{"id":"f2894a7f-2eef-466a-a7a8-f9103b4af9a4","arxiv_id":"2607.20666","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Joint contrastive and classification training improves worst-case cross-domain QoT estimation and few-shot adaptation on three public optical-network datasets, at the cost of peak performance in easy transfers.","lead":"A machine-learning method that trains a QoT classifier with both contrastive and classification losses at once shows more stable cross-network generalization on three optical-network datasets. The paper reports better worst-case transfer and faster few-shot adaptation than retraining baselines, but its few-shot comparison gives the proposed method a warm-start advantage.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot gains over tabular baselines may be due to class-conditional sampling rather than joint contrastive learning; baselines are trained on imbalanced data without the proposed Data Module.","rationale":"The reader's weakest_assumption focused on the stability of cosine-similarity geometry across domains. That is a theoretical risk but not the most immediate threat. The paper's empirical evidence for the central claim is the zero-shot comparison against standard tabular baselines. The proposed method differs from those baselines not only in the contrastive/joint loss but also in a class-conditional sampling scheme that explicitly rebalances each mini-batch. Because the baselines are trained on imbalanced data, the reported MF1 gains in class-prior-shift directions (D1→D2, D1→D3) could be explained by sampling alone. This is a concrete, testable confound that directly threatens the attribution of the improvement to representation learning. The joint-vs-separate comparison within the proposed framework is not confounded by sampling (assuming Separate uses the same Data Module), so the paper still provides some support for joint optimization. But the broad claim of outperforming standard baselines is not cleanly established. This reinforces the reader's CONDITIONAL verdict rather than changing it; the paper needs an ablation isolating the contrastive loss and balanced baselines before acceptance. Thus the verdict remains CONDITIONAL/UNCHANGED, but with a different and more empirically pressing concern than the one identified by the reader.","tokens_in":9258,"tokens_out":4801,"duration_ms":42092,"concrete_test":"Retrain all tabular baselines (RF, XGB, CatBoost, LR) and the MLP using the same class-balanced sampling scheme as the proposed method (e.g., equal positive/negative mini-batches, or class weights). Also run an ablation of the proposed architecture with γ=0 (only cross-entropy, same balanced sampler). If any of these match or exceed the reported Ours results on D1→D2 (MF1 0.644, PR-AUC 0.656) and D1→D3 (0.564/0.756), the central attribution to contrastive learning is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is an experimental confound in the zero-shot comparison. In Sec. III-B (Data Module), the proposed architecture trains on mini-batches with equal numbers of positive and negative samples, explicitly to counter class imbalance. The tabular baselines (RF, XGB, CatBoost, LR) and the supervised MLP described in Sec. IV-A are trained on the original imbalanced datasets, without this class-conditional sampler or equivalent reweighting. Under a class-prior shift (D1 has 24.2% failures; D2/D3 have ~7%), balanced training alone could raise Macro-F1 substantially by preventing the model from defaulting to the majority class. The paper attributes the gains to the contrastive objective (Sec. IV-B: 'These gains stem from the contrastive objective enforcing class-relative geometry...'), but this has not been isolated from the sampling scheme. Since the baselines do not receive balanced sampling, the reported improvements in worst-case MF1/PR-AUC (e.g., D1→D2: MF1 0.534→0.644) may be wholly or partly an artifact of the Data Module, not of the joint contrastive/classification learning.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a joint contrastive-and-classification training scheme for cross-domain QoT estimation in optical networks. The architecture combines a class-conditional mini-batch sampler, a backbone network trained with Multi-Similarity Loss on cosine similarities, and a projection network trained with binary cross-entropy; the two objectives are optimized jointly via a weighted sum. The authors evaluate on three public datasets (D1, D2, D3) from different topologies and operational modes, covering six zero-shot transfer directions and few-shot target-data injection at 1–4%. They report that the proposed method improves worst-case Macro-F1 and PR-AUC relative to tabular baselines, e.g., D1→D2 MF1 from 0.534 to 0.644 and PR-AUC from 0.558 to 0.656, and that fine-tuning with 1–2% target labels closes most of the cross-domain gap. The paper concludes that joint contrastive/classification learning learns task-relevant representations that are stable across network domains.","tokens_in":9559,"tokens_out":3849,"duration_ms":36395,"significance":"Cross-domain QoT estimation is practically important, and public datasets and open-source code strengthen reproducibility. The paper also targets a meaningful operational objective—improving