{"id":"eca263a7-5722-44f1-b259-70653906335c","arxiv_id":"2508.12861","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"CoMuCo fine-tunes CLIP with two complementary expert branches and Laplace/Jeffreys regularizers, achieving state-of-the-art few-shot accuracy on a newly collected cross-domain benchmark.","lead":"This paper introduces CoMuCo, a way to fine-tune CLIP-style vision-language models for few-shot image classification that works better when images come from unusual domains like medical scans, satellite photos, or steel defects. It combines two 'expert' branches that see the image from different angles, plus regularizers that keep the model close to its original knowledge, and it introduces a new seven-dataset benchmark for testing cross-domain few-shot learning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multi-view premise is untested: FI and FR start from identical CLIP weights and the Jeffreys consensus (Eq. 5) actively aligns them, so without a quantitative diversity check the reported gains may just reflect 0.2/0.2/0.6 averaging with zero-shot logits.","rationale":"I agree with the reader that the weakest assumption is the complementarity of the two expert branches. The paper's own formulation makes this fragile: both branches start from the same pretrained weights and the consensus term (Eq. 5) minimizes their divergence, so the method relies entirely on the different fine-tuning depths to create diversity. The ablations (Table 1) show that adding the consensus and prior terms helps, but without error bars and without a measure of inter-branch diversity it is impossible to tell whether the improvement comes from the collaborative mechanism or from the extra capacity and the fixed ensemble with zero-shot logits. The missing Table 11, which is supposed to rule out a simple ensemble explanation, is a concrete gap that the authors themselves acknowledge in Appendix C.5. I would not reject the paper: the experiments are extensive, the theoretical claims (Theorems 1 and 2) are standard and correct, and the proposed cross-domain benchmark is a useful contribution. But the central mechanistic claim should be verified with a direct diversity measurement or the ensemble control before acceptance.","tokens_in":23938,"tokens_out":11063,"duration_ms":113537,"concrete_test":"On the three ablation datasets (ImageNet, Stanford Cars, Galaxy10 DECaLS) plus two cross-domain datasets (TCGA12, NWPU-RESISC45) at 16-shot, after training compute the average absolute difference between the softmax outputs pFR and pFI on the test set, and the fraction of test samples where the two branches disagree on the argmax class. Also compute CKA between the pooled feature embeddings of the two branches. If the disagreement rate is below 5% or the mean absolute logit difference is below 0.01, the branches are functionally identical and the consensus term is vacuous; the gains would then be attributable to the fixed logit averaging with zero-shot CLIP rather than to multi-view collaborative optimization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CoMuCo's central claim is that two functionally complementary experts, FI and FR, provide multi-view features whose consensus (Jeffreys divergence, Eq. 5) improves cross-domain few-shot CLIP transfer (Abstract, Dual-Expert Framework). However, both modules are initialized from the same CLIP weights (Fig. 3) and are trained with the same cross-entropy loss on the same data; the only architectural difference is the fine-tuning depth (last module vs. last layer). The consensus term minimizes the divergence between the two branches' output distributions, actively reducing diversity rather than enforcing it. Nothing in Eq. 9 rewards complementary features, so the branches may become near-duplicates. In that case LD is trivially satisfied, and the reported gains (e.g., +7.04% average on the cross-domain benchmark, Appendix C.2) could be caused by the fixed 0.2/0.2/0.6 logit averaging with zero-shot CLIP, not by multi-view collaboration. The paper's only evidence for complementarity is qualitative (GradCAM and t-SNE in Appendix D), and the ensemble control claimed in Appendix C.5 (Table 11) is missing from the provided text. Without a quantitative diversity measure, the causal role of the proposed mechanism is unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoMuCo, a fine-tuning strategy for CLIP in few-shot settings, especially under cross-domain