{"id":"38ec1745-7ad7-44bc-86ef-c3f744e1aa73","arxiv_id":"2412.10680","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A CLIP prompt-based adapter that generates target prompts for unseen domains/classes by attending over masked source prompts, yielding modest average improvements over ProS.","lead":"UCDR-Adapter adds trainable domain/class prompts and a masked-attention prompt generator to a frozen CLIP image encoder, aiming to retrieve images from domains and classes never seen during training. On standard UCDR benchmarks it reports small average gains over the ProS baseline, while losing on some individual settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own tables do not support the 'consistently outperforms ProS' claim: UCDR-Adapter loses to ProS on Quickdraw and Infograph (UCDR), Sketch and several precision rows (UdCDR), and TU-Berlin (UcCDR), with no significance testing.","rationale":"The reader's conditional verdict is appropriate, and my stress-test supports it. The most load-bearing condition for the central claim is empirical superiority over the ProS baseline, and the paper's own tables contradict a strong reading of 'consistently outperforms ProS': losses on Quickdraw and Infograph in Table 2, on Sketch and several precision rows in Table 3, and on TU-Berlin in Table 4. The average margins are small, and without seeds or significance tests the headline gain could be noise. This is closely related to the reader's weakest assumption about the TPG module, because the only evidence for that mechanism is the same mixed benchmark set. I also flag the Eqs. (7)-(8) matrix swap as a concrete reproducibility issue, and the author overlap with ProS as a reason to require independent reruns rather than copied baseline numbers. None of this requires changing the verdict: the paper remains conditionally acceptable if the claims are weakened, statistical validation is added, and the code is released. I therefore set verdict_should_be to UNCHANGED and agreement_with_reader to partial, since the reader's stated weakest assumption is about TPG's span while my primary concern is the direct empirical overclaim, though the two are intertwined.","tokens_in":14036,"tokens_out":10734,"duration_ms":97760,"concrete_test":"Re-run the full UCDR, UdCDR, and UcCDR protocols from Tables 2-4 at least five times with different random seeds, re-running ProS under the same splits and using the released code, and report per-domain mean +/- std with paired Wilcoxon signed-rank tests. The abstract and conclusion should be revised unless UCDR-Adapter is significantly better than ProS on the majority of per-domain rows; in the same rerun, verify from code whether target prompts are computed as U'wd and V'wc or as the dimensionally inconsistent forms written in Eqs. (7)-(8).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim requires that UCDR-Adapter reliably beats the ProS baseline. The reported numbers do not establish this. In Table 2, ProS beats UCDR-Adapter on Quickdraw (mAP@200 0.2842 vs 0.2794) and Infograph (0.5798 vs 0.5714). In Table 3, ProS beats UCDR-Adapter on Sketch (mAP@200 0.7385 vs 0.7332; Prec@200 0.4911 vs 0.4893) and on several precision rows (Quickdraw, Infograph). In Table 4, ProS beats UCDR-Adapter on TU-Berlin (mAP@All 0.6675 vs 0.6581; Prec@100 0.7442 vs 0.7317). The average gains over ProS are only about +0.19 mAP on UCDR and +0.38 mAP on UdCDR, so run-to-run noise could easily reverse the headline. No error bars, seeds, or paired tests are reported. A secondary reproducibility issue compounds this: Eqs. (7)-(8) define Pd = V'wd and Pc = U'wc, which swaps the masked matrices relative to the attention weights, making the target-prompt computation dimensionally suspect as written. These issues do not prove the method is ineffective, but they mean the central claim is not currently supported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UCDR-Adapter, a method for universal cross-domain retrieval (UCDR) that adapts a frozen CLIP model by learning domain- and class-specific prompt matrices in a first phase and then generating target prompts for unseen domains/classes through soft attention over masked source prompts in a second phase. Training uses a triplet loss with momentum-updated queues and an image-text contrastive loss with a learnable textual semantic template. At test time, only the image branch and the generated prompts are used. Experiments are reported on DomainNet, Sketchy, and