. The central result depends on a gating-training trick whose data source is not defined and may leak the test set, so the gains need verification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cold-start gating augmentation (Eq. 22-23, Algorithm 3) samples pseudo-cold bundles from B_c^{B-int}, a set the paper says only appears after data splitting; if that set is the validation/test split, the reported cold-start gains rest on evaluation-set leakage.","rationale":"The reader's weakest_assumption identifies precisely the load-bearing flaw. The formal definition in Section 3.1 makes B_c^{B-int} a set of bundles with no user interactions, and Section 5.1.3 explicitly says such bundles do not exist in the original data and only appear after data splitting. Since Algorithm 3 samples from B_c^{B-int} without saying which split it uses, the only well-defined source in the experimental protocol is the validation or test portion. The ablation makes the stakes concrete: without this augmentation, cold-start performance is exactly zero, so every reported cold-start gain in Table 3 is attributable to the augmentation. If the augmentation uses evaluation-set bundles, the headline 'state-of-the-art' claim is not supported. This is a genuine correctness risk, not a disagreement with the community's consensus; it concerns internal consistency between the problem definition, the experimental protocol, and the training procedure. The proposed check—regenerating pseudo-cold bundles from training-only masked bundles and re-running the cold-start evaluation—would directly settle the question. No code is released and the backbone CrossCBR is absent from Table 3, so the ambiguity cannot be resolved by inspection. The reader's REJECT verdict is appropriate; my read does not change it.","tokens_in":29198,"tokens_out":4297,"duration_ms":43152,"concrete_test":"Check the provenance of B_c^{B-int}: rerun Algorithm 3 Stage 3 with pseudo-cold bundles generated exclusively from training-split bundles whose interactions are artificially masked (e.g., randomly hold out a fraction of each training bundle's interactions only inside the augmented BPR loss), keeping the published test split untouched. If cold-start Recall@20 in Table 3 drops to approximately the w/o aug level (near 0) or to CoHeat's level, the reported gains depend on evaluation-set leakage. Also verify whether any user u in Eq. 22 has a held-out interaction with the sampled cold bundles; if yes, that is direct test supervision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—MoDiffE sets SOTA on cold-start bundle recommendation (Section 5.2, Table 3)—rests on the cold-start gating augmentation of Section 4.4.2. The paper's own ablation (Table 4) shows that without the augmentation, cold-start Recall@20 and NDCG@20 are exactly 0 on all three datasets, so the augmentation is the entire source of cold-start competence. However, the provenance of the cold bundle set B_c^{B-int} used in Algorithm 3 (line 6-8), Eq. 22, and Eq. 23 is never specified. By the formal definition in Section 3.1, B_c^{B-int} contains bundles with no historical user-bundle interactions; Section 5.1.3 states that the original datasets contain no such bundles and that they 'only appear after a specific data splitting.' If B_c^{B-int} is taken from the validation or test portions of the 7:1:2 split—the only places such bundles exist after splitting—then Stage 3 trains the gating networks on representations and user-bundle labels drawn from the evaluation distribution. That would explain both the large cold-start gains and the ablation pattern, and it would invalidate the comparison against CoHeat and the other baselines in the cold-start and all-bundle scenarios. The paper provides no code, no error bars, and does not report the backbone CrossCBR in Table 3, so the ambiguity cannot be resolved from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MoDiffE, a framework for cold-start bundle recommendation that divides the problem into sub-problems by level and view, generates bundle/item representations with diffusion models, and combines them via a cold-aware hierarchical mixture-of-experts (MoE) gating mechanism. A multi-stage decoupled training pipeline is used, and a cold-start gating augmentation is introduced to synthesize pseudo cold bundles by interpolation. Experiments on Youshu, NetEase, and iFashion claim state-of-the-art results in cold-start, all-bundle, and warm-start scenarios, with up to 0.1027 absolute Recall@20 gain in cold-start and 47.43% relative improvement in all-bundle settings.","tokens_in":29477,"tokens_out":2484,"duration_ms":24006,"significance":"If the empirical results were sound, the paper would make a meaningful contribution by proposing a new architecture for cold-start bundle recommendation that explicitly handles dual-level multi-view complexity, and by demonstrating the usefulness of diffusion models and mixture-of-experts for this task. The divide-and-conquer framing is conceptually appealing, and the multi-stage decoupled training is a practical engineering contribution. However, the central empirical claim rests on a training-data provenance issue that, as written, appears to involve evaluation-set leakage; the absence of error bars, significance tests, and a reported backbone baseline further weakens the evidence. Because the main contribution is empirical state-of-the-art performance, the validity of that claim is the decisive factor, and it is not established by the manuscript.","major_comments":[{"comment":"The source of the cold bundle set B_c^{B-int} used for pseudo-cold