{"id":"09fe6765-2730-4ed6-88e8-3450c24af651","arxiv_id":"2412.19417","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"KALAHash uses CLIP text embeddings of category names to anchor low-rank fine-tuning and hash-code optimization, improving low-shot image retrieval.","lead":"This paper introduces KALAHash, a method that adapts image-hashing models to new categories with as few as one labeled example per class, using text embeddings of category names as anchors. The authors report large retrieval gains over existing hashing baselines and claim a fourfold reduction in required training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The KIDDO loss in Eq. (9) is dimensionally inconsistent and uses one-hot labels for multi-label data, so the described method and the derived DCC update in Eq. (12) do not follow as written.","rationale":"The reader's conditional verdict already accounts for the multi-label label-matrix issue and reproducibility concerns. My stress-test reinforces that with a sharper, checkable defect: the central optimization objective is not well-formed as typeset. This matters more than the text-vision alignment prior because even if CLIP anchors are perfectly aligned, a reader cannot implement KIDDO from Eqs. (9)-(12) as written. It is also independent of unresolved empirical variance (no error bars) and could be settled from the repository in minutes. I do not move the verdict to reject because the code link and self-reported tables may well contain the corrected objective; if so the issue is presentational, though it must be fixed. The reader's emphasis on text-visual alignment partially overlaps with my concern because Eq. (9) is where the label-matrix assumption enters, but my concern is a formal inconsistency rather than a domain-alignment assumption. Therefore UNCHANGED (conditional) is appropriate pending the code check.","tokens_in":14326,"tokens_out":7673,"duration_ms":70278,"concrete_test":"Inspect the released code at https://github.com/Tree-Shu-Zhao/KALAHash.pytorch: print the shapes of Y, T, and B at the KIDDO loss and check whether the code computes Y - T.T@B, Y - T@B.T, or another term, and whether Y uses one-hot or multi-hot labels on NUS-WIDE/MS-COCO. Independently re-derive the DCC update in Eq. (12) from the implemented objective; if the update differs from the paper's formula or if the code's loss is undefined, re-run the 1-shot NUS-WIDE and CIFAR-10 rows of Table 1 with the code as written to see whether the reported mAP values are reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the KIDDO section, Eq. (9) defines L_a = ||Y - T^T B||^2 with Y in R^{C x n}, B in { -1, 1 }^{n x b}, T = G(K), and K in R^{C x d_t}. If G outputs b-dimensional codes, then T is C x b, so T^T is b x C and T^T B has mismatched inner dimensions (C vs n) except in the coincidental case C = n. If T were meant to have n rows, G cannot map from a C-row class-knowledge tensor because n is not known at text-encoding time. The loss is therefore undefined as written. The later DCC update in Eq. (12) is asserted to solve Eq. (11), but since Eq. (11) inherits this dimension mismatch, its derivation cannot be verified and the formula may contain a transpose error (for example T B^T instead of T^T B). Separately, the paper states that Y is a matrix of one-hot label vectors, while NUS-WIDE and MS-COCO are multi-label; a one-hot Y makes KIDDO unable to represent co-occurring labels, so the reported multi-label gains do not follow from the described objective. This is load-bearing because KIDDO is one of two central contributions and the ablations in Table 3 attribute 2.38-11.16 mAP points to it; if the written objective is not what was optimized, the central empirical claim rests on an unspecified method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces KALAHash, a method for low-resource (1-8 shots per class) adaptation of deep hashing models. It builds on CLIP: CLoRA constructs low-rank adaptation matrices for the vision transformer using top-r class text embeddings as anchors, and KIDDO adds a discrete optimization objective that aligns binary codes with text embeddings. Experiments on CIFAR-10, NUS-WIDE, and MS-COCO report consistent mAP improvements over eight hashing baselines, a plug-and-play study adding CLoRA to baselines, ablations, scaling analyses, and inference-time measurements. The central claims are that KALAHash significantly boosts retrieval performance and achieves 4x data efficiency.","tokens_in":14697,"tokens_out":6536,"duration_ms":52823,"significance":"If