{"id":"b64ddd3d-754f-43f1-a542-146688787a95","arxiv_id":"2506.20841","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A plug-and-play contrastive regularization term, FixCLR, repels different pseudo-classes across domains and improves semi-supervised domain generalization accuracy when combined with FixMatch-based methods.","lead":"FixCLR adds a contrastive repelling loss to semi-supervised domain generalization, using pseudo-labels to push different predicted classes apart across all training domains. The paper reports accuracy gains when FixCLR is stacked on FixMatch and other semi-supervised methods, with the largest gains on Terra Incognita and Digits DG.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Excluding Terra Incognita, FixCLR's claimed consistent advantage over FBC-SA largely disappears; the core empirical claim rests on per-dataset best-method cherry-picking without error bars.","rationale":"The reader's verdict is CONDITIONAL, and this stress-test does not move it. The reader identified pseudo-label quality as the weakest assumption; that is a reasonable concern, and the paper's Figure 2 does use the same model's own pseudo-labels as evidence. However, the more load-bearing issue is the empirical consistency claim. FixCLR's advantage over FBC-SA is concentrated in Terra Incognita, and the paper's summary statements ('consistently leads to larger improvements', 'consistently achieves the best results') are contradicted by its own tables on several datasets. Tables 3 and 5 select the best method per dataset after the fact, which inflates the appearance of consistency. No error bars, seeds, or significance tests are reported, so the small reported margins cannot be distinguished from noise. This is a correctness risk in the central claim, not a novelty or framing issue. It is also directly testable: rerunning with multiple seeds and a paired test would settle whether FixCLR's advantage is real and general or specific to Terra Incognita. The paper may still be worth publishing conditionally, because the method is simple, the motivation is clear, and there are plausible gains on some datasets, but the headline claim needs to be scaled back or supported with statistical evidence.","tokens_in":16118,"tokens_out":3351,"duration_ms":32941,"concrete_test":"Re-run all configurations in Tables 2 and 4 with at least 5 seeds, reporting mean and standard deviation. For each base method and dataset, compute the paired delta (FixCLR - FBC-SA) in improvement over the base method. Then check: (i) the fraction of datasets where FixCLR's delta is positive, (ii) the same fraction excluding Terra Incognita, and (iii) a paired Wilcoxon signed-rank test across datasets for each base method and for the pooled set. If the positive fraction drops below 0.5 when Terra is excluded, or the paired test is non-significant, the claim of consistent improvement must be weakened to dataset-specific gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that FixCLR 'consistently' leads to larger improvements than FBC-SA and 'consistently achieves the best results' when combined with semi-supervised methods. The data in Tables 2 and 4 do not support consistency. In Table 2, FixCLR's improvement over the base method beats FBC-SA's improvement by large margins on Terra Incognita (e.g., +13.5 vs +2.3 for SoftMatch, +10.0 vs +3.3 for FreeMatch, +9.2 vs +0.7 for StyleMatch), but on Digits DG, PACS, Office-Home, and FMOW-Wilds the comparison is mixed or favors FBC-SA (e.g., DebiasPL: FixCLR +7.2 vs FBC-SA +7.5 on Digits, +0.7 vs +2.0 on PACS, +1.2 vs +3.0 on Office-Home). Counting all pretrained base-method rows, FixCLR has a larger positive delta than FBC-SA in only roughly half of the comparisons; excluding Terra it is in the minority. Table 4 (non-pretrained) shows the same pattern, with FixCLR deltas frequently negative or smaller than FBC-SA's (e.g., FreeMatch -0.7 on Digits, -2.8 on PACS; SoftMatch +1.6 vs FBC-SA +3.5 on Digits). Tables 3 and 5 compound this by selecting, per dataset, whichever method combination happens to be highest, so the 'consistently best' statement is an artifact of table-mining rather than a property of FixCLR. The pseudo-label quality assumption flagged by the reader is plausible, but the decisive weakness is empirical: without seeded repeats, error bars, or a paired significance test, differences of 0.5-2 accuracy points are within run-to-run noise, and the one dataset carrying the qualitative conclusion is Terra Incognita.