worst-case, not just average, transfer performance—and compares against several standard baselines. If the reported gains are real and attributable to the proposed learning objective, the method would be a useful contribution to optical-network ML. However, the current experimental design contains a central confound: the proposed method uses a class-conditional balanced sampler, while the baselines are trained on the original imbalanced data. This alone can explain part of the Macro-F1 gains under class-prior shift. The few-shot comparison also gives the proposed method a warm-start advantage over baselines that are retrained from scratch. These issues must be resolved with additional ablations before the main claims can be accepted.","major_comments":[{"comment":"The zero-shot comparison is confounded by the class-conditional sampler. The proposed architecture trains on mini-batches with equal numbers of positive and negative samples, while the tabular baselines (RF, XGB, CatBoost, LR) and the supervised MLP are trained on the original imbalanced data. Since D1 has a 24.2% failure rate and D2/D3 have ~7%, balanced sampling alone can substantially increase Macro-F1 by preventing majority-class collapse. The paper attributes the gains to the contrastive objective ('These gains stem from the contrastive objective...'), but it has not isolated the sampling scheme from the learning objective. Please report: (i) the proposed method without the class-conditional sampler, (ii) the baselines trained with the same class-balanced sampler or with class weights, and (iii) an ablation with reweighted cross-entropy only, to establish which component drives the","section":"§III-B (Data Module) and §IV-B/Table I"},{"comment":"The few-shot comparison is not methodologically fair. For baselines, each injection ratio is obtained by retraining from scratch on the union of the full source dataset and i% target data; for the proposed method, the model is warm-started from a source-trained checkpoint and fine-tuned on only the injected target samples. This gives the proposed method two advantages: a better initialization and a different optimization protocol. The observed monotonic improvement and rapid adaptation may therefore reflect fine-tuning from a strong source model rather than the joint contrastive/classification design. Please compare under identical protocols—e.g., fine-tune a source-pretrained baseline as well, or retrain the proposed method from scratch on the same union—and additionally report results with balanced sampling for all methods.","section":"§IV-C and Fig. 2"},{"comment":"No error bars, confidence intervals, or repeated-run statistics are reported. All cross-dataset comparisons appear to be single-point estimates, so it is unclear whether differences such as Ours vs. XGB in D2→D3 (MF1 0.974 vs. 0.992) or Ours vs. CatB in D3→D1 (MF1 0.674 vs. 0.523) are statistically reliable. Given that the paper's central claim is about robustness and stability, the authors should report mean ± std over multiple random seeds and, where possible, paired significance tests across transfer directions.","section":"Table I and Fig. 2"},{"comment":"The Separate baseline is used to motivate joint optimization, but its behavior is not adequately explained. In several settings (e.g., D3 as source) it reports MF1 near 0.333, which is essentially the value of a constant classifier that never predicts one class. The paper does not state whether Separate uses the same class-conditional sampler, the same backbone, or how the classifier is trained on frozen embeddings. Without these details, and without an explanation for the chance-level performance, the comparison does not cleanly support the claim that joint training is superior to a sequentially trained representation.","section":"§IV-B, Separate vs. Joint"}],"minor_comments":[{"comment":"The table formatting is broken in several cells, e.g., '0.9710.994' and '0.9930.990 0.993' in the D2-source D3-target column. Please fix the spacing so all values are readable.","section":"Table I"},{"comment":"Hyperparameters are not reported: the MS-Loss margin m, weighting exponents α and β, loss weights γ and η, the class-conditional sampler counts M and N, and the backbone/projection network architectures, optimizer, learning rate, and epochs. This information is necessary to reproduce the results and to assess sensitivity.","section":"§IV-A / §III-B"},{"comment":"The statement that the representation is 'largely insensitive to dataset-specific amplifier gain profiles and link-length distributions' is an interpretation, not an experimental finding. No embedding visualization or quantitative invariance analysis is provided. Please soften the claim or add evidence.","section":"§IV-B"},{"comment":"The conclusion's operational suggestion—retraining triggered by a PR-AUC threshold—is not evaluated in the paper. It would be better to present this as future work rather than as a demonstrated property.","section":"§IV-C / Conclusion"},{"comment":"The terms 'Online Transceiver Mode' and 'Predefined Transceiver Mode' are used without definition. A one-sentence explanation of what these operational modes mean would help readers who are not specialists in optical transceivers.","section":"General"},{"comment":"The figure caption should specify the exact transfer directions shown in each panel, the meaning of the curves (which method is which line), and whether the plotted values are averages over multiple runs. Currently the reader must infer this from the text.