{"id":"d608f4be-7996-49a8-8953-74e918e38046","arxiv_id":"2505.03242","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A training-free PCA and LLM-based method shifts abstract fashion queries toward concrete caption embeddings, improving text-to-image retrieval in vision-language models.","lead":"Fashion product language is about half abstract words (chic, airy), but vision-language models are trained mostly on concrete descriptions, so they retrieve worse for abstract queries. The authors propose ACT, a training-free method that rewrites the query with an LLM and shifts its embedding along directions learned from concrete captions, improving retrieval over fine-tuned baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (8) leaves the query-standardization statistics unspecified; if they are computed on the held-out query set, ACT is transductive and may lose its margin over SigLIP-ft-df.","rationale":"The reader's verdict is CONDITIONAL and I agree with that verdict; the paper should not be accepted in its current form because Eq. (8) is under-specified. I identify a slightly different load-bearing point than the reader's weakest_assumption. The reader focused on whether h_C - h_A reflects abstractness rather than caption style/length (Eq. 5). That confound is real and would weaken the conceptual claim, but it would not necessarily overturn the practical retrieval result, since a style-transfer interpretation of ACT would still be a working training-free method. The Eq. (8) standardization issue is more numerically decisive: the representation shift is the component that lifts ACT from 0.411 (Language Rewriting only) to 0.437, and the best fine-tuned baseline sits at 0.417. If the unspecified query statistics are taken from the evaluation set, the headline comparison is not apples-to-apples with models that process each query independently. I therefore ask for a single rerun with fixed training-side statistics, plus a written statement of how N(·) is computed. The supplementary's DeepFashion-to-FACAD results (Table 3) already show that language rewriting can degrade performance, reinforcing that the PCA shift carries the method's cross-dataset generality and must be cleanly specified. This does not warrant a REJECT because the issue is checkable and the paper otherwise reports controlled train/test splits for the A-C database; it does warrant retaining CONDITIONAL until the rerun is provided.","tokens_in":19953,"tokens_out":9235,"duration_ms":96073,"concrete_test":"Re-run ACT-df and ACT-facad exactly as in Table 3, but fix the standardization in Eq. (8) to statistics computed only from the training A-C database: replace N(h_q') with (h_q' - mu_A)/sigma_A, where mu_A and sigma_A are the per-dimension mean and standard deviation of the original abstract-query embeddings H_A (or, alternatively, of H_C) over the training split, and apply the same fixed statistics to every query. Do not use any held-out query embeddings when computing N(·). If H@1 drops below 0.417 for ACT-df or the cross-dataset ACT-facad falls behind SigLIP-ft-facad, the reported superiority over fine-tuned models is an artifact of transductive query normalization. If the numbers are unchanged, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim that ACT-df outperforms the best fine-tuned model (H@1 0.437 vs 0.417 for SigLIP-ft-df in Table 3) rests on the representation-shift term in Eq. (8), because the ablation in Table 5 shows Language Rewriting alone reaches only H@1 0.411, below SigLIP-ft-df. Equation (8) computes hDelta_q' = N(h_q') W W^T * sigma_Delta + mu_Delta, where N(·) is \"standardization using the mean and standard deviation among query representations.\" The paper never states where those query mean/std statistics come from. If they are estimated from the evaluation query set, then ACT is transductive: the embedding of each test query is centered and scaled using all other test queries, a form of test-set information that zero-shot and fine-tuned baselines do not receive. The preparation-phase statistics (mu_Delta, sigma_Delta) are not the query statistics, so Eq. (8) cannot be reproduced from the described training data alone. Since the gap over the best fine-tuned baseline is only 0.02 H@1, and the representation-shift component is exactly the transductive part, this ambiguity is load-bearing. The supplementary (Sec. C.2) further shows that the LLM-rewriting component alone can hurt in the DeepFashion-to-FACAD direction, making the shift term necessary for the claimed cross-dataset generality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies abstract-oriented language in vision-language models, focusing on the fashion domain. It presents a statistical analysis of adjectives in DeepFashion, FACAD, and LAION-400M, arguing that abstract adjectives are frequent in fashion descriptions, carry information not captured by concrete attributes, and are useful for