TU-Berlin for UCDR, UdCDR, and UcCDR tasks, comparing with SnMpNet, SCNNet, SASA, Zero-Shot CLIP, and ProS. The paper claims superior performance, with an ablation study in Table 5 and a parameter count of 2.36M trainable parameters in Table 6.","tokens_in":14322,"tokens_out":5461,"duration_ms":44155,"significance":"The proposed idea of synthesizing target prompts by attending to masked source prompts is a plausible and low-cost approach to handling unseen domains and classes, and the two-phase design with momentum queues is well motivated. If the empirical claims were robustly established, the method would be a useful parameter-efficient baseline for UCDR. However, the evidence presented is mixed: the method loses to ProS on several configurations, and no significance testing or variance analysis is given. The paper also contains an apparent typo in the core TPG equations (Eqs. 7-8) and does not describe an explicit adapter module despite the title. The conceptual contribution is still interesting, and the ablation table (Table 5) provides some support for the design choices, but the central performance claim is not fully supported in the current form.","major_comments":[{"comment":"The target prompt computation as written swaps the matrices. With wd = Attn(Ig, U') and wc = Attn(Ig, V'), the weighted combinations should be Pd = U' wd and Pc = V' wc. As written, Pd = V' wd and Pc = U' wc are dimensionally incompatible because U' has |Dtr| rows and V' has |Ctr| rows. Please correct the equations and clarify the intended dimensions of the attention weights and the generated prompts.","section":"Section 4.2, Eqs. (7)-(8)"},{"comment":"The claim of 'consistently outperforms ProS in most cases' is not supported by the reported numbers. ProS is better on Quickdraw and Infograph in Table 2 (e.g., mAP@200 0.2842 vs 0.2794 for Quickdraw and 0.5798 vs 0.5714 for Infograph), on Sketch in Table 3 (mAP@200 0.7385 vs 0.7332), and on TU-Berlin in Table 4 (mAP@All 0.6675 vs 0.6581). The average gains over ProS are only about +0.19 mAP points on the UCDR table and +0.38 mAP points on the UdCDR table, and no error bars, seeds, or paired tests are reported. Please either soften the claims to match the data or add repeated runs with statistical significance tests and discuss the cases where the method underperforms.","section":"Tables 2-4 and Abstract"},{"comment":"The paper's title and Section 4 promise an adapter-based method, but no adapter architecture is actually described. The text describes only learnable prompt vectors added to image features and a learnable text template; Table 6 reports 2.36M training parameters, which is plausible for prompt tuning, but the 'adapter' module itself (e.g., a bottleneck MLP inserted in the ViT) is never specified. Please state exactly which parameters constitute the adapters and how they are inserted and trained, or rename the method to avoid misleading readers.","section":"Sections 4.1-4.3"},{"comment":"There is a contradiction about which components are frozen. Section 4.3 says 'Frozen Image Encoder from the pre-trained CLIP model does not require any extra modifications,' but Section 4.1 states that prompts 'are optimized via backpropagation along with the vision transformer encoder parameters.' Please clarify which parameters are trainable in each phase and whether the image encoder is frozen or fine-tuned.","section":"Section 4.3 vs Section 4.1"},{"comment":"The text calls the pipeline a 'three-phase approach' (Source Prompt Learning, Target Prompt Generation, and test-time), while the abstract and the rest of the paper refer to a 'two-phase training strategy.' This inconsistency should be reconciled, for instance by describing the test-time procedure as a separate inference phase rather than a training phase.","section":"Section 4, first paragraph"}],"minor_comments":[{"comment":"The phrase 'consistently outperforms ProS in most cases' is self-contradictory; either report consistent wins or acknowledge that the method wins in most but not all cases.","section":"Abstract and Section 1"},{"comment":"The text uses 'ternary loss' instead of 'triplet loss' in the description of Eq. (2), and the sentence 'LT ripletloss on these samples' contains a spacing typo; please correct these.","section":"Section 4.1"},{"comment":"The UdCDR results section discusses gains over SASA but does not mention the comparison with ProS in the text; please report the relative differences to ProS and address