{"id":"054f39e8-3d62-44b2-9e6b-1ebca03d430b","arxiv_id":"2412.06284","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"UASA, a prototype-based network with adaptive class thresholds and uncertainty-aware clustering, outperforms prior methods on class-imbalanced cross-domain out-of-distribution detection benchmarks.","lead":"This paper defines a harder version of out-of-distribution detection where training and test data come from different visual domains and classes have unequal sample sizes. It proposes UASA, a network combining prototype-based domain alignment, per-class adaptive thresholds, and uncertainty-aware clustering, and reports large gains over prior methods on three benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (5) as written yields nonpositive thresholds for every target sample, so the published decision rule cannot explain the reported HOS scores; this must be corrected or confirmed with code.","rationale":"The central claim is an empirical SOTA result, so the load-bearing condition is that the proposed modules are actually the ones producing the reported numbers. The adaptive threshold module is the component that separates ID from OOD; if Eq. (5) is taken literally, that module rejects every sample, making the mechanism unexecutable. This is stronger than the reader's entropy-ordering concern: the ordering assumption is a plausible empirical risk that might fail on some target domains, whereas Eq. (5) is a mathematical obstruction that cannot be true for any reasonable entropy values. The reported ablations (Tables 5-9) give some independent support for the module contributions, and the class-balanced comparison (Table 4) is consistent with a working method, but these tables do not reveal the actual threshold formula. Because the issue is fixable by correcting a formula and releasing code, I would keep the CONDITIONAL verdict rather than reject; the revision conditions must include a corrected Eq. (5), consistent sigma values, and a working code link.","tokens_in":21442,"tokens_out":6944,"duration_ms":76352,"concrete_test":"Obtain the released code (or reimplement Section 3.3 exactly) and instrument the first validation mini-batch of Office-Home RE->PR: compute o_i from Eq. (5) and Q(p^t_i) from Eq. (6), then count samples classified as ID. If the count is zero while Table 2 reports HOS 67.48 for that task, the published equation is not the deployed rule; ask the authors for the corrected threshold formula and rerun all three benchmarks with it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.3, Eq. (5) defines o_i = alpha * (T_i - min(T) - max(T)) / (max(T) - min(T)) * log K_s. Because T_i is a mean of entropy values, 0 <= min(T) <= T_i <= max(T); hence T_i - min(T) - max(T) <= -min(T) <= 0. With alpha > 0 and log K_s > 0, every threshold o_i is nonpositive. Eq. (6) then labels x^t_i as ID only if Q(p^t_i) <= o_{argmax_j p^t_ij}. Since sample entropy Q(.) is nonnegative, only exactly zero-entropy samples can pass; under a standard softmax this is measure-zero. If Eq. (5) were actually deployed on Office-Home (K_s = 50), almost all target samples would be classified OOD and HOS would collapse, not reach the reported mean 61.14. Thus the published decision rule is internally inconsistent with the empirical claim. The likely explanation is a typo, but the manuscript does not give the corrected rule and the code link is not provided; moreover, Table 6 identifies sigma = 1.0 as the full model while Section 4 states sigma = 0.05. Without the actual threshold computation, UASA's central OOD mechanism cannot be verified from the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper defines a new task, class-imbalanced cross-domain OOD detection (CCOD), in which a labeled source domain and an unlabeled target domain with possibly different class distributions and unknown classes are used for OOD detection. It proposes the UASA network with four modules: label-driven prototype building, prototype-guided domain alignment, adaptive threshold generation, and uncertainty-aware target clustering. The paper reports large improvements over prior open-set domain adaptation and OOD detection methods on DomainNet, Office-Home, and VisDA-C.","tokens_in":21705,"tokens_out":6179,"duration_ms":60222,"significance":"The proposed CCOD setting is a genuinely useful extension of open-set domain adaptation, and the reported gains (e.g., 70.04 vs. 66.41 mean HOS on DomainNet) are substantial. The paper also provides ablations for each module and hyperparameter studies. However, the published decision rule in Eq. (5) appears to be internally inconsistent with the reported performance, and the sigma inconsistency between Section 4 and Table 6 prevents reproduction. These are local and correctable issues rather than conceptual impossibilities, so the contribution is potentially significant if the equations and implementation details are corrected.","major_comments":[{"comment":"The threshold formula in Eq. (5) is internally inconsistent with the reported results. Since 