text-to-image retrieval. It then proposes ACT (Abstract-to-Concrete Translator), a training-free method that (i) builds an Abstract-Concrete database by captioning fashion images with a frozen VLM, (ii) fits a PCA-based shift between text embeddings of original abstract descriptions and generated concrete captions, and (iii) at inference rewrites the user query with an LLM, embeds it, and applies the learned shift. The paper reports that ACT outperforms zero-shot and fine-tuned VLMs in same-dataset and cross-dataset settings on DeepFashion, and shows consistent gains across multiple VLM families and scales. The main empirical claim is ACT-df reaching H@1 = 0.437 on DeepFashion, above SigLIP-ft-df at 0.417, and ACT-facad reaching H@1 = 0.428 in the FACAD-to-DeepFashion setting.","tokens_in":20275,"tokens_out":7528,"duration_ms":73425,"significance":"If the central claim is validated, the paper is a useful and interesting contribution: it provides a large-scale empirical analysis of abstract versus concrete language in fashion VLMs and demonstrates a plug-and-play, training-free alternative to fine-tuning for closing part of the abstract-language performance gap. The paper ships code and evaluates across many backbones (SigLIP, O-CLIP, EVA-CLIP) and model sizes, which makes the empirical basis unusually broad. The finding that a linear PCA shift in text-embedding space, combined with LLM rewriting, can transfer across datasets is a genuinely useful result. However, the significance is contingent on resolving the reproducibility issue in Eq. (8) and on a more balanced statement of the cross-dataset claims, since the supplementary results in the opposite direction (DeepFashion-to-FACAD) do not show the same superiority.","major_comments":[{"comment":"Equation (8) is not reproducible as written. The standardization N(h_q') is described as using 'the mean and standard deviation among query representations,' but the paper never states where those statistics come from. If they are computed on the evaluation query set, then ACT is transductive: the embedding of each test query is centered and scaled using information from all other test queries, a form of test-set information that is not available to the zero-shot or fine-tuned baselines. This issue is load-bearing: Table 3 shows ACT-df exceeds SigLIP-ft-df by H@1 = 0.437 versus 0.417, while Table 5 shows that the language-rewriting component alone reaches only H@1 = 0.411. The advantage therefore comes from the representation-shift term, which is exactly the potentially transductive part. The authors must specify whether the query mean and standard deviation are computed on the training split, on the A-C database, or on the evaluation query set, and they should report results with query statistics restricted to the training split only.","section":"Sec. 4, Eq. (8)"},{"comment":"The abstract and Section 5 claim that ACT 'outperforms the fine-tuned VLMs in both same- and cross-dataset settings,' but the DeepFashion-to-FACAD cross-dataset results in Supp. Table 3 contradict this: ACT-df achieves H@1 = 0.426 and ACT-df without language rewriting achieves H@1 = 0.460, both below SigLIP-ft-df at H@1 = 0.482. This is not a minor gap; it is a direct counterexample to the unqualified claim and also shows that the LLM-rewriting component can substantially hurt in this direction. The paper should either restrict the cross-dataset claim to the FACAD-to-DeepFashion direction, or add a balanced analysis of the asymmetry and explain under what conditions ACT can be expected to help.","section":"Supp. C.2 and Abstract"},{"comment":"The central interpretation that the PCA shift W captures an abstract-to-concrete semantic shift is not yet established, because the paired differences h_C - h_A also encode the captioning model's shorter length, simpler syntax, restricted vocabulary, and the LLM's stylistic rewriting. The paper does not provide a control that separates abstractness from these confounds, so the retrieval gains could be explained by distribution matching toward caption-style text rather than by translating abstract meaning. A concrete diagnostic would be to apply the shift to queries that are already fully concrete (or whose LLM rewriting contains no abstract adjectives); if W is a genuine abstract-to-concrete map, the shift should be approximately inert on such queries. At minimum, the paper should report length and lexical statistics for original, rewritten, and captioned text, and show that the selected PCA directions are tied to abstract-concrete contrasts rather than to sentence length or caption style.","section":"Sec. 4, Eqs. (4)-(6), and Table 5"}],"minor_comments":[{"comment":"The expression 'W W^T * sigma_Delta + mu_Delta' mixes matrix and vector notation; please specify whether sigma_Delta is a scalar, a per-component vector, or a diagonal matrix, and how the addition broadcasts. This is needed for exact