shifts. The method has two expert modules, a Feature Integrator (FI) trained shallowly and a Feature Refiner (FR) trained more deeply, whose logits are fused with frozen zero-shot CLIP logits. Two regularizers are added: an L1-style prior consistency loss that keeps each branch's logits close to CLIP zero-shot logits, and a Jeffreys-divergence consensus loss that aligns the two branches' predictive distributions. The paper also introduces a new cross-domain few-shot benchmark of seven datasets and reports experiments on this benchmark and on the CLIP Benchmark with ResNet-50 and ViT-B/16, claiming consistent state-of-the-art performance.","tokens_in":24202,"tokens_out":10345,"duration_ms":93411,"significance":"If the empirical claims hold, CoMuCo is a simple and plausible approach for transferring CLIP to domains that differ from its natural-image pretraining distribution. The new cross-domain benchmark addresses a real gap in few-shot CLIP evaluation, and the proposed regularizers are easy to implement. The theoretical components are standard: the Laplace/L1 equivalence is well known, and the Jeffreys divergence is a standard symmetric KL divergence; the geometric framing is motivation rather than a load-bearing derivation. The paper plans to release code and the benchmark, which would aid reproducibility. However, the empirical support as presented is incomplete: no variance estimates are reported, the central multi-view complementarity premise is not quantitatively verified, several referenced tables and proofs are missing, and some headline numbers conflict with the tables. These issues currently limit the significance of the claimed results.","major_comments":[{"comment":"All results are stated to be averaged over three seeds, but no standard deviation, confidence interval, or per-seed result is reported anywhere in Tables 6, 7, 9 or Figures 4 and 5. Several headline gains are small; for example, the ResNet-50 1-shot cross-domain average improvement is 0.9 percentage points (Table 6: 38.19 vs. 37.29), and on individual 1-shot datasets CoMuCo is worse than the best baseline (Table 6: Skin40 27.57 vs. TCP 29.40; TCGA12 18.23 vs. TCP 24.57). Without variance estimates or significance tests, the claim that CoMuCo 'consistently outperforms current methods' is not statistically supported. Please report standard deviations over the three seeds and, ideally, per-dataset win/loss counts.","section":"Experiments, Implementation; Tables 6, 7, 9"},{"comment":"The central mechanism is multi-view collaboration, but the manuscript does not establish that FI and FR actually provide complementary views. Both modules are initialized from the same CLIP weights and are trained with the same cross-entropy objective on the same data; the only architectural difference is fine-tuning depth. The consensus loss in Eq. (5) minimizes the Jeffreys divergence between the two branches, and the total objective in Eq. (9) contains no term that rewards diversity, so the branches could collapse to near-duplicate predictors, making the consensus term vacuous. The only evidence offered in Appendix D is qualitative (GradCAM and t-SNE). Please report a quantitative diversity measure, such as the average cosine similarity between FI and FR features or logits before and after training, or the converged value of the Jeffreys term, and include the ensemble control from Appendix C.5 (Table 11), which is currently missing. Without this, the reported gains over the fixed 0.2/0.2/0.6 logit averaging with zero-shot CLIP cannot be attributed to the proposed collaborative mechanism.","section":"Eq. (5), Eq. (9), Dual-Expert Framework, Appendix C.5"},{"comment":"The numerical claims in the main text are not consistent with the reported tables. For ViT-B/16 on the cross-domain benchmark, the text reports improvements of 2.03%, 3.23%, 3.59%, 2.83%, and 2.78% over the best baseline for 1/2/4/8/16-shot, but Table 7 yields best-baseline margins of 2.26, 2.72, 2.43, 2.47, and 1.75 percentage points. For ResNet-50 on the CLIP Benchmark, the main text reports gains of 1.48%, 2.77%, and 4.65% for 1/4/16-shot, while Appendix C.3 reports 1.55%, 2.82%, and 4.67% for the same settings. The authors should reconcile these numbers and state exactly which baseline and which table each margin is computed against.","section":"Efficacy of the Proposed Method, Appendix C.3; Tables 7 and 9"},{"comment":"The new cross-domain benchmark is self-constructed