{"id":"1c8864f7-3b4b-4e16-924c-deb872245598","arxiv_id":"2509.09935","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"SCoDA combines an EMA teacher-student setup and a space similarity loss, starting from an SSL-pretrained model, to adapt to unlabeled target domains, but reports lower accuracy than prior methods on the tested Office-Home and DomainNet tasks.","lead":"A computer vision paper proposes adapting a model to a new image domain using only unlabeled target images, starting from a self-supervised model instead of a labeled source model. The method combines a geometric similarity loss with a slow-updating teacher network, but its own reported results do not beat existing methods on two benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's SOTA claim is directly contradicted by the paper's own results tables and analysis sections.","rationale":"The reader identified the undefined classifier head and the ImageNet/BYOL initialization mismatch as the weakest assumption. That is a serious protocol problem. However, the more decisive issue is that the paper's own reported results directly contradict its central empirical claim: Section 5.3.1 admits SCoDA is below ViLAaD++ on Office-Home, and Section 5.3.2 admits it is below AaD on DomainNet. This is an internal inconsistency, not merely a disagreement with the field's consensus. The abstract and introduction claim state-of-the-art performance, but the tables and analysis show otherwise. Even if the protocol were fully specified, the presented evidence does not support the headline. The lack of code and the small number of reported tasks make independent verification impossible, but the contradiction is already visible in the manuscript. I agree with the REJECT verdict, though my primary reason is the self-contradictory comparison rather than the protocol ambiguity. The idea of combining SSL initialization with geometric distillation may be worth exploring in a properly specified study, but this preprint does not establish it.","tokens_in":11979,"tokens_out":2371,"duration_ms":27249,"concrete_test":"Reproduce SCoDA on the full standard SFDA splits of Office-Home (all 12 tasks) and DomainNet (all 4 tasks), using a fixed classifier head (e.g., the classifier construction from SHOT/AaD) and a clearly stated source initialization (either labeled-source pretraining or ImageNet SSL pretraining). Compute the average accuracy and compare directly with the ViLAaD++ and AaD numbers cited in Tables 1 and 2. If SCoDA's average is not above ViLAaD++ on Office-Home and above AaD on DomainNet, then the abstract's 'significantly outperforms state-of-the-art SFDA methods' claim is falsified by the same protocol.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim, stated in the Abstract and Introduction, is that SCoDA 'significantly outperforms state-of-the-art SFDA methods.' For that claim to hold, the reported comparisons must show SCoDA ahead of existing methods on the benchmarks used. The paper's own numbers refute this. In Section 5.3.1, the authors state that SCoDA 'falls short of the state-of-the-art ViLAaD++ method, which achieves 90.23% average accuracy,' while SCoDA averages 86.59% on the same three Office-Home tasks (Table 1). In Section 5.3.2, they acknowledge SCoDA 'falls short of the AaD baseline (47.12%) by approximately 9.44 percentage' on DomainNet I→C (Table 2). Thus the central empirical assertion is internally inconsistent with the evidence presented in the manuscript. The underspecified classifier protocol and the use of ImageNet-pretrained BYOL instead of source-domain SSL pretraining (Section 3.1.1 vs Section 4.2) further obscure what is being measured, but the decisive problem is simpler: even taking the paper's numbers at face value, the headline claim is false. The comparison is also incomplete—only three of twelve Office-Home tasks and one DomainNet task are reported, so the claimed average cannot be a general benchmark result. No code or reproducible protocol is provided, so these numbers cannot be independently verified. Because the central claim depends on empirical superiority and the paper's own data contradict it, the argument does not hold as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCoDA, a source-free domain adaptation method that replaces supervised source pretraining with a self-supervised (BYOL) initialization and adapts a teacher-student framework with an EMA update rule. The student is trained on unlabeled target data with a composite loss combining cosine feature similarity (Eq. 1) and a Space Similarity Loss (Eq. 2) imported from unsupervised knowledge distillation. The authors claim, in the abstract and introduction, that SCoDA 'significantly outperforms state-of-the-art SFDA methods.' Experiments are reported on two tasks from Office-Home and