the mixed results explicitly.","section":"Section 5.2"},{"comment":"The caption says 'Where Pd and Pc are target prompts generated for unseen domains and classes,' but the figure does not clearly label Pd and Pc; please add labels to the figure to match the text.","section":"Figure 4 caption"},{"comment":"The paper gives a project page but no code or configuration details; consider adding a reproducibility appendix with dataset splits, hyperparameter values, and the source of the ProS baseline numbers.","section":"Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The strongest baseline, ProS, is co-authored by one of the current authors (Zhi-Qi Cheng). While this is not inherently problematic, the paper should state how the ProS numbers were obtained (official code, re-implementation, or reported values) and ideally release code to allow verification. The central empirical claim would be strengthened by including standard deviations over multiple seeds. The paper is within the scope of the journal, though its significance is incremental relative to ProS."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid incremental paper with a real architectural combination, but the central claim is over-egged. The authors combine momentum-updated class/domain prompts, a learnable text template, triplet loss, and a masked-attention target prompt generator for universal cross-domain retrieval with frozen CLIP. Each ingredient is known, but putting them together for UCDR with only 2.36M trainable parameters is a legitimate new configuration, and the ablations in Table 5 show each component buys something. That is real work and worth acknowledging.\n\nWhat is new and good: the two-phase training is clean, the TPG idea of synthesizing target prompts by attending over masked source prompts is plausible, and the parameter-efficiency story is well made. The paper is clearly written enough to follow, and the ablation study is more honest than many in this space.\n\nNow the soft spots, in order of severity. First, the abstract and conclusion claim UCDR-Adapter 'consistently outperforms' ProS. The paper's own tables contradict that. In Table 2, ProS beats it on Quickdraw and Infograph for the unseen-class gallery; in Table 3, ProS is better on Sketch mAP@200; in Table 4, ProS wins on TU-Berlin for both metrics. The average gains over ProS are under 0.4 mAP, and there are no error bars, seeds, or significance tests, so the headline claim is not supported. This is a load-bearing flaw in the framing, not in the method itself.\n\nSecond, Eqs. (7)-(8) look like a typo that swaps which masked matrix gets which attention weights. The text says wd = Attn(Ig, U') and wc = Attn(Ig, V'), then defines Pd = V'wd and Pc = U'wc. As written, the dimensions and semantics don't line up; likely Pd = U'wd and Pc = V'wc. Probably a notation slip, but it makes the core TPG formula hard to trust without code.\n\nThird, the main baseline ProS is co-authored by one of this paper's authors, and the paper does not disclose that author overlap. Overlap with a baseline is not disqualifying, but reviewers and readers should know, especially when the reported advantage is thin.\n\nMinor point: calling the prompts 'adapters' is imprecise; they are prompts, not a bottleneck adapter in the usual sense.\n\nBottom line: the method deserves a serious referee and probably a place at a decent venue after revision. The engineering is sound, the ablations are honest, and the masked-attention prompt generation is worth testing. But the authors need to soften the 'consistent' claim, fix the equation, release code, and ideally add variance estimates. Conditional accept for me, not a reject, and not a strong accept either.","headline":"A reasonably built parameter-efficient prompt-tuning variant for UCDR, but the 'consistently outperforms' headline doesn't survive contact with its own tables.","tokens_in":14914,"tokens_out":1803,"would_cite":false,"duration_ms":15782,"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":"UCDR-Adapter claims that dynamic prompts, synthesized by attending over masked source prompts, let a frozen CLIP model retrieve across unseen domains and classes more accurately than static-prompt baselines.","keywords":["universal cross-domain retrieval","prompt tuning","vision-language models","CLIP","dynamic prompt generation","adapter modules","image retrieval","momentum contrast"],"falsifier":"The claim that UCDR-Adapter 'consistently outperforms