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FixCLR, a contrastive regularization term for semi-supervised domain generalization (SSDG). FixCLR is added to a FixMatch-style base objective (L = LS + LU + LC) and is designed to explicitly encourage domain-invariant representations by repelling samples that have different pseudo-labeled classes, regardless of domain, while deliberately omitting positive attraction. The paper claims that FixCLR is a plug-and-play regularizer that, when combined with various semi-supervised and SSDG methods, consistently improves accuracy over both the base methods and the FBC-SA baseline, especially on datasets with many domains. The authors also present efficiency measurements, pseudo-label quality/quantity analyses, an ablation removing positive attraction, and comparisons of pretrained versus non-pretrained backbones across six datasets.","tokens_in":16471,"tokens_out":5381,"duration_ms":49808,"significance":"If the empirical claims were supported, FixCLR would be a useful, simple, and efficient plug-and-play regularizer for SSDG, and the paper's broad benchmarking—including many-domain datasets, multiple pseudo-label improvement methods, and both pretrained and non-pretrained settings—would be a valuable contribution to the community. The idea of using negative-class repelling based on pseudo-labels is coherent and mechanistically plausible, and the efficiency advantage over FBC-SA and StyleMatch is clearly demonstrated in Table 6. However, the central empirical claims of consistency and superiority over FBC-SA are not supported by the paper's own tables, and the absence of variance estimates or significance tests makes the reported differences unreliable.","major_comments":[{"comment":"The claim that 'FixCLR consistently leads to larger improvements compared to FBC-SA' is contradicted by the paper's own results. In the pretrained Table 2, FBC-SA gives larger positive deltas than FixCLR for DebiasPL on Digits (+7.5 vs +7.2), PACS (+2.0 vs +0.7), and Office-Home (+3.0 vs +1.2); SoftMatch on PACS (+1.5 vs +1.3); and StyleMatch on Digits (+6.3 vs +5.8). In the non-pretrained Table 4, the pattern is even weaker: DebiasPL on Digits (+5.0 vs +1.0) and PACS (+3.7 vs +1.5), SoftMatch on Digits (+3.5 vs +1.6), and StyleMatch on Digits (+5.0 vs +1.5) all favor FBC-SA, and several FixCLR deltas are negative (e.g., DeFixMatch on Digits -3.5, FreeMatch on PACS -2.8). Excluding Terra Incognita, where FixCLR does show large gains, the claimed consistent advantage largely disappears. The 'consistently' wording should be removed or the statement needs to be restricted to specific datasets with statistical support.","section":"Section 4, Tables 2 and 4"},{"comment":"The statement that FixCLR 'consistently achieves the best results' is an artifact of per-dataset selection: Table 3 chooses a different combination for each dataset (FixCLR+SoftMatch for Digits, Terra, IMG-R; FixCLR+StyleMatch for PACS, Office-Home, FMOW), and Table 5 does the same. Since the 'best' base method differs by dataset and by pretraining condition, the tables do not demonstrate that any one combination is consistently best; at most they show that some FixCLR-augmented method can be selected to top each column. A proper test would either fix a single combination across all datasets or report a pairwise ranking (e.g., average rank of FixCLR vs FBC-SA across all base methods and datasets).","section":"Tables 3 and 5"},{"comment":"The paper reports only 'average accuracy across four runs' with no standard deviations, confidence intervals, or significance tests. Many of the claimed improvements are small (e.g., 0.3–2 percentage points), and paired differences of this size across four leave-one-domain-out runs are within normal run-to-run noise. For example, Table 2 shows FixMatch at 60.3% on Digits and DebiasPL at 60.3%, while Table 4 shows SoftMatch +1.6 vs FBC-SA +3.5 on Digits; without error bars or a paired test, these numbers cannot support the conclusion that FixCLR is consistently better. The authors should report per-run or per-target-domain results, standard deviations, and ideally a paired significance test (e.g., Wilcoxon signed-rank) comparing FixCLR and FBC-SA over the same base methods and datasets.","section":"Section 3.1 and Section 4"},{"comment":"The pseudo-label quality improvement claim is partly circular. Figure 2 measures the accuracy of pseudo-labels produced by the very same model that is being trained with LC (Eq. 1), and those pseudo-labels are used to compute LC itself. The observed 'improvement in quality' could simply reflect the model becoming more confident on samples it already labels correctly, rather than an independent benefit of domain-invariant representations. To validate the mechanism, the authors should either compute pseudo-label accuracy using a fixed pseudo-label set generated by a reference model, or evaluate on a held-out labeled subset that is not used for pseudo-label generation or training. Without such a check, the claim that FixCLR 'improves pseudo-label quality' is not