one from DomainNet, along with ablations over SSL initializations and loss components. The paper's own results, however, show SCoDA below several published baselines on the same benchmarks.","tokens_in":12318,"tokens_out":2693,"duration_ms":32903,"significance":"The idea of replacing supervised source training with a self-supervised initialization and preserving geometric manifold structure during source-free adaptation is interesting and could be a useful direction for the SFDA community. The paper also has some strengths: it explicitly reports the pre-adaptation versus post-adaptation performance, it includes an ablation of the two loss terms, and it documents a catastrophic forgetting failure mode with DINO, which is informative. However, the central empirical claim is contradicted by the paper's own tables. The evaluation protocol is also under-specified, and the benchmark coverage is too limited to support the stated conclusions. If the method were properly evaluated and the claims scaled back, the core idea might merit further study, but as presented the evidence does not support the headline result.","major_comments":[{"comment":"The abstract and introduction claim that SCoDA 'significantly outperforms state-of-the-art SFDA methods.' This is directly contradicted by the paper's own results. Section 5.3.1 states that SCoDA 'falls short of the state-of-the-art ViLAaD++ method, which achieves 90.23% average accuracy,' while Table 1 reports SCoDA at 86.59% on the same three Office-Home tasks. The claim is therefore not supported by the evidence presented, and this is a load-bearing issue for the paper's central contribution.","section":"Abstract and Section 5.3.1"},{"comment":"On DomainNet I→C, Table 2 shows SCoDA achieves 37.68%, while the AaD baseline achieves 47.12%. The text acknowledges a gap of approximately 9.44 percentage points. This further contradicts the abstract's claim of state-of-the-art performance. Since the paper's own numbers place SCoDA behind multiple existing methods, the empirical case for the method is not made.","section":"Section 5.3.2 and Table 2"},{"comment":"There is a critical inconsistency in the initialization protocol. Section 3.1.1 says the model is 'pre-trained on the source domain using a self-supervised learning objective,' but Section 4.2 says the backbone is 'pretrained on ImageNet.' These are different data distributions, and the paper never states how a classifier for 65 or 345 classes is attached to the SSL feature extractor. Without a clear description of the classifier head and how it is initialized or trained, the reported accuracies are difficult to compare with SFDA methods that use a supervised source classifier. This undermines the claim that the method performs SFDA in the standard sense.","section":"Sections 3.1.1 and 4.2"},{"comment":"The first claimed contribution—that replacing supervised source pretraining with SSL is beneficial—is not tested. The paper compares SCoDA (with BYOL initialization) against published SFDA methods that use supervised source models, but it does not include an ablation where SCoDA is initialized with a supervised source model and run with the same adaptation procedure. Such a controlled comparison is necessary to support the claim that the SSL initialization, rather than the adaptation loss or EMA framework, is responsible for the observed results. Without this baseline, the contribution is not validated.","section":"Table 1 and Section 1"}],"minor_comments":[{"comment":"The paper has many typos and formatting issues, including inconsistent capitalization of 'SCoDA'/'SCODA', incomplete ACM CCS placeholder text ('Do Not Use This Code'), and inconsistent reference formatting (some entries use full author names, others initials). These should be cleaned up.","section":"General"},{"comment":"The confusion matrices are not legible in the provided figure and are not described with class labels. This makes the qualitative analysis very difficult to follow.","section":"Figure 4"},{"comment":"The Space Similarity Loss is said to be 'adapted from [16]', but the adaptation to the SFDA setting is not discussed in detail. In particular, the role of the teacher's feature space versus the student's feature space, and whether the loss is computed on all dimensions or a subset, would benefit from clarification.","section":"Equation (2)"},{"comment":"The teacher's BatchNorm statistics are updated via EMA, but the exact procedure and the momentum used are not specified. The main EMA momentum m is given in Section 4.2, but the BN-specific update is not described.","section":"Algorithm 1"},{"comment":"The domain pair I→C is not expanded in the text; Section 5.3.2 refers to 'Infograph→Clipart' but the dataset description only defines DomainNet broadly. A clear statement of the domain