ProS in most cases' is tested directly by the paper's Tables 2 and 4: on the Quickdraw holdout ProS achieves 0.2842 mAP@200 against 0.2794 for UCDR-Adapter, and on TU-Berlin ProS achieves 0.6675 mAP@All against 0.6581. An additional experiment that would localize the cause is to substitute uniform attention weights for the learned weights in Eqs. 7-8; if retrieval accuracy is unchanged, the attention weighting itself is not the source of the gains.","tokens_in":13771,"feed_emoji":"🔍","tokens_out":7397,"duration_ms":57562,"temperature":0.7,"pith_summary":"The paper introduces UCDR-Adapter, a two-phase method to adapt a frozen CLIP model for Universal Cross-Domain Retrieval (UCDR), where test queries can come from unseen domains and classes. Phase one learns class- and domain-specific prompts under momentum updates, triplet loss, and image-text contrastive loss; phase two generates dynamic target prompts for novel inputs by attending over masked source prompts. The central claim is that this dynamic prompt generation outperforms the static-prompt ProS baseline and earlier methods on the UCDR, UdCDR, and UcCDR benchmarks, while using only the image branch at inference and adding 2.36M trainable parameters. Success would make cross-domain retrieval practical in settings where new categories and visual domains appear after deployment, without needing text descriptions at query time.","feed_headline":"Dynamic prompt generation lifts CLIP retrieval on unseen data","feed_subtitle":"Two-phase masked-prompt attention beats static prompts on DomainNet, Sketchy, and TU-Berlin.","key_machinery":"The load-bearing object is the Target Prompt Generation (TPG) module. Given masked source prompts $U' = U \\circ (1-\\delta_d)$ and $V' = V \\circ (1-\\delta_c)$, it computes attention weights $w_d = \\text{Attn}(I_g, U')$ and $w_c = \\text{Attn}(I_g, V')$, then forms target prompts as weighted combinations of the masked rows (the paper's Eqs. 5-8). These are concatenated, projected to the encoder input dimension, and added to the image features of a frozen CLIP ViT-B/32 encoder. The same module serves all three retrieval subtasks without modification, and it is what distinguishes UCDR-Adapter from static-prompt methods like ProS.","core_discovery":"The paper claims that adapting a pre-trained vision-language model to unseen domains and classes in retrieval is better done with dynamic, image-conditioned prompts than with static prompts. The mechanism is a Target Prompt Generation (TPG) module that masks the rows of the learned domain prompt matrix U and class prompt matrix V, computes soft-attention weights between the input image features and the masked rows, and forms target prompts as the weighted combinations of those masked rows. These generated prompts are injected into the frozen image encoder, and the resulting features are aligned to text embeddings through the contrastive loss from phase one. At test time, the text encoder and all textual inputs are discarded, so retrieval uses only the image branch. The authors report consistent improvements over the static-prompt ProS method 'in most cases' and larger gains over SASA, with the full system reaching 0.6591 mAP@200 on the DomainNet UCDR Sketch holdout versus 0.6457 for ProS.","pith_inferences":["The largest ablation gain comes from adding the two-phase TPG schedule (+6.82% mAP over one-phase training), while isolating the masking itself gains only +0.66% (Table 5); this suggests the benefit may come more from the two-phase optimization than from the attention-weighted masking, a distinction the paper does not make.","A natural extension, untested in the paper, is whether the same mask-and-attend prompt synthesis transfers to very different visual distributions such as medical or satellite imagery; the current benchmarks are all natural-image domains.","Because the paper reports single-run results without variance or significance tests, the margins under one mAP point separating UCDR-Adapter from ProS on several configurations would need repeated-run evidence to support the claim of consistent superiority."],"forward_implications":["If the central claim is right, a frozen CLIP model with 2.36M trainable parameters can serve retrieval when both domain and class are unseen, eliminating test-time text input entirely.","The TPG masking strategy becomes a general recipe for synthesizing prompts for novel categories and domains from a fixed source-prompt bank, applicable