and currently lacks external validation, and it contains internal inconsistencies that block reproduction. The RFMiD dataset is called 'RFMiD12' in Tables 5-7 and the main text, but Appendix B.1 says the benchmark uses a subset of the 15 most populous classes and Table 3 lists 15 classes; the relation between the name 'RFMiD12' and the 15-class subset is unclear. The exact train/validation protocol is also not specified for every dataset (e.g., TCGA12 and Skin40 are listed without a validation split, while IP102 uses its official split). Please clarify the class counts and splits, and provide the exact sampling procedure or image indices to enable external checks.","section":"Appendix B.1, Table 3, Appendix C.2"}],"minor_comments":[{"comment":"The first row of Table 1 has no components checked and its values (58.18, 55.61, 13.90) coincide exactly with the zero-shot CLIP numbers in Tables 9 and 6. If this row is the zero-shot CLIP baseline, label it as such; if it is an LCE-only trained model, explain why training with 16-shot cross-entropy leaves the accuracy unchanged to two decimals.","section":"Table 1"},{"comment":"The text contains an unresolved cross-reference '?? shows' when discussing the fine-tuning depth study; this should point to the relevant figure (apparently Fig. 9).","section":"Impact of Fine-tuning Layer Configurations in FR"},{"comment":"Theorem 1 and Theorem 2 refer to proofs in Appendix A.1 and A.2, but the appendices included in the manuscript start at B; the proofs should be included or the references corrected.","section":"Theorem 1 and Theorem 2"},{"comment":"Appendix C.3 states that results are presented in 'Tab. 9 and Tab. 10', and Appendix C.5 refers to 'Tab. 11'; Tables 10 and 11 do not appear in the manuscript. Please include them.","section":"Appendix C.3 and C.5"},{"comment":"The abstract and conclusion say CoMuCo 'consistently outperforms current methods', but Tables 6 and 7 show several per-dataset cases where it is below the best baseline (e.g., Table 6, 1-shot Skin40 and TCGA12; Table 7, 1-shot IP102). Please qualify the claim as average performance or report per-dataset win/loss counts.","section":"Abstract and Conclusion"},{"comment":"In Eq. (7), the coefficients alpha, beta, and gamma are introduced without constraints; the text later fixes alpha = beta = 0.2, implying gamma = 0.6, but it would be clearer to state the simplex constraint gamma = 1 - alpha - beta explicitly.","section":"Eq. (7)"},{"comment":"Figures 4 and 5 show a legend that omits LP++, although LP++ results appear in Tables 6, 7, and 9; please either add it to the figures or explain why it is omitted.","section":"Figures 4 and 5"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible mechanism and a useful benchmark idea, but the empirical claims are not yet fully supported by the provided artifacts: no variance estimates, inconsistent margins between text and tables, missing tables and proofs, and an unverified complementarity assumption. These issues are fixable with additional experiments and careful reporting, so I recommend major revision rather than rejection. The authors should be asked to add uncertainty quantification, a diversity/complementarity analysis, the missing tables, and a reconciliation of all reported improvement numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know about arXiv:2508.12861. First, it is a genuinely useful recipe for CLIP fine-tuning in cross-domain few-shot settings, evaluated over an unusually broad set of backbones and benchmarks, plus a new seven-dataset cross-domain benchmark that the community would likely adopt. Second, the paper's central story — that two 'functionally complementary' expert branches learn multi-view features — is not actually supported by the evidence. The branches start from identical CLIP weights, the consensus loss actively pulls their output distributions together, and no quantitative diversity measure or ensemble control is provided. The reported gains may come from the fixed 0.2/0.2/0.6 weighted averaging with zero-shot logits as much as from any collaboration between the branches.