independently established.","section":"Section 4, 'Effects of FixCLR' and Figure 2"}],"minor_comments":[{"comment":"The numerator 'exp(1/tau)' in Eq. (1) is unexplained; it appears to be a constant placeholder replacing the positive-attraction term. Please clarify this notation, since a reader might otherwise think there is a fixed positive similarity of 1, and state whether exp(1/tau) was chosen as a constant that does not affect the gradient direction.","section":"Section 3, Eq. (1)"},{"comment":"There is a typo in 'Terra Incongnita'; it should be 'Terra Incognita'. Also, 'signficiant' appears in Section 4.","section":"Section 3.1"},{"comment":"The text states 'We did not include StyleMatch in these combinations because its use of a pretrained network for style-augmentation may provide an unfair advantage,' but Table 2 and Table 4 both list rows for 'StyleMatch +FBC-SA' and 'StyleMatch +FixCLR'. This is an internal contradiction that needs correction or clarification.","section":"Section 4, paragraph after Table 2"},{"comment":"Reference [32] is cited as the source of the ResNet-18 architecture, but the listed paper ('Resnet in resnet: Generalizing residual architectures') is not the standard ResNet-18 reference; the correct citation is He et al., 'Deep Residual Learning for Image Recognition' (CVPR 2016).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict is appropriate. The proposed method is simple and the experiments are extensive, but the paper's central empirical claims ('consistently larger improvements', 'consistently achieves the best results') are not supported by the presented data, and the lack of error bars or significance testing makes the quantitative comparisons unverifiable. The manuscript is not ready for acceptance in its current form; a major revision that reanalyzes the experiments, adds variance measures and significance tests, and revises the claims accordingly would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the core idea is neat and the paper is honest about its method, but the 'consistently larger improvements' claim doesn't hold up when you actually count wins in Tables 2 and 4. The advantage is concentrated in Terra Incognita; elsewhere FixCLR and FBC-SA are roughly a wash, and on non-pretrained models FBC-SA sometimes does better.\n\nWhat's new: the repelling-only contrastive term with pseudo-label class grouping is a sensible adaptation of SimCLR to SSDG. SelfReg uses same-class attraction, PCL needs labels and proxies; FixCLR fills a real gap. The decision to avoid positive attraction is well-motivated and the ablation (Table 7) supports it—adding positives consistently hurts, which is a useful empirical datapoint consistent with PCL's findings. The efficiency comparison is also in FixCLR's favor: no extra forward passes, and the wall-clock numbers show real savings versus FBC-SA and StyleMatch.\n\nWhere it gets soft: no error bars anywhere. The tables report averages over four runs, and with differences of 0.5–2 accuracy points, run-to-run noise alone could flip many of these comparisons. The paper's central sentence—'FixCLR consistently leads to larger improvements compared to FBC-SA'—is contradicted by the paper's own tables. Counting the pretrained table, FixCLR wins roughly half the row/dataset comparisons; excluding Terra Incognita it's in the minority. The non-pretrained table is similar, and in several rows FBC-SA clearly dominates. Tables 3 and 5 make it worse by picking the best method per dataset; that is table-mining, not evidence of consistency.\n\nThe loss definition (Eq. 1) is also under-specified. It's not clear whether 'any sample' means one randomly chosen sample per (domain, class) pair or all samples in the batch, and the sum over domains with a single log for each is confusing. This matters for reproducibility, especially since no code is provided.\n\nOn the pseudo-label quality story: Figure 2 measures accuracy against ground truth, so it's not circular in the narrow sense. The concern is more that the causal claim—that domain-invariant regularization improves pseudo-label quality—is plausible but not independently validated; the quantity drop is consistent with their story, so I'd call this a soft spot, not a fatal flaw.