the empirical claims were supported, the paper would make a useful contribution: it is among the first to address low-resource adaptation specifically for deep hashing, and the idea of using class-level text embeddings both as a parameter-efficient adaptation mechanism (CLoRA) and as optimization anchors (KIDDO) is practical and reasonably novel. The authors provide a public code link and extensive experiments across three datasets, several backbones, and multiple shot settings, which is a strength. However, the current manuscript contains serious inconsistencies in the formulation of KIDDO, inconsistent baseline numbers across tables, missing statistical validation, and an unsubstantiated 4x data-efficiency claim, so the central claims are not yet supported as written.","major_comments":[{"comment":"The alignment loss La = ||Y - T^T B||^2 is dimensionally inconsistent. With Y in R^{C x n}, B in {±1}^{n x b}, and T = G(K) with K in R^{C x d_t}, if G outputs b-dimensional rows then T is C x b, so T^T B is (b x C)(n x b), which is not conformable unless C = n. The expression T B^T, giving C x n, would be conformable, suggesting a missing transpose. Since Eq. (11) and the DCC update in Eq. (12) inherit this mismatch, the derivation cannot be verified, and the formula for S and the update Bi = sign(S_i - B'^T T'_i) contain further dimensional inconsistencies. This is load-bearing because KIDDO is a central contribution and the ablation in Table 3 attributes 2.38-11.16 mAP points to it. Please correct the objective, re-derive the DCC update, and verify the equations.","section":"Knowledge-Guided Discrete Optimization, Eq. (9)"},{"comment":"Y is defined as a matrix of one-hot label vectors, yet the experiments on NUS-WIDE and MS-COCO treat these as multi-label datasets, and the evaluation protocol states that two images are similar if they share at least one label. With a one-hot Y, the alignment loss cannot represent co-occurring labels, so the reported multi-label gains do not follow from the described objective. Please specify whether Y is multi-hot for these datasets and ensure the equations and implementation match this choice.","section":"Knowledge-Guided Discrete Optimization, Eq. (9)"},{"comment":"The baseline numbers are inconsistent between the two tables. For MS-COCO 1-shot, HSWD is 60.15 in Table 1 but 58.55 in Table 2; for CIFAR-10 1-shot, OrthoHash is 46.68 in Table 1 but 49.50 in Table 2. This inconsistency undermines the plug-and-play gains in Table 2, since the improvements may be computed against different baseline values. Please unify the tables and re-report the results consistently.","section":"Tables 1 and 2"},{"comment":"None of the reported results include error bars, standard deviations, or multiple random splits. In 1-shot to 8-shot settings, the training set selection is stochastic and variance is typically high; without repeated runs, the claim that KALAHash \"significantly boosts\" retrieval performance is not statistically supported. Please report mean ± std across at least 3 seeds (or provide confidence intervals) for the main tables and for the plug-and-play results.","section":"Experiments (overall)"},{"comment":"The claim of \"4x data efficiency\" is not quantified anywhere in the manuscript. There is no experiment showing that KALAHash at k shots matches or exceeds baselines at 4k shots, and no explicit definition of the 4x factor. Please add a clear analysis, e.g., using Figure 4 or a dedicated table, to substantiate this claim, or remove it from the abstract and introduction.","section":"Abstract and Experiments"},{"comment":"All baselines are trained without the text embeddings (CLIP class-name knowledge) that KALAHash uses in both CLoRA and KIDDO. The statement that \"all methods, including baselines, use the same backbone model, optimizer, training hyperparameters, etc.