bundle synthesis is unspecified and, by the paper's own definitions, cannot be part of the training split. Section 3.1 defines B_c^{B-int} as bundles with no historical user interactions, and Section 5.1.3 states that such bundles 'only appear after a specific data splitting' — meaning they are not in the training set. If the sampled bundle pairs come from the validation or test portions of the 7:1:2 split, then the gating networks are trained on the evaluation distribution, which would explain the large cold-start gains and invalidate the comparison against CoHeat and all other baselines. The ablation in Table 4 shows that without this augmentation, cold-start Recall@20 and NDCG@20 are exactly 0 on all datasets, so the augmentation is the entire source of cold-start competence. The manuscript must specify the exact provenance of the sampled cold bundles and, if they are drawn from validation/test, the experiments must be redesigned.","section":"Section 4.4.2, Eq. 23, Algorithm 3"},{"comment":"No error bars, multiple seeds, or significance tests are reported, so the claim that MoDiffE 'significantly outperforms state-of-the-art baselines' is not statistically supported. This matters particularly where the absolute differences are small, e.g., on NetEase cold-start Recall@20 MoDiffE is 0.0235 versus CoHeat's 0.0191, and on iFashion warm-start Recall@20 the difference is 0.1182 versus 0.1156. At minimum, standard deviations over several random seeds and paired significance tests (e.g., paired t-test or Wilcoxon) are required for the central claims.","section":"Section 5.2, Table 3"},{"comment":"Hyperparameters are tuned per scenario and per dataset (e.g., interpolation ratio η = 0, 0.3, 0.5 for warm-start, all-bundle, and cold-start scenarios; noise schedule; learning rates; top-n for similarity-based sampling). Without a clear validation-based selection protocol that is held fixed before evaluating on the test set, the reported results risk overfitting to the test set. Additionally, the backbone CrossCBR is not reported in Table 3, so the reader cannot verify the claimed improvements over the prior-embedding model from which MoDiffE is built.","section":"Section 5.1.5 and Table 3"}],"minor_comments":[{"comment":"The captions and panel labels for Figures 3 and 4 appear scrambled: the same sub-figure images (a-d) are duplicated across the two figures, and the caption of Figure 4 lists panels that do not match the text description. This makes the hyperparameter experiments hard to interpret.","section":"Section 5.4.1, Figure 3/4"},{"comment":"There are repeated typos, e.g., 'MoDiifE' in Section 5.2, 'CoHEAT' in the iFashion table, 'dependency on' in Section 5.3, and an incomplete sentence 'However, However, this definition contributes by explicitly emphasizes...' in Section 3.1. The paper would benefit from a careful proofread.","section":"Throughout"},{"comment":"Equation 22 defines S with notation '<b_x, b_y>' and '<b'_x, b'_y>' that is not explained; the reader cannot tell whether negative samples are also interpolated or only positives. Please clarify the notation and the construction of negative samples for the BPR loss on the augmented set.","section":"Section 4.4.1, Eq. 22"}],"recommendation":"reject","confidential_remarks":"The decisive issue is the potential evaluation-set leakage in the cold-start gating augmentation: the paper's own definitions and ablation imply that the augmentation is both necessary and sourced from bundles that are not in the training set, which would invalidate the headline results. This is not a minor fix; it requires re-running the entire experimental pipeline under a correctly specified protocol. Even if the leakage were resolved, the lack of error bars and the per-scenario hyperparameter tuning would need substantial additional validation. I recommend rejection, though a future version with a clean protocol and rigorous statistical reporting could be worthy of reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The framework is a real combination nobody has put together—per-view diffusion experts, a cold-aware hierarchical MoE, and a multi-stage decoupled pipeline with interpolation-based gating augmentation—and the divide-and-conquer framing of dual-level multi-view cold-start is a useful way to organize the problem. The paper does decent work motivating why bundle cold-start is harder than item cold-start, and the experiments cover three standard datasets.\n\nThe problem is the cold-start evaluation. The augmentation in Section 4.4.2 samples bundle pairs from B_c^{B-int}, the set of bundles with no user interactions. Section 5.1.3 says the original datasets have no such bundles; they only appear after the 7:1:2 cold-start split, which means they live in the validation and test portions. The paper never says where the sampled cold bundles come from, and if they come from the evaluation split, Stage 3 trains the gating networks on the test distribution. The ablation makes this load-bearing: without the augmentation, cold-start Recall@20 is exactly 0 on all three datasets. So the entire cold-start gain—up to 0.10 in Recall@20—comes from this mechanism, and its provenance is undefined. That is not a minor omission.\n\nThere are smaller issues. No error bars or significance tests anywhere, so 'significantly outperforms' is doing work it can't support. The main comparison omits CrossCBR, the backbone, which appears only in the ablation. Hyperparameters are tuned per scenario (eta=0/0.3/0.5), which is fine but worth stating. No code is released.