names would improve readability.","section":"Table 2"}],"recommendation":"reject","confidential_remarks":"The paper appears to be an early or workshop-style draft: it uses an ACM template with placeholder fields, reports only a small subset of benchmark tasks, and does not provide code. The most serious issue is that the abstract's central claim is contradicted by the paper's own tables. Even if the experimental issues were fixed, the results would require substantial new experiments to support the stated contributions. Given the scope of the claims and the current evidence, I do not see a path to acceptance without major revision and a re-framing of the contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know that the abstract says SCoDA \"significantly outperforms state-of-the-art SFDA methods,\" but the paper's own Section 5.3.1 admits it falls short of ViLAaD++ by 3.64 points on Office-Home, and Section 5.3.2 admits it trails AaD by 9.44 points on DomainNet I→C. The central empirical claim is false on the evidence in front of you. That is the punchline.\n\nWhat's actually new: the hook—replacing supervised source pretraining with a self-supervised (BYOL) pretrained backbone for source-free adaptation—is worth a few minutes of thought. The paper is also honest in places: the ablation across SSL models (Table 3) shows BYOL works, DINO collapses, and Section 5.3.5 lists plausible reasons for that failure. That is a useful data point.\n\nThe soft spots are hard to ignore. The only genuinely new ingredient is the SSL initialization; the Space Similarity loss is imported verbatim from Singh et al. and the EMA teacher-student framework is from CoSDA. More importantly, the protocol is under-specified: Section 3.1.1 says pretrained on the source domain via SSL, Section 4.2 says ImageNet pretrained—those are different distributions. And there is no statement of how the 65- or 345-class classifier is attached to the SSL backbone, so how the \"pre-adaptation\" accuracy of 74.80% is even computed is a mystery. Only three of twelve Office-Home tasks and one DomainNet task are reported, which makes \"extensive experiments\" hard to take at face value. No code.\n\nAll that said, the ideas aren't crazy. If someone re-does this carefully—BYOL initialization, a proper classifier protocol, the full benchmark matrix, and a comparison that doesn't overclaim—the combination might produce a solid workshop or even a main-track result. As it stands, the paper's own numbers break its headline claim, so I would not send this to peer review in its current form.","headline":"The paper's headline claim of state-of-the-art performance is contradicted by its own results tables; the SSL-initialization idea is worth exploring, but this manuscript is not ready for peer review.","tokens_in":12821,"tokens_out":3069,"would_cite":false,"duration_ms":32713,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SCoDA claims that replacing supervised source pretraining with a self-supervised BYOL teacher and adding a Space Similarity Loss that preserves latent geometry yields strong source-free domain adaptation, with 86.59% average accuracy on thr","keywords":["Source-free domain adaptation","Self-supervised learning","Knowledge distillation","Geometric manifold alignment","Teacher-student framework","BYOL","Space similarity loss","Catastrophic forgetting"],"falsifier":"A decisive test: evaluate SCoDA on all 12 Office-Home domain pairs (the paper only reports Ar→Cl, Ar→Pr, Ar→Rw) and compare the average to the three-task subset. A second test: initialize the teacher with a supervised ImageNet-pretrained ResNet-50 (with its classifier) and run the same adaptation; if the supervised start performs equally well, the paper's claim that self-supervised initialization is key would be undercut.","tokens_in":11842,"feed_emoji":"🎯","tokens_out":6152,"duration_ms":64815,"temperature":0.7,"pith_summary":"SCoDA is a source-free domain adaptation (SFDA) framework that discards the standard supervised-pretrained source model and instead starts from a teacher pretrained entirely with self-supervised learning (BYOL). The student adapts to unlabeled target data by minimizing a composite loss that combines instance-level cosine matching with a Space Similarity Loss, which aligns the transposed feature matrices to preserve the teacher's latent manifold geometry. The teacher is updated as a slow EMA of the student, preventing catastrophic forgetting. The paper reports 86.59% average accuracy on three Office-Home adaptation tasks (Ar→Cl, Ar→Pr, Ar→Rw) and 37.68% on a DomainNet I→C task, and argues these results support the two departures: label-agnostic initialization and geometric manifold alignment.","feed_headline":"SSL pretraining plus geometry loss lifts source-free adaptation to 