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid empirical setup in terms of public datasets and code, but the main claims currently rest on experimental comparisons that are not apples-to-apples. The class-imbalance confound and the warm-start vs. retrain-from-scratch asymmetry in the few-shot experiment are the two issues that most need addressing. If the authors can add the requested ablations and repeated-run statistics, the paper could become a valuable contribution; without them, the evidence does not currently support the attribution of the gains to joint contrastive learning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a decent empirical study of joint contrastive + classification training for QoT estimation across three public datasets. The authors report meaningful worst-case MF1/PR-AUC gains on hard transfer directions. But the comparison is confounded: the proposed pipeline uses a class-conditional sampler that balances the mini-batches, while every baseline trains on the imbalanced data as-is. Since D1 has 24.2% failures vs ~7% in D2/D3, balanced sampling alone could explain much of the improvement. The paper nowhere isolates the contrastive objective from the sampling scheme, and explicitly credits the contrastive loss for the gains. That's the load-bearing issue.\n\nWhat's genuinely useful: the paper clearly documents the domain shift across the three topologies, reports six transfer directions, and shows the method doesn't collapse in easy transfers. Public code is promised, and the datasets are external. If the confound were fixed with an ablation (balanced sampling for baselines, or unbalanced training for the proposed method), the result would be worth serious attention.\n\nOther soft spots are minor in comparison but worth noting. No error bars anywhere; all numbers are single runs, so the 0.03–0.11 improvements could be noise. The few-shot comparison gives the proposed method a warm-start fine-tuning advantage while baselines are retrained from scratch, so the rapid-adaptation claim is not an apples-to-apples comparison. Hyperparameters for MS-Loss and the loss weights are not fully specified.\n\nOn the citation pattern: it's honestly situated in the optical-network ML literature, and the self-citation to the Siamese zero-day anomaly detection work is relevant. No problem there.\n\nBottom line: the paper's core evidence does not yet support its central attribution. It's a promising direction and a fair referee could help the authors tighten the experimental design. I'd send it to review, but I would not cite the zero-shot claim as it stands.","headline":"The paper has a real idea, but the zero-shot gains are confounded by balanced sampling, so the central attribution to contrastive learning doesn't yet stand.","tokens_in":10013,"tokens_out":1998,"would_cite":false,"duration_ms":17803,"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":"A model trained with joint contrastive and classification objectives on one optical network keeps detecting failing lightpaths on unseen networks, and fine-tunes to near-target performance with just 1–2% labeled samples.","keywords":["contrastive learning","deep metric learning","multi-similarity loss","cross-domain generalization","quality of transmission estimation","optical networks","transfer learning","few-shot adaptation"],"falsifier":"Train the same architecture with the MS-Loss term removed (γ=0) on D1 and evaluate zero-shot on D2 and D3; if Macro-F1 does not drop, the contrastive objective is not what carries the generalization, and the paper's central claim is unsupported.","tokens_in":9167,"feed_emoji":"🌐","tokens_out":6739,"duration_ms":54615,"temperature":0.7,"pith_summary":"This paper tries to establish that a model trained on one optical network can keep working on a different, unseen network if its feature space is shaped by contrastive and classification objectives at the same time, rather than by classification alone or by a frozen representation followed by a classifier. The proposed method—joint Multi-Similarity Loss plus binary cross-entropy on a projection head—learns embeddings organized by relative QoT similarity, which the authors argue are less sensitive to domain-specific amplifier profiles and link-length distributions. On lightpath quality-of-transmission (QoT) estimation, the method improves zero-shot Macro-F1 on the two hardest cross-topology transfers (D1→D2: 0.534 to 0.644; D1→D3: 0.487 to 0.564) and establishes a worst-case PR-AUC floor of 0.656 across six transfer directions, higher than the best tree-ensemble baseline's 0.542. With 1–2% labeled target lightpaths, fine-tuning the joint model closes most of the cross-domain gap, whereas retraining baselines from scratch can get worse before they get better. If correct, the result gives network operators a deployable zero-shot QoT pre-screening block and a data-efficient adaptation path.","feed_headline":"Joint contrastive learning keeps network AI accurate on new nets","feed_subtitle":"Zero-shot