\n\nI think the reader's take is essentially right. The paper is not incoherent, and the framework is salvageable: sample pseudo-cold pairs from warm bundles, or clarify that the cold bundles come from a legitimately held-out-in-training source. But as written, the central empirical claim cannot be accepted. A serious referee should see it, mostly to force the authors to pin down the provenance and re-run the experiments without potential test leakage, plus add error bars and report the backbone. I'd bring this to a reading group as a cautionary example, but I wouldn't cite it in this form.","headline":"Novel framework, but the cold-start results hinge on an augmentation whose training data source is, read literally, the test split; the central claim does not hold as written.","tokens_in":30069,"tokens_out":3858,"would_cite":false,"duration_ms":34921,"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 divide-and-conquer diffusion framework beats existing cold-start bundle recommenders.","keywords":["cold-start bundle recommendation","divide-and-conquer","diffusion model","mixture of experts","cold-aware gating","representation generation","collaborative filtering","cold-start gating augmentation"],"falsifier":"Inspect the released code or re-run the cold-start protocol with the interpolation sampling restricted to bundles that are never used for evaluation; if Recall@20 in the cold-start scenario falls toward the zero values reported for the no-augmentation ablation, the central empirical claim fails. A second check: count how many of the interpolated pairs in Stage 3 come from the validation or test bundle IDs.","tokens_in":28922,"feed_emoji":"📦","tokens_out":6775,"duration_ms":61625,"temperature":0.7,"pith_summary":"This paper proposes MoDiffE, a framework that treats cold-start bundle recommendation as a divide-and-conquer problem: split the cold-start difficulty by level (bundle and item) and by view (e.g., interaction views), solve each piece by training a diffusion model to generate the missing embedded representation directly, and combine the pieces with a cold-aware hierarchical mixture-of-experts gate. The authors claim this is the first work to formalize the dual-level multi-view complexity of the bundle cold-start problem, and that on three real-world datasets the framework outperforms existing cold-start recommenders, with up to 0.1027 absolute gain in Recall@20 in cold-start scenarios and up to 47.43% relative improvement in all-bundle scenarios. If right, it means new bundles without interaction history can be recommended immediately, without relying on hand-crafted content features for every view.","feed_headline":"Diffusion experts crack cold-start bundles","feed_subtitle":"New framework splits the task into views and levels, generates missing representations, and fuses experts for up to 0.1027 Recall@20 gain.","key_machinery":"The central mechanism is the divide-and-conquer pipeline with three load-bearing parts. First, a 'prior-embedding model' (a graph collaborative filtering backbone) produces embedded representations, which fail for feature-missing cold entities. Second, a denoising diffusion probabilistic model per view is trained to map warm embedded representations back to themselves with an L2 loss, and at inference denoises from a similarity-based anchor representation (top-$n$ most compositionally similar warm bundles' mean representation) to output a 'diffusion representation' for any input bundle, deterministic rather than diverse. Third, a cold-aware hierarchical MoE combines experts: view-layer gating (Softmax over cold-aware features such as interaction counts) fuses embedded and diffusion representations per view, and output-layer gating (Tanh) fuses view predictions. The cold-start gating augmentation creates pseudo cold bundles by interpolating pairs of cold-bundle representations (with cold-aware feature set to 0), enabling gating training for entities absent from the training set.","core_discovery":"The paper's central claim is that the bundle cold-start problem decomposes into independent sub-problems—one per level and view—and that all of them share one root failure: a prior-embedding model, given missing features, produces no usable representation ($r = f^v_\\theta(\\emptyset) = \\emptyset$). MoDiffE overcomes this by training a denoising diffusion model on the representations of warm bundles to capture their distribution, then, at inference, denoising from a similarity-based anchor representation of the cold bundle to generate a diffusion representation without requiring the missing features. A cold-aware hierarchical mixture of experts fuses embedded and diffusion representations per view (Softmax gating on cold-aware features like interaction counts) and then fuses view predictions (Tanh gating), so different cold-start situations get different expert mixes. A cold-start gating augmentation synthesizes pseudo cold bundles by linear interpolation so the gates can learn to route cold entities during training. The paper reports that this system outperforms the existing state of the art on Youshu, NetEase, and iFashion across cold-start, all-bundle, and warm-start scenarios.","pith_inferences":["One consequence the paper leaves implicit: the augmentation step requires access to genuine cold bundles during training, so a strict deployment to truly unseen bundles would need either a source of cold bundles or an augmentation scheme that synthesizes them from warm bundles; the paper does not demonstrate that warm-only augmentation preserves the gains.","A testable extension would be to replace the diffusion expert with a non-diffusion