86.6%","feed_subtitle":"SCoDA uses a BYOL teacher and a space-similarity loss to adapt without source labels, reporting 86.6% average on three Office-Home tasks.","key_machinery":"The composite loss L_total = L_cos + λ L_space: L_cos is the standard L2-normalized per-sample cosine similarity between teacher and student features, and L_space is the Space Similarity Loss, which applies cosine similarity to the transposed feature matrices (d×B), comparing each feature dimension's activation pattern over the batch. The teacher is a slowly evolving exponential moving average (EMA) of the student, providing a stable target that mitigates catastrophic forgetting.","core_discovery":"The central claim is that two modifications to the SFDA recipe—starting with a self-supervised rather than supervised teacher, and preserving the teacher's feature-space geometry rather than only matching per-sample L2-normalized vectors—lead to better adaptation and lower catastrophic forgetting. The Space Similarity Loss (adapted from Singh et al., 2022) is the pivotal mechanism: instead of comparing individual feature vectors, it compares the rows of the transposed feature matrices, i.e., the activation of each feature dimension across the mini-batch. This captures the 'internal grammar' of the latent space. In an EMA teacher-student loop, the student learns to match both instance identit","pith_inferences":["If an ImageNet-pretrained SSL encoder suffices as the 'source,' SFDA becomes closer to test-time adaptation, and the term 'source-free' could be reinterpreted as 'label-free' rather than 'access-free.'","The reported 86.59% average is computed on only three of the twelve Office-Home direction pairs; a full 12-task evaluation might change the ranking and should be the next check.","The DINO failure suggests the space-similarity loss may be incompatible with representations built on prototypes or centering/sharpening; adding such stabilization could extend the method to transformers.","Comparing accuracies against supervised-source baselines (e.g., SHOT, AaD with Co-learn) is complicated by the different classifier protocols; a controlled comparison using the same classifier head would clarify the source of the gains."],"forward_implications":["SFDA can be performed without any labeled source data, removing the need for source supervision.","Preserving latent geometry via L_space avoids the information loss caused by L2-normalized instance matching.","The EMA teacher-student design yields near-zero source forgetting (only 0.11% drop on DomainNet I→I), balancing plasticity and stability.","The method transfers across CNN-based SSL initializations (BYOL, SimCLR) but fails with transformer-based DINO, defining a compatibility frontier."],"fun_headline_variants":["Self-supervised teacher + geometry loss beats source-free domain adaptation","No source labels? SCoDA uses SSL pretraining and space similarity to adapt","Forgetting less: SCoDA's EMA teacher and geometry loss improve SFDA","Drop supervised pretraining: SCoDA's SSL teacher and manifold alignment","Geometry-aware adaptation: SCoDA hits 86.6% without source data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that an ImageNet-pretrained BYOL model, without any source-domain training, can serve as the teacher for Office-Home and DomainNet, and that the reported accuracies are comparable to methods that train on labeled source data—a premise the paper does not fully substantiate because it never specifies how a classifier for 65 or 345 classes is attached.","fun_headline_variants_meta":{"raw":{"variants":["Self-supervised teacher + geometry loss beats source-free domain adaptation","No source labels? SCoDA uses SSL pretraining and space similarity to adapt","Forgetting less: SCoDA's EMA teacher and geometry loss improve SFDA","Drop supervised pretraining: SCoDA's SSL teacher and manifold alignment","Geometry-aware adaptation: SCoDA hits 86.6% without source data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000708,"raw_usage":{"total_tokens":3022,"prompt_tokens":734,"completion_tokens":2288,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2202}},"tokens_in":478,"tokens_out":2288,"duration_ms":18969,"temperature":1.0,"reasoning_tokens":2202,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T18:25:56.413010+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test: evaluate SCoDA on all 12 Office-Home domain pairs (the paper only reports Ar→Cl, Ar→Pr, Ar→Rw) and compare the average to the three-task subset. A second test: initialize the teacher with a supervised ImageNet-pretrained ResNet-50 (with its classifier) and run the same adaptation; if the supervised start performs equally well, the paper's claim that self-supervised initialization is key would be undercut.","supporting_citations":[],"review_version":1}