\n\nWhat is new: the specific combination of a conservative feature integrator (last-module fine-tuning) and a more adaptive feature refiner (last-layer fine-tuning) with L1 logit-drift regularization (Laplace prior) and Jeffreys divergence consensus. The components are individually known, but not in this combination. The new benchmark, with datasets like TCGA12, RFMiD12, NEU-CLS, and Galaxy10, fills a real gap: most CLIP few-shot evaluations are on natural-image benchmarks with limited domain shift. The paper deserves credit for the breadth of experiments — CLIP Benchmark, the new benchmark, domain generalization to ImageNet-V2/Sketch, ResNet-50/101, ViT-B/32/B/16, and a hyperparameter sensitivity study.\n\nThe soft spots are mostly empirical. No standard deviations are reported despite averaging over three seeds; the 'consistently outperforms' claim in the abstract is too strong given that on several datasets (e.g., Skin40 1-shot) CoMuCo is below the best baseline. More serious is the ablation table: the LCE-only row reports 58.18 on ImageNet, exactly equal to zero-shot CLIP, and the same pattern on Stanford Cars and Galaxy10. That suggests that without the regularizers the branches do not move, which is either a bug or a sign that the regularizers are doing all the work — in either case, it undercuts the multi-view story and needs an explanation. The self-constructed benchmark has no external validation yet, and code and data are promised but not released. The appendix ensemble control (Table 11) is referenced but not present in the manuscript text, and one figure reference is a broken '??'.\n\nThe multi-view premise is the weakest conceptual point. Nothing in the objective rewards complementarity; the Jeffreys term only penalizes divergence. The authors offer GradCAM and t-SNE as qualitative evidence, but that is not enough to establish that the two branches learn different functions. This is fixable: report disagreement rates or a feature-similarity measure, and provide the missing ensemble control.\n\nWho this is for: anyone working on efficient CLIP transfer, especially in medical, remote sensing, or industrial domains. It deserves a serious referee, but with a request for major revision — release the code and benchmark, give error bars, explain or fix the LCE-only ablation, and add a quantitative diversity analysis. If the numbers hold, it's a solid contribution; as written, the main mechanism is unverified.","headline":"Useful fine-tuning recipe and a good new benchmark, but the multi-view story is untested and an ablation row looks off.","tokens_in":24782,"tokens_out":4855,"would_cite":false,"duration_ms":49466,"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":"The paper proposes CoMuCo, a two-expert fine-tuning strategy with Laplace and Jeffreys regularizers, and claims it consistently beats existing prompt- and adapter-tuning methods on few-shot benchmarks, especially under cross-domain shift.","keywords":["few-shot learning","cross-domain","vision-language models","CLIP","efficient transfer learning","Jeffreys divergence","Laplace prior","multi-view learning"],"falsifier":"Train CoMuCo on a cross-domain dataset such as Galaxy10 DECaLS at 16-shot, then measure the Jeffreys divergence between the FR and FI prediction distributions on the test set; if the divergence is near zero on almost every sample, the two views have collapsed and the consensus term is not doing the work. A second decisive check is to ablate the fusion weights: if setting $\\alpha=\\beta=0$ (using only zero-shot CLIP logits) and training FI/FR separately reproduces most of the accuracy gains, the multi-view mechanism is not responsible.","tokens_in":23691,"feed_emoji":"🎯","tokens_out":8706,"duration_ms":76547,"temperature":0.7,"pith_summary":"CoMuCo is a fine-tuning strategy for vision-language models aimed at few-shot classification when the target domain is far from natural images—skin lesions, pathology slides, retinal fundus images, remote sensing scenes, steel defects, crop pests, and galaxy images. It trains two experts from the same CLIP checkpoint: a Feature Integrator that changes only the final attention-pooling block, and a Feature Refiner that tunes deeper layers, so the two views favor stability and adaptation respectively. Two regularizers steer the training: an L1 penalty on each expert's logit deviation from frozen zero-shot CLIP (equivalent to a Laplace prior) that preserves general knowledge, and a Jeffreys-divergence consensus term that pulls the two experts' prediction distributions together. On both the established CLIP benchmark and a new seven-dataset cross-domain benchmark, the paper reports consistent gains over existing prompt- and adapter-tuning methods, with margins that grow as the number of