\" is therefore incomplete as a fairness argument, because the proposed method has access to additional semantic side information. Since the method's gains could stem primarily from this extra prior, please add a baseline that also uses text knowledge (e.g., a CLIP zero-shot hash adapted variant, or a variant with text features appended) or explicitly discuss this limitation.","section":"Implementation Details and Comparison Protocol"}],"minor_comments":[{"comment":"The text contains typos such as \"konwledge\" for \"knowledge\"; please proofread the manuscript.","section":"Class-Level Textual Knowledge Generation"},{"comment":"In the similarity loss Ls, it is not explicitly stated whether H or B is used in θij; please clarify that H is the real-valued hash output before binarization.","section":"Eq. (8)"},{"comment":"The definitions of S, S', T', and B' are ambiguous and contain possible notation errors (e.g., \"S = βYT + γH\" appears to lack a transpose and is dimensionally unclear). The formulas need to be rewritten consistently, even after correcting Eq. (9).","section":"Eq. (12)"},{"comment":"The caption says \"mean Average Precision scores\" but should use \"mean Average Precision (mAP)\" for consistency with the tables.","section":"Figure 1"},{"comment":"The text states \"FFT achieves a Silhouette Score of 50.0%\" without noting that the raw score is normalized from [-1,1] to [0,100]; please add that clarification in the main text as well as the caption.","section":"Introduction"},{"comment":"The notation \"w.o.CLoRA\" and \"w.CLoRA\" is unusual; please spell out \"without\" and \"with\" in the table caption or use standard abbreviations.","section":"Table 5"},{"comment":"The sentence \"As the number of training samples increases, our approach consistently improves the retrieval performance\" should read \"As the number of bits increases\" based on the figure and context.","section":"Appendix: Scalability of the Number of Bits"},{"comment":"Tables 8 and 9 are not explicitly referenced in the main text; please add cross-references or integrate this content into the main experimental section.","section":"Appendix Tables 8 and 9"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible core idea and the public code link is a positive sign, but the current version has several load-bearing inconsistencies that prevent verification of the central claims. The KIDDO objective as written is dimensionally invalid, the multi-label use of one-hot labels is unexplained, baseline numbers differ across tables, no statistical significance is reported, and the 4x data-efficiency claim is unquantified. These issues are likely fixable with additional experiments and corrections, so I recommend major revision rather than rejection. I would ask the authors to check their code carefully to ensure the described equations match the actual implementation, and to add error bars and multi-seed results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhao et al. attack a real gap: adapting deep hashing models with 1-8 examples per class. The CLoRA module, which replaces the learned left LoRA factor with CLIP class-text embeddings, is a non-obvious twist on LoRA, and the plug-and-play results in Table 2 are consistent across baselines. The benchmark itself—8 baselines, 3 datasets, 4 shot levels—is a useful contribution. Credit where due: they ship code, and the ablations and sensitivity analyses are reasonably thorough.\n\nThe soft spots are serious, though. Equation (9) defines La = ||Y - T^T B||^2 with T ∈ R^{C×b}, B ∈ R^{n×b}, and Y ∈ R^{C×n}. Then T^T is b×C, so T^T B is undefined unless C=n. The natural alignment loss would be ||Y - T B^T||^2 or ||Y - B T^T||^2, but as written the objective doesn't parse. The DCC update in Eq. (12) inherits this, so the derivation can't be verified. On top of that, Y is defined as one-hot label vectors, yet NUS-WIDE and MS-COCO are multi-label; a one-hot Y cannot represent co-occurring labels, so the reported multi-label gains don't follow from the described objective. This is load-bearing because the ablations credit KIDDO with 2.38-11.16 mAP points. The authors need to state exactly what objective they optimized and confirm the code matches the corrected equation.\n\nWeaker issues: no error bars anywhere, an inconsistent HSWD baseline between Tables 1 and 2, and an unquantified '4x data efficiency' claim. These are fixable but shouldn't be ignored.\n\nSo what's the bottom line? The CLoRA idea is worth serious consideration and the empirical trend is probably real, but the KIDDO section is mathematically unclear as written. I'd send it to reviewers with a request for major revision, not desk-reject it. If you work in low-resource retrieval, read the CLoRA part and wait for the corrected KIDDO version before relying on those numbers.","headline":"Promising CLoRA idea and a broad empirical study, but the KIDDO loss as written is dimensionally undefined, which undercuts the paper's central contribution.","tokens_in":15203,"tokens_out":5095,"would_cite":false,"duration_ms":38957,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Class-name text embeddings serve as anchors that let deep hashing adapt from as