reproducibility.","section":"Eq. (8)"},{"comment":"The manual reclassification of context-dependent adjectives such as 'cool' and 'sharp' is a reasonable domain adaptation, but the number of reclassified words and a robustness check without the manual edits would help assess how much the abstract/concrete statistics depend on this choice.","section":"Supp. B.2"},{"comment":"The main comparisons are reported without error bars or a significance test; the headline advantage over SigLIP-ft-df is 0.020 in H@1, so a confidence interval, per-seed variance, or a paired test would substantially strengthen the claim that ACT 'outperforms' the best fine-tuned model.","section":"Sec. 5, Table 3"},{"comment":"The DeepFashion-to-FACAD results and the associated failure analysis of language rewriting should be moved from the supplementary into the main text, because they materially qualify the cross-dataset claim made in the abstract and in Section 5.1.","section":"Sec. 5.1 and Supp. C.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the reader's report and the stress-test note both correctly identify Eq. (8) as a load-bearing ambiguity, and I concur. The manuscript's strongest advertised result—outperforming fine-tuned models in both cross-dataset directions—is not supported by the supplementary's opposite-direction experiment. Both issues are fixable by re-analysis and rewording, so I do not recommend rejection, but the authors should not proceed without addressing them."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper earns its place. It quantifies something concrete: fashion descriptions are rich in abstract adjectives, while LAION (and likely other pre-training corpora) is concrete-heavy, and current VLMs retrieve better from concrete captions than from the original abstract ones. The proposed fix, ACT, is a sensible training-free combination of LLM rewriting plus a PCA shift in text-embedding space, and the experiments support it across several backbones and ablations. The main-table result (ACT-df H@1 0.437 vs SigLIP-ft-df 0.417) is a real, if modest, win, and the cross-dataset direction FACAD-to-DeepFashion also holds up. This is worth referee time.\n\nNow the soft spots, in proportion. The biggest is Eq. (8). The paper standardizes the query embedding using \"the mean and standard deviation among query representations\" and never says which query set. If those statistics are computed on the evaluation queries, the method is transductive, and the 0.02 H@1 margin over the best fine-tuned model could evaporate. This is a genuine ambiguity and load-bearing because the ablation shows the representation shift is what pushes the method past the fine-tuned baseline. The fix is easy—use training-set statistics or state clearly that they are estimated from the A-C database—but it has to be done.\n\nThe second issue is an overclaim. The abstract says ACT outperforms fine-tuned VLMs in same- and cross-dataset settings, but the supplement's reverse direction (DeepFashion-to-FACAD) shows ACT-df below SigLIP-ft-df, and the LLM rewriting alone hurts there. That result is disclosed honestly, but the abstract and conclusions don't reflect it. The paper should be reworded to say \"one cross-dataset direction.\"\n\nThe oracle analysis in Fig. 3-left lacks frequency control: abstract adjectives may look more discriminative because they are rarer, not because of abstractness. That is a minor concern given the rest of the evidence, but worth a correction. The k=600 choice also gets no sensitivity analysis; minor.\n\nCitation pattern is fine; the related work is fair and the novelty claim is honest. This is a solid paper with one fixable technical ambiguity and one framing problem. I'd send it to serious peer review, require the Eq. (8) statistics to be specified and the cross-dataset claim to be narrowed, and then it should be acceptable.","headline":"A real and measurable abstract-language bias in VLM text encoders, with a mostly working training-free corrective shift that needs one equation clarified and one overclaim reined in.","tokens_in":20802,"tokens_out":1915,"would_cite":true,"duration_ms":20279,"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":"CLIP-style models under-represent abstract fashion language, and a training-free PCA shift can close most of that gap, beating fine-tuned models on retrieval.","keywords":["vision language models","abstract language","text-to-image retrieval","fashion retrieval","representation shift","training-free adaptation","LLM rewriting","PCA translation"],"falsifier":"Build a control set in which the captioning-model captions are length- and style-matched to the original descriptions, for example by truncating or padding them, and recompute the PCA shift. If the retrieval gain disappears, the shift encodes captioning style rather than abstractness. As a second check, apply the same shift to the LLM-rewritten, already-concrete