training images per class increases.","feed_headline":"Two CLIP experts beat prior few-shot tuners on cross-domain data","feed_subtitle":"Laplace priors preserve zero-shot knowledge; Jeffreys consensus aligns the two experts. Gains widen as shots grow.","key_machinery":"The load-bearing objects are the two expert modules—Feature Integrator (FI) and Feature Refiner (FR)—and the two regularization terms that coordinate them. FI is the invariant expert, updated only at the final attention-pooling block; FR is the adaptive expert, fine-tuned in deeper layers. The prior consistency constraint models each expert's logit deviation from frozen zero-shot CLIP as a zero-mean Laplace distribution, which is equivalent to $\\ell^1$ regularization on the logit offset, a mechanism that permits sparse class-level changes without erasing general knowledge. The multi-view consensus constraint uses Jeffreys divergence, the symmetric KL divergence $D_J(p\\parallel q)=D_{KL}(p\\parallel q)+D_{KL}(q\\parallel p)$, as a fourth-order approximation to the squared geodesic distance between the two experts' prediction distributions on the statistical manifold; minimizing it aligns the two views. These pieces work together: FI supplies stability, FR supplies adaptation, the Laplace term prevents forgetting, and the Jeffreys term makes the two views compatible, with final logits fused as $s_i=\\alpha s_{FR}(x_i)+\\beta s_{FI}(x_i)+\\gamma s_{ZS}(x_i)$.","core_discovery":"The paper's central claim is that the two-branch architecture, plus the two logit-space regularizers, is what lets CLIP transfer to genuinely different imaging domains. The Feature Integrator is the conservative expert: it keeps most of the pre-trained model intact and refines only attention pooling at the last block, preserving knowledge that is already relevant. The Feature Refiner is the adaptive expert: it tunes deeper layers to learn task-specific patterns the pre-training never saw. The final prediction is a weighted fusion of the two experts' logits and the frozen zero-shot CLIP logits ($\\alpha=0.2$, $\\beta=0.2$, $\\gamma=0.6$). The prior consistency term penalizes the $\\ell^1$ norm of each expert's logit offset from zero-shot CLIP, enforcing sparse changes; the consensus term minimizes the Jeffreys divergence between the experts' predictive distributions, which the paper proves approximates the squared geodesic distance on the statistical manifold to fourth order. Empirically, on the new cross-domain benchmark with ResNet-50, the reported average improvements over the best baseline are 0.9, 2.25, 2.78, 5.27, and 7.04 points at 1, 2, 4, 8, and 16 shots, with analogous gains under ViT-B/16 and on the standard CLIP benchmark.","pith_inferences":["If the two experts are truly complementary, a testable extension is to replace the Jeffreys consensus term with another symmetric divergence (for example, the Bhattacharyya distance) and compare; similar gains would suggest the specific geometry matters less than the presence of a two-view agreement constraint.","A further extension is to initialize FI and FR from a prompt-tuned or adapter-tuned CLIP rather than from the frozen checkpoint, to see whether the collaborative mechanism stacks with existing transfer methods instead of only replacing them.","The paper's own tiering of the benchmark by zero-shot CLIP accuracy implies a prediction the authors leave implicit: gains should be largest on high-shift datasets such as the pathology and fundus sets, and smaller on low-shift sets such as remote sensing; readers can check this pattern in the reported per-dataset numbers.","Because the consensus term is only meaningful if the two branches diverge, a direct check of FI/FR representation similarity after training would settle whether the reported gains come from collaboration or from the fixed 0.2/0.2/0.6 fusion with zero-shot CLIP."],"forward_implications":["Efficient CLIP transfer to distant domains does not require large prompt or adapter modules; two modestly tuned branches plus two logit-space regularizers are enough to improve on existing methods.","The gains are not tied to one visual backbone: the paper reports consistent improvements with ResNet-50, ResNet-101, ViT-B/32, ViT-B/16, and ConvNeXt-Base, so the mechanism transfers across architectures.","The new seven-dataset cross-domain benchmark gives the community a common testbed where zero-shot CLIP is weak, allowing future few-shot methods to