few as one image per class.","keywords":["deep hashing","low-resource adaptation","parameter-efficient fine-tuning","LoRA","vision-language models","CLIP embeddings","few-shot retrieval","discrete optimization"],"falsifier":"Take a 1-shot retrieval benchmark from a domain CLIP was not trained on, such as medical or satellite imagery, and compare KALAHash against plain LoRA on the same frozen CLIP backbone; if mAP does not improve, the text-anchor assumption fails. A cleaner controlled test: permute the category-name text embeddings randomly before training, and if retrieval mAP stays high after permuting, the semantic content of the anchors is not what carries the result.","tokens_in":14141,"feed_emoji":"🔎","tokens_out":6967,"duration_ms":57173,"temperature":0.7,"pith_summary":"The paper takes on a setting most deep-hashing methods ignore: adapting a pre-trained model to a new retrieval task when only one to eight labeled images per class are available. It claims that in this regime, existing methods degrade badly because the few samples shift the learned hash-code distribution away from what the pre-trained model knows. To fix this, KALAHash treats the CLIP text embedding of each category name as an anchor point, injecting that knowledge into both the fine-tuning adapter and the hash-code optimization. Concretely, CLoRA builds LoRA's low-rank update from class-level text embeddings, and KIDDO adds an alignment loss that pushes binary codes to reproduce the one-hot label matrix through the text knowledge. The paper reports that the combination beats eight deep-hashing baselines on CIFAR-10, NUS-WIDE, and MS-COCO at 1 to 8 shots and claims a 4x data efficiency.","feed_headline":"Text anchors let deep hashing learn from a single image per class","feed_subtitle":"Injecting CLIP class-name embeddings into LoRA and hash optimization beats eight baselines at 1-8 shots.","key_machinery":"The carrying mechanism is a pair of knowledge-injection devices built on the CLIP text encoder. CLoRA constrains a LoRA weight update to the span of class text embeddings: for each input, it averages the vision tokens, picks the top text embeddings by cosine similarity, and forms $\\Delta W = \\eta \\sum_i \\hat{k}_i q_i^T$, so the adapter can only move weights along semantic directions the category names define. KIDDO adds an alignment loss $\\|Y - T^\\top B\\|_2^2$ to the standard pairwise similarity and quantization losses and solves for binary codes $B$ column by column with discrete cyclic coordinate descent, injecting the same textual knowledge into the binary codes themselves.","core_discovery":"On the paper's own terms, the central discovery is that class-level textual knowledge from a pre-trained vision-language model can stand in for the missing visual variety in low-resource deep hashing. The hash function is anchored by category-name embeddings: CLoRA replaces the random left factor of a LoRA update with a text embedding selected by cosine similarity to the pooled visual tokens, and KIDDO aligns the learned binary codes with the same text knowledge while optimizing them with discrete cyclic coordinate descent. The paper's experiments report consistent gains over eight baselines at 1, 2, 4, and 8 shots, with the largest margins in the 1-shot case, and it shows CLoRA can be dropped onto existing hashing losses to improve them.","pith_inferences":["Beyond the paper: the anchor construction should transfer to any retrieval head built on CLIP features, not just hash layers, because it only changes how adapter weights are formed.","Beyond the paper: if anchor selection is the true source of gain, replacing single category-name prompts with richer per-class descriptions should help further on fine-grained benchmarks, which the paper does not test.","Beyond the paper: the top-1 cosine anchor selection in Equation (6) could misfire for visually similar classes, so a soft mixture of anchors or a learned selection rule is a natural stress test of that assumption.","Beyond the paper: the 4x data-efficiency claim suggests a direct testable rule, namely that KALAHash at one shot should match the retrieval mAP of existing methods at roughly four shots, which can be checked against the scaling curves in Figure 4."],"forward_implications":["At the 1-shot and 2-shot settings, KALAHash reports mAP gains over all eight baselines on CIFAR-10, NUS-WIDE, and MS-COCO, with the largest margins at 1 shot.","The plug-and-play CLoRA component improves mAP for every baseline it is inserted into, with gains up to 