denoiser (for example, an MLP trained with the same L2 objective and anchor sampling) to isolate whether multi-step denoising, rather than representation generation in general, drives the reported improvement.","The divide-conquer-combine pattern is not tied to interaction views; the same diffusion-plus-gating structure could be applied to content or knowledge-graph views, where feature-missing entities also produce empty prior-embedding representations, though the paper only sketches this as future work."],"forward_implications":["Cold bundles with no interaction history can still receive personalized ranking, because the diffusion expert generates bundle-level and item-level representations without relying on missing features.","The same framework can be dropped into other dual-level multi-view bundle models: any prior-embedding backbone can serve as the embedding expert, with diffusion experts trained on its warm representations.","Because the gating is conditioned on cold-aware features, different bundles in different cold-start situations (cold at bundle level, cold at item level, or both) receive different weightings of embedded versus generated representations.","In the all-bundle scenario, the model improves overall Recall@20 and NDCG@20 while shifting some hits from warm bundles to cold bundles, a fairness trade-off the paper's hit analysis documents.","The cold-start gating augmentation is necessary for the framework's cold-start performance: without it, the ablation shows Recall@20 and NDCG@20 drop to 0 in the cold-start scenario on all three datasets."],"supporting_citations":[{"why":"Supplies the cold-start bundle recommendation task, the three datasets, and the strongest baseline that MoDiffE must beat.","marker":"[20]"},{"why":"The dual-level multi-view backbone (prior-embedding model) that MoDiffE extends with diffusion experts.","marker":"[36]"},{"why":"Provides the denoising diffusion probabilistic model used to generate representations from warm ones.","marker":"[19]"},{"why":"Implements the collaborative filtering signal in both interaction views via graph convolution.","marker":"[16]"},{"why":"Provides the BPR ranking loss used to train the prior-embedding model and the gating networks.","marker":"[39]"},{"why":"The mixup-style linear interpolation that cold-start gating augmentation adapts to synthesize pseudo cold bundles.","marker":"[58]"},{"why":"Fast ODE solver used to speed up diffusion inference to 20 steps.","marker":"[33]"},{"why":"Establishes the dual-level multi-view architecture for bundle recommendation that the paper formalizes.","marker":"[4]"}],"fun_headline_variants":["Diffusion experts break cold-start bundle barrier","Split by view, denoise, and fuse: cold-start bundles tamed","MoDiffE: diffusion mixture conquers bundle cold start","Cold-start bundles: divide, denoise, and decide"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The cold-start gating augmentation samples bundle pairs from the set of bundles with no user interactions, but the paper never states where those cold bundles come from at training time; if they come from the validation or test split, the reported cold-start gains would be inflated by training the gates on the evaluation distribution.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion experts break cold-start bundle barrier","Split by view, denoise, and fuse: cold-start bundles tamed","MoDiffE: diffusion mixture conquers bundle cold start","Cold-start bundles: divide, denoise, and decide"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00034,"raw_usage":{"total_tokens":1957,"prompt_tokens":1106,"completion_tokens":851,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":722,"completion_tokens_details":{"reasoning_tokens":783}},"tokens_in":722,"tokens_out":851,"duration_ms":8030,"temperature":1.0,"reasoning_tokens":783,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:15:00.007921+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the released code or re-run the cold-start protocol with the interpolation sampling restricted to bundles that are never used for evaluation; if Recall@20 in the cold-start scenario falls toward the zero values reported for the no-augmentation ablation, the central empirical claim fails. A second check: count how many of the interpolated pairs in Stage 3 come from the validation or test bundle IDs.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the cold-start bundle recommendation task, the three datasets, and the strongest baseline that MoDiffE must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The dual-level multi-view backbone (prior-embedding model) that MoDiffE extends with diffusion experts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the denoising diffusion probabilistic model used to generate representations from warm ones."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Implements the collaborative filtering signal in both interaction views via graph convolution."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the BPR ranking loss used to train the prior-embedding model and the gating networks."},{"cited_title":"Dauphin, and David Lopez-Paz","cited_arxiv_id":null,"evidence_quote":"The mixup-style linear interpolation that cold-start gating augmentation adapts to synthesize pseudo cold bundles."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Fast ODE solver used to speed up diffusion inference to 20 steps."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the dual-level multi-view architecture for bundle recommendation that the paper formalizes."}],"review_version":1}