\n\nBottom line: the method is reasonable, the experiments are extensive, and the central direction is probably useful for SSDG. But the empirical claims need statistical grounding and a more honest summary. It should go to review—the idea is novel enough and the experiments are broad enough to deserve referee time—but the authors need to add error bars, use paired tests, and stop saying 'consistently' until the data support it.","headline":"A clean regularizer idea for SSDG whose empirical case is real but narrower than the paper claims; the advantage over FBC-SA mostly lives on Terra Incognita and, without error bars, the consistency claim is unsupported.","tokens_in":17038,"tokens_out":7296,"would_cite":false,"duration_ms":63645,"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":"FixCLR adds a repulsion-only contrastive loss to FixMatch to train domain-invariant representations for semi-supervised domain generalization.","keywords":["semi-supervised domain generalization","negative-class contrastive learning","domain-invariant representations","pseudo-labeling","FixMatch","contrastive regularization","out-of-distribution generalization","representation learning"],"falsifier":"Train FixCLR and FixMatch on the same SSDG datasets while injecting increasing noise into the pseudo-labels used only by $\\mathcal{L}_C$; if target-domain accuracy and domain-cluster removal do not degrade faster with noise than they do for FixMatch, then pseudo-label accuracy is not the mechanism driving the improvement.","tokens_in":15893,"feed_emoji":"🧲","tokens_out":8909,"duration_ms":83023,"temperature":0.7,"pith_summary":"Semi-supervised domain generalization (SSDG) trains on a few labeled and many unlabeled examples from known domains so the model can handle an unseen domain. The paper claims that existing SSDG methods, which build on FixMatch, only implicitly encourage domain invariance, leaving domain-specific clusters in the representation space. FixCLR adds a loss term that groups samples by their predicted class across all domains and repels every other predicted class regardless of domain, with no positive attraction between same-class samples. The paper argues this explicit repulsion removes domain clusters and improves accuracy, and that FixCLR can be stacked on top of most SSDG or semi-supervised methods for complementary gains.","feed_headline":"Repelling wrong classes boosts domain generalization with few labels","feed_subtitle":"FixCLR adds a contrastive repulsion term to FixMatch, improving accuracy across six datasets when labels are scarce.","key_machinery":"The central object is the FixCLR repulsion loss in Eq. (1): $\\mathcal{L}_C = \\sum_i -\\log\\left( \\exp(1/\\tau) \\; / \\; \\sum_j \\exp(\\operatorname{sim}(\\operatorname{DOM}^i_{-j}, \\operatorname{CLS}_j)/\\tau) \\right)$, with temperature $\\tau = 0.5$. Here $\\operatorname{CLS}_j$ is the set of latent representations of samples predicted to belong to class $j$ in any domain, and $\\operatorname{DOM}^i_{-j}$ is a sample from domain $i$ predicted not to belong to class $j$. Because the numerator is a fixed constant, the loss only repels: the model is trained to make every class group dissimilar to all samples of other predicted classes across all domains. This term is added on top of FixMatch's labeled and unlabeled cross-entropy losses and requires no extra forward pass.","core_discovery":"The central claim is that adding the FixCLR loss $\\mathcal{L}_C$ to the FixMatch objective, $\\mathcal{L} = \\mathcal{L}_S + \\mathcal{L}_U + \\mathcal{L}_C$, explicitly regularizes the model to learn domain-invariant representations given limited labels. The loss minimizes the cosine similarity between each pseudo-label class group and all samples predicted to belong to other classes in every domain. Unlike SimCLR, same-class attraction is deliberately omitted because the cross-entropy losses already pull classes together and because attracting positive pairs from vastly different domains is reported to hurt generalization. The paper presents t-SNE evidence that domain clusters disappear when FixCLR is used, and reports consistent accuracy improvements when FixCLR is combined with other semi-supervised methods, particularly on datasets where domain invariance matters.","pith_inferences":["A natural extension would weight the repulsion by pseudo-label confidence, so low-confidence samples contribute less to $\\mathcal{L}_C$; the paper's own quality-quantity curves suggest this could stabilize early training.","The same repulsion-only principle could transfer to label-scarce domain adaptation and to fully supervised domain generalization, where attracting positive pairs across far-apart domains is known to be risky.","The paper's caution about pretrained models implies that comparisons on ImageNet-similar datasets may overstate methods that rely on pretrained features; a domain-leakage-controlled benchmark would make SSDG comparisons fairer.","The observed pseudo-label quality gain suggests pairing FixCLR with debiasing methods beyond the ones tested, which could matter for class-imbalanced datasets such as FMOW-Wilds."],"forward_implications":["FixCLR can be added to most existing SSDG and semi-supervised methods, with the largest gains appearing on datasets such as Terra Incognita where domain invariance is most