0 <= min(T) <= T_i <= max(T), the numerator T_i - min(T) - max(T) is <= -min(T) <= 0, and with alpha > 0 and log K_s > 0 every threshold o_i is nonpositive. Eq. (6) then labels a sample as ID only if Q(p_i) <= o_i, which for nonnegative entropy is possible only for exactly zero entropy; under a standard softmax this is measure-zero. Thus the published rule would mark almost all target samples as OOD and cannot yield the reported HOS values. Please correct the formula, give the exact deployed decision rule, and provide code or pseudocode for the threshold computation.","section":"Section 3.3, Eq. (5)"},{"comment":"The implementation details state sigma = 0.05 in Eq. (2), but the ablation in Table 6 reports the full model ('Ours') with results identical to the sigma = 1.0 row and identifies sigma = 1.0 as best. This ambiguity affects all reported results because sigma controls the softmax temperature in Eq. (2) and Eq. (3). Please state explicitly which sigma value was used for Tables 1-4 and the ablations, and why Table 6 uses sigma = 1.0.","section":"Section 4 vs. Table 6"},{"comment":"The entire ID/OOD decision rule rests on the assumption that 'ID samples always have lower entropy than OOD samples,' but the paper never validates this ordering on the target domains. If hard OOD samples receive overconfident low-entropy predictions, Eq. (5)-(6) cannot separate them from ID samples. Please add an empirical analysis of the entropy distributions for ID vs. OOD target samples (e.g., histograms or the AUROC of entropy as a detector) and discuss the failure mode. Also clarify how the pseudo-labels in H_i are initialized and updated, since Eq. (4)-(6) form a coupled system.","section":"Section 3.3, entropy-ordering assumption"},{"comment":"In Eq. (2), the denominator sums over j = 1 to K_t, but the source classifier has only K_s prototypes and source labels range over {1,...,K_s}. Unless K_t is a typo for K_s, or the sum is over the full label set of size K_s + K_t with a different notation, the source classification probability in Eq. (1) is not well-defined. Please fix the indexing and define K_t before its first use.","section":"Eq. (2)"}],"minor_comments":[{"comment":"The text says 'Our codes are available in Github' but no URL is given; please provide a working link or a supplement, as the central method cannot otherwise be verified.","section":"Section 4, code availability"},{"comment":"For DomainNet and Office-Home, no explicit class-imbalance statistics are reported; please state the class-size distribution or the imbalance factor used, since the CCOD setting is defined by class imbalance.","section":"Section 4, datasets"},{"comment":"Table 8 appears to duplicate the row label 'A = 2.4K_s' twice; one of the rows is presumably a different cluster count and should be relabeled.","section":"Table 8"},{"comment":"The claim that UASA 'outperforms compared methods by 7.07%' in the RE->PR task should state the specific baseline and table entry, since the sentence is ambiguous as written.","section":"Section 4.2"},{"comment":"The header formatting in Table 3 is difficult to read; please use separate columns for the imbalance factor mu and the source class count K_s.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea and reported benchmark gains are interesting, but I could not verify the method from the manuscript alone: the sign inconsistency in Eq. (5), the sigma discrepancy between Section 4 and Table 6, and the undefined K_t in Eq. (2) are exactly the kind of load-bearing details that must be corrected or confirmed with code before acceptance. I recommend major revision rather than rejection because these issues appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper identifies a real gap—cross-domain OOD detection with class-imbalanced target data—and the reported gains are consistent and large. But the central OOD threshold rule, Eq. (5), cannot produce those gains as written. Since T_i is a class-wise mean entropy inside [min(T), max(T)], the numerator T_i − min(T) − max(T) is always ≤ 0; with α>0 and log K_s>0 every threshold o_i is nonpositive. Eq. (6) then only accepts exactly-zero-entropy samples as ID, so on any standard softmax the model would tag essentially everything OOD and HOS would collapse. That has to be a typo, but the corrected formula is not given and the link to code is missing.\n\nWhat the paper does well: the CCOD setting is new and worth studying—earlier work handles domain shift, class imbalance, or open sets separately, and UASA is a reasonable combination of prototype alignment, per-sample adaptive thresholds, and uncertainty-weighted clustering. The ablations show each module contributes, and the gains over DANCE/Ovanet are large and fairly consistent across tasks and datasets. The entropy-order assumption (ID lower entropy than OOD) is stated explicitly, though not validated on the target domains; that is a real risk but not fatal.