queries: if it lifts retrieval as much as it lifts abstract queries, the shift is not specifically translating abstract meaning.","tokens_in":19719,"feed_emoji":"👗","tokens_out":9061,"duration_ms":81434,"temperature":0.7,"pith_summary":"Fashion descriptions are rich in abstract adjectives—words like 'sexy', 'airy', and 'chic'—that occur about as often as concrete ones, carry information concrete attributes do not, and make text queries more discriminative. The paper argues that CLIP-style vision-language models under-represent this abstract language because their pre-training text is concrete-heavy, and that fine-tuning does not fully repair the gap. It proposes ACT, a training-free pipeline that has an LLM rewrite an abstract query toward concrete wording and then adds a PCA-derived shift in the VLM's text-embedding space, moving the query toward the embeddings of concrete captions. On text-to-image retrieval in the fashion domain, ACT beats zero-shot and fine-tuned baselines in both same-dataset and cross-dataset settings. If right, this means much of the abstract-language gap in such models is a systematic, low-dimensional embedding shift that can be corrected without retraining.","feed_headline":"Shifting embeddings beats fine-tuning for abstract fashion queries","feed_subtitle":"A training-free PCA shift moves vague fashion descriptions toward concrete ones, lifting retrieval without weight updates.","key_machinery":"The load-bearing mechanism is the Abstract-to-Concrete Translator (ACT), a two-phase, training-free procedure. In preparation, paired embeddings $h^C_s$ and $h^A_s$ are obtained from captioning-model captions and original abstract descriptions of the same items; their standardized difference, $\\Delta_{\\text{A-C}} = H_C - H_A$, is projected with Principal Component Analysis onto its top $k = 600$ directions, giving a shift projector $W$ along with mean and standard-deviation statistics. At inference, the query is first rewritten by a frozen LLM, embedded, then shifted by $\\hat{h}^C_q = h_{q'} + (N(h_{q'}) W W^\\top \\sigma_\\Delta + \\mu_\\Delta)$. This addition is what carries the argument: a single linear correction, learned once per dataset and VLM, moves abstract query embeddings toward the concrete caption region of the latent space, and cosine retrieval with the shifted embedding outperforms fine-tuned models.","core_discovery":"The central claim is that current VLMs, both general-purpose and fashion-specific, systematically under-represent abstract-oriented language, and that the gap can be largely closed without weight updates. The paper builds an Abstract-Concrete database by captioning fashion images with a frozen captioning model, then computes the embedding difference between the captioning-model captions and the original abstract descriptions for the same images. The dominant directions of this difference, extracted with PCA, define a shift that is added to the query representation at inference, after an LLM has already rewritten the query in more concrete terms. On DeepFashion retrieval, ACT-df reaches H@1 of 0.437, above the best fine-tuned model (SigLIP-ft-df at 0.417) and far above zero-shot SigLIP at 0.311; in the cross-dataset setting, ACT-facad reaches 0.428, above every fine-tuned baseline. The paper also claims the improvement is consistent across model families and scales, with an average +4.9% H@1 gain.","pith_inferences":["If the shift is a genuine semantic correction, the same preparation could be run for other under-represented registers—metaphor, emotive, subculture-specific—turning ACT into a generic 'domain dialect' adapter for CLIP-like encoders.","The dependence on a captioning model means ACT's notion of 'concrete' is whatever that model's vocabulary captures; using human-written concrete paraphrases or a captioner with controlled length/style would test how much of the gain is style matching.","The supplementary's FACAD failure suggests the LLM rewriting step, not the PCA shift, is the bottleneck for terse, ungrammatical queries; a rewriter trained for concise fashion text would likely extend ACT's gains there.","Since the shift is computed once per VLM and dataset and applied as an embedding addition, it could be shipped as a lightweight per-domain plug-in, potentially transferring to composed retrieval or text-to-video search without fine-tuning."],"forward_implications":["Same-dataset retrieval on DeepFashion would rise from 0.311 to 0.437 H@1 with SigLIP, beating the best fine-tuned model (0.417) with no weight updates.","Cross-dataset use, building the shift on FACAD and testing on DeepFashion, would reach 0.428 H@1, nearly matching the same-dataset version and beating all fine-tuned baselines.","The gain would generalize across model families and scales, with an average +4.9% H@1 over SigLIP, CLIP, O-CLIP and EVA-CLIP backbones.","Both pipeline components would be necessary: LLM rewriting