be compared under genuine domain shift.","Because the reported margins on the cross-domain benchmark widen with sample size (from 0.9 to 7.04 points on ResNet-50), the method is most valuable when a modest number of images is available, not only in the extreme 1-shot regime.","The domain-adaptation results (ImageNet trained, ImageNet-V2 and ImageNet-Sketch evaluated) indicate that the features learned by the two experts remain general enough to transfer beyond the training domain."],"supporting_citations":[{"why":"Supplies the CLIP backbone whose frozen zero-shot logits anchor the prior consistency term and the final fused prediction.","marker":"(Radford et al. 2021)"},{"why":"Defines the prompt-tuning baseline (CoOp) and the established CLIP Benchmark protocol that the paper evaluates against.","marker":"(Zhou et al. 2022b)"},{"why":"Provides the Tip-Adapter baseline and the text templates used for ImageNet-series datasets.","marker":"(Zhang et al. 2022)"},{"why":"Provides the TaskRes baseline and the pre-defined text templates used for text encoding.","marker":"(Yu et al. 2023)"},{"why":"Provides the TCP baseline, the strongest competitor in several comparisons.","marker":"(Yao, Zhang, and Xu 2024)"},{"why":"Provides the MaPLe multi-modal prompt-tuning baseline.","marker":"(Khattak et al. 2023)"},{"why":"Provides the DePT prompt-tuning baseline.","marker":"(Zhang et al. 2024)"},{"why":"Provides the SkipT adapter baseline.","marker":"(Wu et al. 2025)"},{"why":"Supplies the Skin40 dataset for the new cross-domain benchmark.","marker":"(Yang et al. 2023)"},{"why":"Supplies the TCGA12 pathology dataset for the new cross-domain benchmark.","marker":"(Chen et al. 2022)"}],"fun_headline_variants":["CoMuCo: two CLIP experts with sparse priors and geometric consensus","Dual CLIP experts, Jeffreys consensus: cross-domain few-shot gains","Conservative + adaptive CLIP heads: cross-domain few-shot winner","CoMuCo: two experts, one geometry, better cross-domain few-shot","Sparse priors, Jeffreys consensus: CLIP that learns across domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Feature Integrator and Feature Refiner, both initialized from the same CLIP weights and both pulled toward the zero-shot CLIP logits, learn genuinely different and complementary features; if the two branches collapse into near-identical representations, the Jeffreys consensus term becomes vacuous and the reported gains would reduce to the fixed logit averaging with frozen CLIP.","fun_headline_variants_meta":{"raw":{"variants":["CoMuCo: two CLIP experts with sparse priors and geometric consensus","Dual CLIP experts, Jeffreys consensus: cross-domain few-shot gains","Conservative + adaptive CLIP heads: cross-domain few-shot winner","CoMuCo: two experts, one geometry, better cross-domain few-shot","Sparse priors, Jeffreys consensus: CLIP that learns across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3524,"prompt_tokens":1011,"completion_tokens":2513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":2415}},"tokens_in":627,"tokens_out":2513,"duration_ms":19881,"temperature":1.0,"reasoning_tokens":2415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:18:06.692345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CoMuCo on a cross-domain dataset such as Galaxy10 DECaLS at 16-shot, then measure the Jeffreys divergence between the FR and FI prediction distributions on the test set; if the divergence is near zero on almost every sample, the two views have collapsed and the consensus term is not doing the work. A second decisive check is to ablate the fusion weights: if setting $\\alpha=\\beta=0$ (using only zero-shot CLIP logits) and training FI/FR separately reproduces most of the accuracy gains, the multi-view mechanism is not responsible.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Tip-Adapter baseline and the text templates used for ImageNet-series datasets."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the TCP baseline, the strongest competitor in several comparisons."},{"cited_title":"U.; Rasheed, H","cited_arxiv_id":null,"evidence_quote":"Provides the MaPLe multi-modal prompt-tuning baseline."},{"cited_title":"T.; and Song, J","cited_arxiv_id":null,"evidence_quote":"Provides the DePT prompt-tuning baseline."},{"cited_title":"J.; Chen, C.; Li, Y.; Chen, T","cited_arxiv_id":null,"evidence_quote":"Supplies the TCGA12 pathology dataset for the new cross-domain benchmark."}],"review_version":2}