\n\nOther soft spots, in decreasing severity: the implementation text says σ=0.05 while Table 6 identifies σ=1.0 as the best setting and labels it \"Ours\"; the class-imbalance construction for DomainNet and Office-Home is not described (only VisDA-C gives the imbalance factor μ); there are no error bars or significance tests; and the code is mentioned but not linked. These are all fixable.\n\nBottom line: this deserves referee time rather than desk rejection, because the setting is valuable and the results might be real once the threshold typo is fixed. As-is, though, the central mechanism cannot be verified. I would send it back with a strong request for the corrected formula, the code, and a precise protocol for the imbalanced benchmarks. I would not cite it until then.","headline":"The CCOD setting is a real gap and the reported gains are plausible, but Eq. (5) as written makes the OOD threshold nonpositive for every sample, so the published method cannot produce the reported HOS scores.","tokens_in":22293,"tokens_out":6756,"would_cite":false,"duration_ms":64395,"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":"This paper introduces a realistic OOD detection setting that combines semantic, domain, and class-imbalance gaps, and proposes a prototype-based network that outperforms prior methods on three benchmarks.","keywords":["out-of-distribution detection","cross-domain detection","class-imbalanced data","prototype alignment","adaptive threshold","uncertainty-aware clustering","HOS score"],"falsifier":"Measure the entropy distributions of ID and OOD samples in a target domain after training UASA, for example the sketch domain on DomainNet; if a substantial fraction of OOD samples have entropy below the threshold assigned by Equation (5), then Equation (6) misclassifies them as ID and the reported HOS advantage would disappear on that task.","tokens_in":21218,"feed_emoji":"🎯","tokens_out":6551,"duration_ms":55466,"temperature":0.7,"pith_summary":"This paper argues that real-world out-of-distribution (OOD) detection must handle three gaps at once: the semantic gap between known and unknown classes, the domain gap between training and test data, and the class-imbalance gap between categories with different sizes. It introduces the class-imbalanced cross-domain OOD detection (CCOD) setting, in which a labeled source domain trains a detector and an unlabeled target domain provides both in-distribution and unknown classes with arbitrary class sizes. To solve CCOD, the paper proposes the UASA network, which builds label-driven prototypes from the source classifier, aligns target features to those prototypes, applies sample-wise adaptive entropy thresholds for OOD decisions, and clusters target samples with uncertainty-derived weights. On three benchmarks, UASA reports higher mean HOS scores than previous state-of-the-art methods, including 70.04 versus 66.41 on DomainNet.","feed_headline":"New network beats OOD detectors on three cross-domain benchmarks","feed_subtitle":"UASA handles unknown classes, domain shifts, and class imbalance together, lifting mean HOS to 70.04 on DomainNet.","key_machinery":"The key machinery is the label-driven prototype: each column of the source classifier's weight matrix serves as a class prototype, giving a bijective label-to-prototype mapping without extra parameters. These prototypes drive three mechanisms: prototype-guided domain alignment, which minimizes the entropy of target-feature similarities to prototypes and memory-bank neighbors; adaptive threshold generation, which computes a class-wise entropy baseline and rescales it into a per-sample OOD threshold; and uncertainty-aware target clustering, which weights pairwise KL-divergence losses by confidence scores so that semantically similar target samples form tight clusters. Together they target the semantic gap, the domain gap, and the class-imbalance gap respectively.","core_discovery":"UASA's central claim is that class-imbalanced cross-domain OOD detection is tractable with a prototype-based alignment network that avoids fixed decision thresholds. The network treats the columns of the source classifier weight matrix as label-driven prototypes, stores L2-normalized target features in a memory bank, and aligns target samples to prototypes and their neighbors by minimizing the entropy of the similarity distribution. For each target sample it generates an adaptive threshold from class-wise entropy statistics, marking a sample as OOD when its own entropy exceeds that threshold. Uncertainty-aware clustering then pulls together target samples that share pseudo-labels, weighting each pair by confidence. The paper reports mean HOS scores of 70.04 on DomainNet, 61.14 on Office-Home, and 52.39 on VisDA-C, exceeding the best prior methods by 3.63, 3.05, and 3.74 points respectively.","pith_inferences":["A direct test of the entropy assumption would be to train UASA on a target domain where OOD classes are designed to be confusable with ID classes and check the fraction of OOD samples below their sample-wise thresholds.","Because the prototypes are built from classifier weights, a source-free