alone adds about +10% H@1 and the representation shift alone about +3.7% H@1, with the best result when combined.","Fashion text-to-image search would work with natural, abstract user phrasing rather than requiring users to supply concrete attribute lists."],"supporting_citations":[{"why":"Supplies DeepFashion, the main evaluation dataset and the source of abstract-oriented seller descriptions for ACT-df.","marker":"[24]"},{"why":"Supplies FACAD, the cross-dataset source for the FACAD-to-DeepFashion setting and its captions.","marker":"[39]"},{"why":"Supplies the large web-crawled corpus whose concrete-heavy text motivates the under-representation claim.","marker":"[28]"},{"why":"Supplies the CLIP-style text and image encoders that ACT modifies and that define the latent space.","marker":"[26]"},{"why":"Supplies the F-CLIP fashion-specific baseline that ACT must beat.","marker":"[8]"},{"why":"Supplies the OF-CLIP baseline and the retrieval protocol, including H@K metrics, that the paper follows.","marker":"[6]"},{"why":"Supplies the concreteness lexicon and the 3.0 threshold used to classify adjectives as abstract or concrete.","marker":"[5]"},{"why":"Supplies the Qwen2-VL captioning model used to generate concrete paired descriptions for the A-C database.","marker":"[35]"},{"why":"Supplies the Llama-3.1 LLM used for rewriting abstract queries toward concrete language at inference.","marker":"[11]"},{"why":"Supplies the PCA method used to extract the dominant abstract-to-concrete representation shifts.","marker":"[29]"}],"fun_headline_variants":["Training-free shift beats fine-tuning for abstract fashion queries","PCA shift, not fine-tuning, boosts abstract fashion retrieval","Abstract language gap in VLMs closed by training-free shift","Shift, don't fine-tune: training-free win for abstract queries","VLM abstract text under-represented? Simple shift fixes retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the difference between captioning-model captions and original abstract descriptions is dominated by the abstract-to-concrete semantic shift, and that one linear direction learned from that difference transfers to new queries, rather than reflecting the captioning model's shorter length, repetitive phrasing, or narrower vocabulary.","fun_headline_variants_meta":{"raw":{"variants":["Training-free shift beats fine-tuning for abstract fashion queries","PCA shift, not fine-tuning, boosts abstract fashion retrieval","Abstract language gap in VLMs closed by training-free shift","Shift, don't fine-tune: training-free win for abstract queries","VLM abstract text under-represented? Simple shift fixes retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000743,"raw_usage":{"total_tokens":3344,"prompt_tokens":1004,"completion_tokens":2340,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":2256}},"tokens_in":620,"tokens_out":2340,"duration_ms":17624,"temperature":1.0,"reasoning_tokens":2256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:56:47.260341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a control set in which the captioning-model captions are length- and style-matched to the original descriptions, for example by truncating or padding them, and recompute the PCA shift. If the retrieval gain disappears, the shift encodes captioning style rather than abstractness. As a second check, apply the same shift to the LLM-rewritten, already-concrete queries: if it lifts retrieval as much as it lifts abstract queries, the shift is not specifically translating abstract meaning.","supporting_citations":[{"cited_title":"Deepfashion: Powering robust clothes recog- nition and retrieval with rich annotations","cited_arxiv_id":null,"evidence_quote":"Supplies DeepFashion, the main evaluation dataset and the source of abstract-oriented seller descriptions for ACT-df."},{"cited_title":"Fashion captioning: Towards generating accurate descrip- tions with semantic rewards","cited_arxiv_id":null,"evidence_quote":"Supplies FACAD, the cross-dataset source for the FACAD-to-DeepFashion setting and its captions."},{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the CLIP-style text and image encoders that ACT modifies and that define the latent space."},{"cited_title":"Contrastive language and vi- sion learning of general fashion concepts","cited_arxiv_id":null,"evidence_quote":"Supplies the F-CLIP fashion-specific baseline that ACT must beat."},{"cited_title":"Openfash- ionclip: Vision-and-language contrastive learning with open- source fashion data","cited_arxiv_id":null,"evidence_quote":"Supplies the OF-CLIP baseline and the retrieval protocol, including H@K metrics, that the paper follows."},{"cited_title":"Concreteness ratings for 40 thousand generally known en- glish word lemmas","cited_arxiv_id":null,"evidence_quote":"Supplies the concreteness lexicon and the 3.0 threshold used to classify adjectives as abstract or concrete."}],"review_version":1}