12.52 mAP points on CIFAR-10.","Ablation results indicate both components matter: removing CLoRA costs 2.38 to 11.16 mAP points, and removing KIDDO costs 3.72 to 6.65 mAP points across the three datasets.","When the number of training samples grows to 500 shots on CIFAR-10, KALAHash remains competitive with state-of-the-art methods, suggesting the textual anchor does not harm performance once data is plentiful.","CLoRA adds negligible inference overhead across tested backbones, with measured increases of 0.01 to 0.05 milliseconds per image."],"supporting_citations":[{"why":"Supplies the pre-trained CLIP text and vision encoders whose class-level text embeddings are the knowledge anchors for both CLoRA and KIDDO.","marker":"Radford et al. (2021)"},{"why":"Defines the LoRA low-rank adaptation formulation that CLoRA modifies by replacing one low-rank factor with class text embeddings.","marker":"Hu et al. (2022)"},{"why":"Provides the discrete cyclic coordinate descent method used to optimize the binary hash codes under the KIDDO objective.","marker":"Shen et al. (2015)"},{"why":"Provides the OrthoHash baseline, the evaluation protocol, and the hash layer design that the paper follows.","marker":"Hoe et al. (2021)"},{"why":"Provides the HashNet baseline and the mAP evaluation convention used in the low-resource comparisons.","marker":"Cao et al. (2017)"},{"why":"Supplies the NUS-WIDE dataset used for multi-label retrieval evaluation.","marker":"Chua et al. (2009)"},{"why":"Supplies the MS-COCO dataset used for multi-label retrieval evaluation.","marker":"Lin et al. (2014)"},{"why":"Supplies the CIFAR-10 dataset used for single-label retrieval evaluation.","marker":"Krizhevsky and Hinton (2009)"}],"fun_headline_variants":["Text anchors let deep hashing learn from 1 image per class","CLoRA: text knowledge anchors low-shot deep hashing","KALAHash: 4x data efficiency with class text anchors","Class-name embeddings rescue low-resource hashing","Vision-language anchors yield 4x data efficiency in hashing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the CLIP text embedding of a category name is a faithful anchor for that category's visual hash codes, so if text and vision embeddings do not line up for a target domain, both CLoRA's anchor selection and KIDDO's alignment loss lose their grounding.","fun_headline_variants_meta":{"raw":{"variants":["Text anchors let deep hashing learn from 1 image per class","CLoRA: text knowledge anchors low-shot deep hashing","KALAHash: 4x data efficiency with class text anchors","Class-name embeddings rescue low-resource hashing","Vision-language anchors yield 4x data efficiency in hashing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000618,"raw_usage":{"total_tokens":2849,"prompt_tokens":910,"completion_tokens":1939,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":1854}},"tokens_in":526,"tokens_out":1939,"duration_ms":11564,"temperature":1.0,"reasoning_tokens":1854,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:37:23.988891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a 1-shot retrieval benchmark from a domain CLIP was not trained on, such as medical or satellite imagery, and compare KALAHash against plain LoRA on the same frozen CLIP backbone; if mAP does not improve, the text-anchor assumption fails. A cleaner controlled test: permute the category-name text embeddings randomly before training, and if retrieval mAP stays high after permuting, the semantic content of the anchors is not what carries the result.","supporting_citations":[{"cited_title":"W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained CLIP text and vision encoders whose class-level text embeddings are the knowledge anchors for both CLoRA and KIDDO."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the discrete cyclic coordinate descent method used to optimize the binary hash codes under the KIDDO objective."},{"cited_title":"T.; Ng, K","cited_arxiv_id":null,"evidence_quote":"Provides the OrthoHash baseline, the evaluation protocol, and the hash layer design that the paper follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the HashNet baseline and the mAP evaluation convention used in the low-resource comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NUS-WIDE dataset used for multi-label retrieval evaluation."},{"cited_title":"J.; Hays, J.; Perona, P.; Ramanan, D.; Doll \\' a r, P.; and Zitnick, C","cited_arxiv_id":null,"evidence_quote":"Supplies the MS-COCO dataset used for multi-label retrieval evaluation."}],"review_version":1}