Macro-F1 improves from 0.534 to 0.644 on hard topology shifts; 1–2% target labels close most of the gap.","key_machinery":"The central object is the jointly optimized latent space: a backbone deep network maps lightpath features to embeddings, and two losses act on them. The Multi-Similarity Loss (MS-Loss) is a contrastive loss that mines hard positive and negative pairs inside each mini-batch using multiple pairwise cosine similarities, pulling same-class lightpaths together and pushing different-class lightpaths apart. A projection network applies binary cross-entropy on the same embeddings, keeping the representation aligned with the QoT decision boundary. A class-conditional sampler balances each mini-batch so the contrastive signal is not dominated by healthy lightpaths. The MS-Loss encodes the domain-invar","core_discovery":"Central claim: jointly training a contrastive loss and a classification head in one loop produces a latent representation that stays aligned with the QoT decision boundary when the network changes. The contrastive loss acts only on relative pairwise cosine similarities, so lightpaths with different absolute amplifier gains or link lengths but comparable SNR margin cluster together, and the classifier reads the boundary from that class-relative geometry. On six cross-topology transfers, the hardest shifts (D1→D2, D1→D3) improve Macro-F1 from 0.534 to 0.644 and from 0.487 to 0.564, and worst-case PR-AUC is 0.656 versus 0.542 for the best tree-ensemble baseline. Warm-start fine-tuning with 1–2%","pith_inferences":["If pairwise cosine-similarity geometry is the true carry-over, the same joint contrastive-plus-classification objective should transfer to other optical-network tasks with a natural similarity notion, such as OSNR regression or modulation-format identification; this extension is not tested in the paper.","The observed non-monotonicity of retrain-from-scratch baselines under small target injections suggests that naively pooling source and target data can hurt early adaptation; the paper's warm-start fine-tuning is a safer default, but this deserves a direct comparison on more dataset pairs.","A testable refinement would be to measure how the claimed invariance degrades as class-prior shift increases; the paper's hardest cases combine prior shift with configuration changes, so isolating the two would clarify whether the robustness comes from the contrastive geometry or from the class-conditional sampler."],"forward_implications":["Zero-shot deployment: a model trained only on one network's labeled data can be dropped into an unseen network and still detect failing lightpaths at a usable rate, without any target labels.","Worst-case floor: across six transfer directions the method's PR-AUC never falls below 0.656, whereas the best tree-ensemble baseline drops to 0.542, so the method protects against catastrophic generalization failures.","Rapid adaptation: with only 1–2% labeled target lightpaths, Macro-F1 closes most of the cross-domain gap (e.g., D1→D2 rises from 0.644 to roughly 0.94 at 2% injection).","Joint optimization beats the sequential pipeline: freezing the representation first and then training a classifier collapses under shift (MF1 0.33–0.35 in some transfers), while joint training stays robust.","The resulting model can serve as a QoT pre-screening block fed by telemetry, with retraining triggered only when injected-sample PR-AUC drops below a deployment-defined threshold."],"fun_headline_variants":["Joint contrastive learning lifts cross-domain optical AI by 20%","Unseen networks: joint contrastive learning keeps AI accurate","Optical AI generalizes across topologies with joint contrastive loss","1-2% retraining closes domain gap for optical network AI","Cross-domain optical AI: contrastive coupling beats domain shift"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the relative pairwise cosine-similarity structure learned on the source network remains meaningful in the target network—lightpaths with comparable QoT margin cluster together regardless of topology, amplifier gains, or class balance—so target decision boundaries coincide with source clusters.","fun_headline_variants_meta":{"raw":{"variants":["Joint contrastive learning lifts cross-domain optical AI by 20%","Unseen networks: joint contrastive learning keeps AI accurate","Optical AI generalizes across topologies with joint contrastive loss","1-2% retraining closes domain gap for optical network AI","Cross-domain optical AI: contrastive coupling beats domain shift"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000556,"raw_usage":{"total_tokens":2448,"prompt_tokens":676,"completion_tokens":1772,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":420,"completion_tokens_details":{"reasoning_tokens":1699}},"tokens_in":420,"tokens_out":1772,"duration_ms":11471,"temperature":1.0,"reasoning_tokens":1699,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:41:21.397860+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture with the MS-Loss term removed (γ=0) on D1 and evaluate zero-shot on D2 and D3; if Macro-F1 does not drop, the contrastive objective is not what carries the generalization, and the paper's central claim is unsupported.","supporting_citations":[],"review_version":1}