variant could ship only the weight matrix and adapt to a new target domain without any source images, which the paper suggests but does not evaluate.","The uncertainty-weighted clustering loss could be transferred to open-set domain adaptation tasks beyond CCOD, such as open-set semantic segmentation.","If the entropy assumption degrades under label shift, an alternative could calibrate thresholds using a small validation set of target ID samples."],"forward_implications":["CCOD becomes a realistic benchmark for OOD detection in multi-domain, class-imbalanced applications such as autonomous driving and medical imaging.","Sample-wise adaptive thresholds offer a general replacement for fixed entropy thresholds in other OOD detectors.","Prototype-based alignment can be carried out with only the classifier weights, so the source data need not be revisited during target adaptation.","The method's success on DomainNet, where many classes are visually close, suggests adaptive thresholds help most in fine-grained OOD settings."],"supporting_citations":[{"why":"Supplies the MSP baseline and the entropy-as-OOD-score premise that the adaptive threshold extends.","marker":"(Hendrycks and Gimpel, 2017)"},{"why":"Provides the ODIN baseline using temperature-scaled softmax scores for OOD detection.","marker":"(Liang et al., 2018)"},{"why":"Provides the Energy baseline and energy-based OOD scoring.","marker":"(Liu et al., 2020)"},{"why":"Provides the CIDER baseline using hyperspherical embeddings for OOD detection.","marker":"(Ming et al., 2023)"},{"why":"Supplies the DANCE baseline for universal domain adaptation with self-supervision.","marker":"(Saito et al., 2020)"},{"why":"Provides the Ovanet baseline, the strongest prior method on DomainNet and Office-Home.","marker":"(Saito and Saenko, 2021b)"},{"why":"Defines the HOS score used to measure joint ID accuracy and OOD recall.","marker":"(Bucci et al., 2020)"},{"why":"Supplies the DomainNet dataset used for the largest benchmark comparison.","marker":"(Peng et al., 2019)"},{"why":"Supplies the Office-Home dataset for cross-domain evaluation.","marker":"(Venkateswara et al., 2017a)"},{"why":"Supplies the VisDA-C dataset, used in its class-imbalanced version.","marker":"(Peng et al., 2018)"}],"fun_headline_variants":["UASA tops OOD detection on three hard benchmarks","Prototype alignment network beats OOD baselines","Cross-domain OOD: adaptive thresholds beat fixed ones","UASA lifts mean HOS to 70 on DomainNet","New network handles domain gaps and imbalanced classes for OOD"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The OOD decision rule rests on the assumption that in-distribution samples always have lower prediction entropy than out-of-distribution samples in the target domain; if some unknown-class samples receive overconfident low-entropy predictions, the adaptive threshold will label them as in-distribution.","fun_headline_variants_meta":{"raw":{"variants":["UASA tops OOD detection on three hard benchmarks","Prototype alignment network beats OOD baselines","Cross-domain OOD: adaptive thresholds beat fixed ones","UASA lifts mean HOS to 70 on DomainNet","New network handles domain gaps and imbalanced classes for OOD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3044,"prompt_tokens":993,"completion_tokens":2051,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1972}},"tokens_in":609,"tokens_out":2051,"duration_ms":15642,"temperature":1.0,"reasoning_tokens":1972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:49:58.575969+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the entropy distributions of ID and OOD samples in a target domain after training UASA, for example the sketch domain on DomainNet; if a substantial fraction of OOD samples have entropy below the threshold assigned by Equation (5), then Equation (6) misclassifies them as ID and the reported HOS advantage would disappear on that task.","supporting_citations":[{"cited_title":", author Gimpel, K","cited_arxiv_id":null,"evidence_quote":"Supplies the MSP baseline and the entropy-as-OOD-score premise that the adaptive threshold extends."},{"cited_title":", author Li, Y","cited_arxiv_id":null,"evidence_quote":"Provides the ODIN baseline using temperature-scaled softmax scores for OOD detection."},{"cited_title":", author Wang, X","cited_arxiv_id":null,"evidence_quote":"Provides the Energy baseline and energy-based OOD scoring."},{"cited_title":", author Sun, Y","cited_arxiv_id":null,"evidence_quote":"Provides the CIDER baseline using hyperspherical embeddings for OOD detection."},{"cited_title":", author Kim, D","cited_arxiv_id":null,"evidence_quote":"Supplies the DANCE baseline for universal domain adaptation with self-supervision."},{"cited_title":", author Bai, Q","cited_arxiv_id":null,"evidence_quote":"Supplies the DomainNet dataset used for the largest benchmark comparison."},{"cited_title":", author Usman, B","cited_arxiv_id":null,"evidence_quote":"Supplies the VisDA-C dataset, used in its class-imbalanced version."}],"review_version":1}