beyond retrieval to classification or detection.","The reported UcCDR gains on Sketchy and TU-Berlin suggest the method transfers to fine-grained, sketch-based retrieval and would apply to other sketch-photo benchmarks.","The momentum queues and triplet loss are claimed to improve inter-class separation for unseen classes, implying the adapted features are more linearly separable than raw CLIP features."],"supporting_citations":[{"why":"Defines the UCDR task, the DomainNet/Sketchy/TU-Berlin splits, and the evaluation metrics used throughout.","marker":"[33]"},{"why":"The static-prompt baseline (ProS) that UCDR-Adapter is designed to beat and the main comparison in Tables 2-4.","marker":"[12]"},{"why":"Provides the frozen CLIP image and text encoders that the adapters and prompts augment.","marker":"[37]"},{"why":"Supplies the momentum encoder and per-class queue design used for hard sample mining in phase 1.","marker":"[16]"},{"why":"Provides the learnable textual template formulation that the paper adapts into its Learnable Textual Semantic Template.","marker":"[52]"},{"why":"A prior state-of-the-art UCDR method (SASA) used as a principal comparison and to motivate the gap in static prompt methods.","marker":"[41]"},{"why":"An earlier UCDR baseline (SCNNet) that helps calibrate the improvement over prior art in the benchmark tables.","marker":"[1]"}],"fun_headline_variants":["Dynamic prompts beat static for cross-domain retrieval","Image-only retrieval: dynamic prompts win","CLIP adapters with dynamic prompts top static on unseen domains","No text at test time: dynamic prompts improve retrieval","UCDR-Adapter: dynamic prompts lift cross-domain retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise that carries the method is that prompts useful for an unseen class or domain can be synthesized as a soft-attention-weighted combination of the masked rows of the source prompt matrices; if the masked rows do not span a space containing useful prompts for novel inputs, the Target Prompt Generation stage contributes nothing beyond the phase-one prompts.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic prompts beat static for cross-domain retrieval","Image-only retrieval: dynamic prompts win","CLIP adapters with dynamic prompts top static on unseen domains","No text at test time: dynamic prompts improve retrieval","UCDR-Adapter: dynamic prompts lift cross-domain retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1208,"prompt_tokens":952,"completion_tokens":256,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":181}},"tokens_in":568,"tokens_out":256,"duration_ms":2795,"temperature":1.0,"reasoning_tokens":181,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:43:24.536311+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The claim that UCDR-Adapter 'consistently outperforms ProS in most cases' is tested directly by the paper's Tables 2 and 4: on the Quickdraw holdout ProS achieves 0.2842 mAP@200 against 0.2794 for UCDR-Adapter, and on TU-Berlin ProS achieves 0.6675 mAP@All against 0.6581. An additional experiment that would localize the cause is to substitute uniform attention weights for the learned weights in Eqs. 7-8; if retrieval accuracy is unchanged, the attention weighting itself is not the source of the gains.","supporting_citations":[{"cited_title":"Universal cross-domain retrieval: Generalizing across classes and do- mains","cited_arxiv_id":null,"evidence_quote":"Defines the UCDR task, the DomainNet/Sketchy/TU-Berlin splits, and the evaluation metrics used throughout."},{"cited_title":"Pros: Prompting-to-simulate generalized knowledge for universal cross-domain retrieval","cited_arxiv_id":null,"evidence_quote":"The static-prompt baseline (ProS) that UCDR-Adapter is designed to beat and the main comparison in Tables 2-4."},{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Provides the frozen CLIP image and text encoders that the adapters and prompts augment."},{"cited_title":"Structure-aware semantic-aligned network for universal cross-domain retrieval","cited_arxiv_id":null,"evidence_quote":"A prior state-of-the-art UCDR method (SASA) used as a principal comparison and to motivate the gap in static prompt methods."},{"cited_title":"Contrastive learning of semantic concepts for open-set cross-domain retrieval","cited_arxiv_id":null,"evidence_quote":"An earlier UCDR baseline (SCNNet) that helps calibrate the improvement over prior art in the benchmark tables."}],"review_version":1}