{"id":"bafaf31e-44e8-4cf6-94e6-ab210d93424b","arxiv_id":"2608.10080","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A supervised weighted contrastive learning pipeline, zlbg, improves contaminant rejection in DESI Lyman Break Galaxy spectra while matching the redshift accuracy of the existing lbgNET pipeline.","lead":"This paper trains a neural network to sort faint high-redshift galaxies from quasars and low-redshift impostors in DESI spectra, and to estimate galaxy distances, using a machine learning trick that pulls similar spectra together in a learned space. The new pipeline, called zlbg, removes contaminants from a Lyman Break Galaxy sample better than the current DESI pipeline while matching its redshift accuracy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Redshift parity rests on unvalidated template-based z-augmentation: without it zlbg drops to AUC 0.737 vs 0.789 for lbgNET, so the z>3.8 claim needs an independent high-z test.","rationale":"The reader's weakest assumption points to the same mechanism, and Figure 8 gives it quantitative force: removing z-augmentation changes the redshift comparison from roughly even to a 0.05 AUC deficit for zlbg. The paper is transparent about the small test sample and about hyperparameter tuning, and Appendix C shows sensitivity to the chosen prior width, so the correct response is not rejection. The classification improvement stands. The condition should be an independent high-z validation before the redshift-comparability claim is used for DESI Run 2 science. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":29080,"tokens_out":7984,"duration_ms":80935,"concrete_test":"Hold out every visually inspected LBG with z>3.8 and recompute the redshift purity/efficiency AUC for zlbg and lbgNET; if fewer than about 50 such spectra exist, build an independent mock test set by taking low-z VI spectra, shifting them to z in [3.8, 4.5], completing them with a template library not derived from lbgNET classifications (for example BPASS or CIGALE composite spectra), adding realistic DESI noise and sky residuals, and re-running both fiducial pipelines. If the high-z AUC difference matches the no-z-aug gap (about 0.05) rather than the full-sample gap (about 0.01), the redshift-comparability claim fails in the target regime; if it stays within the reported 2% seed scatter, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The contaminant-classification claim (AUC 0.997 vs 0.988) is well supported and nearly independent of redshift augmentation. The load-bearing weakness is the redshift half of the central claim. In Figure 8, with the same training set but no redshift augmentation, zlbg's redshift AUC is 0.737 versus 0.789 for lbgNET; with the fiducial z-augmentation it is 0.790 versus 0.799. The 'comparable redshift' result therefore depends on the Section 4.3 augmentation, which completes spectra shifted beyond their observed coverage with four templates (Section 4.1) built from VI stacks plus lbgNET classifications at tau=0.99. For z > 3.8, where VI coverage is explicitly sparse, the blue side of an augmented spectrum is mostly template rather than data, so the embedding and KNN redshifts in that regime are learned from, and can inherit systematics from, those templates. The test set has little independent coverage there, so the reported AUCs do not validate the z=3.8-4.5 regime that DESI Run 2 needs. This does not undermine the classification result, but it is load-bearing for the claim that zlbg can replace or complement lbgNET with equal redshift completeness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes zlbg, a supervised weighted contrastive-learning pipeline for classifying DESI Lyman Break Galaxy (LBG) spectra and estimating their redshifts. The encoder is the same backbone as QuasarNET/lbgNET, but the line-finder head is replaced by a projection head trained with a continuously weighted contrastive loss that encodes both galaxy class similarity and redshift proximity. After training, a small MLP provides per-class probabilities and a K-nearest-neighbors regressor provides redshifts, with a redshift-quality flag q_z used to form the final confidence threshold. Training data come from DESI visual-inspection (VI) campaigns, supplemented by SNR and redshift augmentations; the redshift augmentation splices LBG templates into observed spectra to extend coverage to z=4.5. On a held-out test set of 781 spectra, zlbg achieves contaminant-selection AUC 0.997 versus 0.988 for lbgNET and redshift AUC 0.790 versus 0.799 under the fiducial ℓ_z=0.025(1+z) criterion. Appendices explore kernel choices, intra-class weighting, prior-width sensitivity, and treatment of bad spectra.","tokens_in":29360,"tokens_out":5721,"duration_ms":54463,"significance":"If the results hold, the paper offers a genuinely useful alternative to lbgNET for DESI Run 2 LBG processing: stronger contaminant rejection at comparable redshift completeness would improve the purity of the LBG sample without sacrificing redshift yield. The work has clear strengths: the code is publicly released, lbgNET is retrained on the same training set for a fair architecture comparison, the paper performs controlled-seed runs, and the appendices document sensitivity to kernel shape, class weights, prior width, and bad spectra. The main caveat is that the redshift-comparability claim is not yet established independently of the template-based redshift augmentation, and the test set is small enough that statistical uncertainties on the headline AUC differences should be quantified.","major_comments":[{"comment":"The headline result of \"comparable redshift identification\" is load-bearing on the redshift augmentation described in Section 4.3. In Figure 8, the 'Base (no z-aug)' configuration gives zlbg a redshift AUC of 0.737 versus 0.789 for lbgNET, while the fiducial 'Base' configuration gives 0.790 versus 0.799. Since the templates used for augmentation are built from VI stacks plus lbgNET classifications at tau=0.99 (Section 4.1), and since the VI sample has sparse coverage at z>3.8, the high-redshift portion of the augmented training set is largely template rather than observed data. The reported test AUCs therefore do not independently validate the z~3.8-4.5 regime that DESI Run 2 requires. Please add an independent high-z validation set (for example, published LBG spectra or DESI pilot data not used in template construction) or report redshift-binned purity/efficiency and redshift accuracy specifically for z>3.8.","section":"Section 5.3, Figure 8, and Section 4.3"},{"comment":"The redshift kernel width sigma_z=0.025 is set equal to the evaluation success criterion ell_z=0.025(1+z), and the same ell_z enters the quality flag q_z in Equation (3.8). This couples the training objective, the confidence threshold, and the success metric to the same scale. It is not full circularity because evaluation is on held-out spectra, but it weakens the force of the \"comparable redshift\" claim: the comparison is made under a metric matched to the training kernel. Please report a decoupled evaluation (for example, KNN point-estimate nMAD and 3nMAD/5nMAD outlier fractions as in Figure 13 but applied directly to the KNN redshifts) and, ideally, retrain with several sigma_z values from Figure 10 and evaluate at a fixed, independently chosen ell_z.","section":"Section 3.3, Section 5.2, and Appendix C"},{"comment":"The test set is small (781 spectra, including only 72 ELGs and 63 QSOs), and the reported AUC differences are not accompanied by confidence intervals. The text states that seed-to-seed variation is about 0.2% for contaminant selection and 2% for redshift performance, but this does not characterize the sampling uncertainty of the test set itself. Please provide bootstrap or DeLong confidence intervals for the headline AUC values (0.997 vs 0.988 for classification, 0.790 vs 0.799 for redshift) and for the four configurations in Figure 8, so the reader can judge whether the improvements and the augmentation dependence are statistically significant.","section":"Section 5.1, Section 5.2, and Figure 8"},{"comment":"The conclusion that zlbg requires redshift augmentation to match lbgNET is not consistent across training configurations: in 'