{"id":"62d023c9-e8be-480b-9d0e-2a5283027d2a","arxiv_id":"2608.12570","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MM-slotgate factorizes Fashion-CLIP embeddings into four named attribute slots with per-slot modality gates, improving constrained fashion retrieval on H&M over equal-weight multimodal fusion.","lead":"MM-slotgate splits a product's text and image into four named attribute slots, each with a learned preference for text or image evidence. On H&M data it reports better constrained fashion retrieval than equal-weight multimodal fusion, with the color slot leaning on images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unreported alpha in Eq. (4) could dominate the combined score, so the headline CS@10 gains may measure the supervised alignment heads rather than the learned gates; reporting alpha* and an alpha=0 ablation is required.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: alpha is tuned by cross-validation but never reported, so the combined score may be dominated by the supervised logit term. This is a genuine attribution problem because CS@10 is exactly the criterion the logit heads are trained to optimize. The paper explicitly says alpha is tuned by 5-fold CV but gives no value and no alpha=0 ablation, so the reader cannot determine whether the headline gains come from the learned gates and slot similarity or from the alignment heads. The negative-control and gate-interpretability results are real supporting evidence, but they do not resolve the magnitude of alpha. The correct response is to keep the CONDITIONAL verdict and require the missing alpha reporting and ablation, not to reject the paper outright. I therefore recommend UNCHANGED relative to the reader's verdict.","tokens_in":9019,"tokens_out":4348,"duration_ms":49377,"concrete_test":"Report alpha* from the 5-fold CV for every method, and rerun Tables 1 and 2 with alpha=0 (similarity-only) and with a common alpha fixed across methods. Also compute CS@10 using only the cosine term and only the logit term. If the learned-vs-fixed_half gap (0.005) or the color gain (0.568) collapses under alpha=0, the headline should be reattributed to the supervised logit heads rather than to the learned gates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on the combined retrieval score of Eq. (4): score = cos(w(q),w(i)) + alpha * sum_{s in C} log P(y_s = c* | i), with alpha* tuned by 5-fold CV. The paper never reports alpha* and never ablates alpha=0. Because CS@10 only checks whether all constrained slots in the top-10 match the query's ground-truth labels, a sufficiently large alpha makes the ranking an item-side classifier ranking: top items are those whose alignment heads assign high probability to the query's target classes, and the query-specific cosine similarity becomes nearly irrelevant. Under that regime, the 0.7566 headline and the color 0.321->0.889 improvement reflect the supervised alignment heads' classification accuracy (color acc 0.762 in Table 4) rather than the learned multimodal gates or the slot-similarity representation. The fCLIP text-only baseline has color classifier accuracy near chance (0.304), so the color gain could be entirely a classification effect of adding image input, not a retrieval or controllability effect. The learned-vs-fixed_half margin is also only 0.005 with no error bars. Without reporting alpha*, an alpha=0 ablation, or a per-term score breakdown, the attribution of the headline result to gated slots is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"MM-slotgate factorizes Fashion-CLIP text and image embeddings into four named attribute slots (pattern, color, category, demographic). Each slot has a learnable scalar gate over the two modalities, a VQ codebook, and a supervised alignment head. Retrieval uses a combined score of slot-weighted cosine similarity plus a tuned weighted sum of constrained-slot log probabilities. On a 50K-item H&M subset with a 5K held-out gallery, the paper reports macro CS@10 of 0.7566, beating MM-global fusion (0.7142) and fCLIP text-only (0.4755/0.4765), with the largest gain on color; it also reports learned gate values, a shuffled-image negative control, linear-probe leakage measurements, and codebook intervention lifts.","tokens_in":9290,"tokens_out":5643,"duration_ms":47787,"significance":"If the attribution is established, the paper contributes a simple, interpretable mechanism for attribute-conditioned modality selection in fashion retrieval, with a clear evaluation metric and useful diagnostics (oracle upper bound, negative control, label-prior leakage baseline). The architecture and training protocol are specified in enough detail to be reimplemented. However, the headline comparison currently depends on an unreported tuning weight, and the key learned-versus-fixed ablation lacks error bars, so the central claim is not yet fully supported.","major_comments":[{"comment":"The combined retrieval score is cos(w(q),w(i)) + alpha * sum_{s in C} log P(y_s = c*_s | i), with alpha tuned by 5-fold cross-validation, but the paper never reports the selected alpha* and never ablates the logit term (alpha=0). The concern that this term could dominate the ranking is well-founded: with large alpha, the top-10 items are those whose alignment heads assign high probability to the query's target classes, so the headline CS@10 gains, including color 0.321 to 0.889, could reflect the supervised alignment heads' classification accuracy rather than the learned modality gates or the slot-similarity representation. Please report alpha*, add an alpha=0 ablation, and ideally provide a per-term score breakdown so that the cosine-only and logit-only contributions can be compared.","section":"Section 3, Eq. (4), and Section 4"},{"comment":"All results are point estimates from a single training run and a single 5K gallery, with no error bars, multiple seeds, or significance tests. The central ablation claim that learned gates beat fixed equal-weight fusion rests on a 0.005 CS@10 difference (0.7566 vs 0.7516) and on very close validation losses and color accuracies in Table 4. This margin is not interpretable without variance estimates; please provide multiple seeds or a significance test for the main comparisons, at minimum for learned versus fixed_half, MM-slotgate versus MM-global, and MM-slotgate versus fCLIP-text-only.","section":"Tables 1-4, Sections 4-5"}],"minor_comments":[{"comment":"The fCLIP-text-only macro CS@10 is reported as 0.4755 in the abstract and conclusion but as 0.4765 in Tables 1 and 2, and the relative gain in Section 1 (59.1%) is inconsistent with the 58.8% computed from the table; please reconcile all occurrences.","section":"Abstract, Section 1, Tables 1-2"},{"comment":"The shuffled-image negative control is reported only through validation loss and gate values; please also report CS@10 under the shuffle, since the claim that correctly aligned visual content drives the retrieval gain depends on showing that the retrieval metric degrades, not just that the model leans more on text.","section":"Section 5, negative control"},{"comment":"The intervention metrics Hit@10, Null@10, and PreserveDelta are described only in prose; please give formal definitions and, if space permits, error estimates, since the color lift of 15.3x is a headline number.","section":"Section 7, intervention"},{"comment":"The orthogonality penalty Lorth is mentioned but its exact form (which Gram matrix, which off-diagonal terms, how normalized) is not specified; please state it precisely.","section":"Section 3, Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The unreported alpha* in Eq. (4) is the main risk: if the authors can show the headline results hold for alpha=0 or when the cosine term is dominant, the paper would be substantially strengthened. The small learned-versus-fixed margin also needs error bars. The topic fits RecSys, and the experimental design is otherwise clear and well-scoped."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the per-slot text-image gate is a real, easy-to-describe idea that I haven't seen in the semantic-ID or fashion-retrieval literature, and the paper has some careful validation (shuffled-image negative control, label-prior leakage baseline). But the headline retrieval numbers lean on a combined score whose second term is weighted by alpha*, tuned by CV and never reported, and there is no alpha=0 ablation. That is the soft spot that matters.\n\nThe architecture is straightforward: fCLIP text and image embeddings are projected per slot, mixed by a learned scalar gate, then quantized. The gate converges to sensible values (color image-leaning, category text-leaning) without modality supervision. That is a nice, clean result. The negative control -- shuffling images makes all gates text-leaning -- is the kind of check too many papers skip. And the leakage analysis, comparing off-diagonal probe AUC to the label-prior baseline, is the right way to think about disentanglement.\n\nThe retrieval evaluation is where I would want changes. Eq. (4) adds alpha times the query-constrained slot logits to the cosine term. The paper says alpha* is tuned by 5-fold CV but does not report it, and never shows the alpha=0 case. If alpha is large, the ranking reduces to item-side classifier confidence and the cosine term is decoration. The stress-test note worries this means the color gain (0.321 to 0.889) is mostly the alignment head's image classification ability, not the gate. That is a legitimate worry, but note that all methods use the same combined score, so the relative order between MM-slotgate and MM-global might survive. Still, without alpha, the absolute attribution is unclear.\n\nThe other issues are smaller: no error bars on any CS@10 number; the learned-gates-versus-fixed_half margin is 0.005 (almost certainly noise); and the abstract says fCLIP-text-only is 0.4755 while Table 1 says 0.4765. No code release either.\n\nThe paper itself is honest about the small learned-gate ablation margin, so I do not think there is misconduct; I think it is an under-reported hyperparameter that needs a simple ablation.\n\nWho is it for? Anyone working on controllable multimodal retrieval or semantic IDs. It is a useful incremental contribution, not a field-reorganizing one. With alpha reported and an alpha=0 row added, the claims would be much easier to evaluate. I would send it to peer review and ask for that before acceptance.","headline":"A genuinely new per-slot modality gate, but the headline CS@10 numbers lean on an unreported cross-validated logit weight that needs an ablation before the main claim is trusted.","tokens_in":9817,"tokens_out":3942,"would_cite":false,"duration_ms":38673,"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":"MM-slotgate factorizes Fashion-CLIP embeddings into four gated attribute slots and shows on H&M that this beats global fusion and text-only retrieval for constraint-satisfying fashion retrieval.","keywords":["controllable retrieval","multimodal fusion","slot factorization","semantic IDs","vector quantization","fashion retrieval","attribute-conditioned","ConstraintSatisfied@10"],"falsifier":"Run the same retrieval evaluation with the alpha term in the combined score set to zero (slot similarity only) and with the learned gates replaced by text-only; if slot-similarity-only MM-slotgate drops toward the text-only baseline of 0.4765 or no longer beats equal-weight fusion, then the claimed gating contribution is not established.","tokens_in":8826,"feed_emoji":"👗","tokens_out":4826,"duration_ms":40774,"temperature":0.7,"pith_summary":"The paper tries to establish that fashion retrieval should represent an item as a set of named attribute slots, pattern, color, category, and demographic, rather than as one global embedding, and that each slot should learn its own mixture of text and image evidence. On the H&M catalog, the proposed MM-slotgate achieves 0.7566 macro ConstraintSatisfied@10, beating equal-weight multimodal fusion (0.7142) and text-only Fashion-CLIP retrieval (0.4755). The largest single gain is color, which rises from 0.321 to 0.889 as the learned color gate assigns 57.4 percent weight to image evidence. If correct, this means controllable multi-attribute fashion retrieval can be built from typed, addressable slots instead of opaque item-level codes.","feed_headline":"Four gated attribute slots outrank single-vector fashion retrieval","feed_subtitle":"Constraint satisfaction jumps from 0.321 to 0.889 on color when the model learns to lean on images.","key_machinery":"The central object is the per-slot scalar gate $g_s = \\sigma(a_s)$, with $g_s$ as the text weight and $1-g_s$ as the image weight, in the fusion equation that mixes slot-specific text and image projections before L2 normalization and vector quantization. The gates let the model allocate modality evidence attribute-by-attribute, and they are trained end-to-end with the alignment loss on four attribute labels and the VQ commitment loss. Retrieval uses the continuous slots through a slot-weighted similarity plus the sum of the constrained slots' classifier logits, while intervention uses the quantized slot codes from each slot's codebook.","core_discovery":"MM-slotgate claims that supervised slot factorization plus per-slot learnable text-image gates gives better constraint-satisfying fashion retrieval than either a single global multimodal embedding or text-only retrieval. Each of the four slots receives a scalar gate $g_s = \\sigma(a_s)$, initialized at 0.5 and trained end-to-end, that controls how much text versus image enters that slot's vector-quantized bottleneck. The learned gates converge without modality supervision: color is image-leaning, category is text-leaning, pattern is balanced, and demographic is slightly text-leaning. On H&M the method reaches 0.7566 macro CS@10, outperforms equal-weight fusion (0.7142) and text-only retrieval (0.4755), improves color from 0.321 to 0.889, and supports codebook-level intervention with a 15.3x lift for color while adding no measured excess cross-slot leakage beyond label correlations.","pith_inferences":["Because the gate values are a function of each catalog's label-text-image correlations, the specific learned preferences (color image-leaning, category text-leaning) may not transfer to catalogs with different taxonomy or image quality; a testable extension is to re-learn gates per dataset and compare.","The combined score's alpha is tuned by cross-validation; an editorial extension would be to test whether a fixed alpha (including alpha=0) preserves the ordering, which would determine whether the slot-similarity term or the logit term drives the headline gains.","The same slot-gating mechanism could be applied to richer taxonomies such as occasion, style, or brand, though retrieval and intervention complexity would grow with the number of slots.","A direct follow-up experiment would ablate image availability per attribute class to verify that the missing-image text fallback does not systematically bias the learned gates toward text."],"forward_implications":["Multi-attribute fashion queries can be satisfied without hand-tuning modality weights: each slot learns whether text or image should dominate.","Color retrieval, the weakest text-only attribute, becomes the strongest after image grounding: CS@10 for color rises from 0.321 to 0.889.","Both modalities are necessary; text-only collapses on color accuracy (0.304, near chance) and image-only degrades on category and demographic.","Quantized slots permit targeted attribute steering at retrieval time without recomputing embeddings, with color lift rising from 1.5x in the text-only encoder to 15.3x.","The learned gates are interpretable and match intuitive modality needs, and the representation adds no measured excess leakage beyond label co-occurrence."],"supporting_citations":[{"why":"Fashion-CLIP backbone supplies the text and image embeddings that each slot projects and gates.","marker":"[2]"},{"why":"H&M catalog supplies the 50K-item subset, attribute labels, and train/held-out split used for evaluation.","marker":"[6]"},{"why":"Straight-through estimator makes the vector-quantization bottleneck trainable, enabling the slot codebooks.","marker":"[1]"},{"why":"Label-correlation probe baseline is used to measure excess cross-slot leakage.","marker":"[4]"},{"why":"CLIP's contrastive image-text alignment is the foundation of the Fashion-CLIP backbone used here.","marker":"[9]"},{"why":"TIGER is the semantic-ID approach the paper contrasts with its named attribute slots.","marker":"[10]"}],"fun_headline_variants":["Gated attribute slots beat single-vector fashion retrieval","Color retrieval improves from 0.32 to 0.89 via learned gate","Four named slots with learned gates control fashion search","MM-slotgate: interpretable slots outperform fused embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline numbers come from a combined score that adds a tuned amount of the slot classifiers' confidence scores to the slot similarity score; the paper never reports how large that tuned amount is, so if it is large the gains may come from the classifiers rather than from the learned gates.","fun_headline_variants_meta":{"raw":{"variants":["Gated attribute slots beat single-vector fashion retrieval","Color retrieval improves from 0.32 to 0.89 via learned gate","Four named slots with learned gates control fashion search","MM-slotgate: interpretable slots outperform fused embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":2056,"prompt_tokens":1064,"completion_tokens":992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":923}},"tokens_in":680,"tokens_out":992,"duration_ms":9377,"temperature":1.0,"reasoning_tokens":923,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:05:09.188540+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same retrieval evaluation with the alpha term in the combined score set to zero (slot similarity only) and with the learned gates replaced by text-only; if slot-similarity-only MM-slotgate drops toward the text-only baseline of 0.4765 or no longer beats equal-weight fusion, then the claimed gating contribution is not established.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"H&M catalog supplies the 50K-item subset, attribute labels, and train/held-out split used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Label-correlation probe baseline is used to measure excess cross-slot leakage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TIGER is the semantic-ID approach the paper contrasts with its named attribute slots."}],"review_version":1}