needed.","When combined with SoftMatch or StyleMatch, FixCLR achieves the highest accuracy on most datasets in both the 10-label and 5-label settings.","FixCLR improves pseudo-label quality while reducing pseudo-label quantity, because the repulsion regularizer lowers overconfidence and keeps only the more confidently classified samples.","Because FixCLR does not require extra forward passes, it trains faster per epoch than StyleMatch and FBC-SA.","Using positive same-class attraction in the contrastive loss consistently reduces performance, supporting the paper's repulsion-only design."],"supporting_citations":[{"why":"FixMatch is the base semi-supervised framework whose objective FixCLR augments with the repulsion loss.","marker":"[31]"},{"why":"SimCLR supplies the contrastive learning formulation and temperature setting that FixCLR adapts by dropping positive attraction.","marker":"[8]"},{"why":"FBC-SA is the main SSDG baseline FixCLR compares against and whose two-domain prototype regularization it aims to improve.","marker":"[11]"},{"why":"StyleMatch is the other SSDG baseline, providing the stochastic-classifier and style-augmentation comparison and the pretrained-network leakage concern.","marker":"[45]"},{"why":"PCL reports that attracting positive pairs from different domains degrades generalization, which the paper uses to justify its repulsion-only design.","marker":"[39]"},{"why":"Negative learning supplies the complementary-label interpretation of FixCLR's repelling-all-other-classes mechanism.","marker":"[18]"}],"fun_headline_variants":["Repel wrong classes, learn domain-invariant features with few labels","Few labels? Repel wrong classes for cross-domain robustness","Contrastive repulsion for domain generalization under label scarcity","FixCLR: repelling wrong classes for domain invariance","Push wrong classes away: better generalization from few labels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the model's own machine-generated labels are accurate enough that repelling every other predicted class across domains will push representations toward class structure rather than amplify early labeling mistakes.","fun_headline_variants_meta":{"raw":{"variants":["Repel wrong classes, learn domain-invariant features with few labels","Few labels? Repel wrong classes for cross-domain robustness","Contrastive repulsion for domain generalization under label scarcity","FixCLR: repelling wrong classes for domain invariance","Push wrong classes away: better generalization from few labels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000685,"raw_usage":{"total_tokens":3092,"prompt_tokens":917,"completion_tokens":2175,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2093}},"tokens_in":533,"tokens_out":2175,"duration_ms":17752,"temperature":1.0,"reasoning_tokens":2093,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:39:23.063561+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FixCLR and FixMatch on the same SSDG datasets while injecting increasing noise into the pseudo-labels used only by $\\mathcal{L}_C$; if target-domain accuracy and domain-cluster removal do not degrade faster with noise than they do for FixMatch, then pseudo-label accuracy is not the mechanism driving the improvement.","supporting_citations":[{"cited_title":"Fixmatch: Simplifying semi-supervised learning with consistency and confidence","cited_arxiv_id":null,"evidence_quote":"FixMatch is the base semi-supervised framework whose objective FixCLR augments with the repulsion loss."},{"cited_title":"Big self-supervised models are strong semi-supervised learners","cited_arxiv_id":null,"evidence_quote":"SimCLR supplies the contrastive learning formulation and temperature setting that FixCLR adapts by dropping positive attraction."},{"cited_title":"Towards Generalizing to Unseen Domains with Few Labels","cited_arxiv_id":"2403.11674","evidence_quote":"FBC-SA is the main SSDG baseline FixCLR compares against and whose two-domain prototype regularization it aims to improve."},{"cited_title":"Semi-supervised domain generalization with stochastic stylematch","cited_arxiv_id":null,"evidence_quote":"StyleMatch is the other SSDG baseline, providing the stochastic-classifier and style-augmentation comparison and the pretrained-network leakage concern."},{"cited_title":"Pcl: Proxy-based contrastive learning for domain generalization","cited_arxiv_id":null,"evidence_quote":"PCL reports that attracting positive pairs from different domains degrades generalization, which the paper uses to justify its repulsion-only design."},{"cited_title":"Nlnl: Negative learning for noisy labels","cited_arxiv_id":null,"evidence_quote":"Negative learning supplies the complementary-label interpretation of FixCLR's repelling-all-other-classes mechanism."}],"review_version":1}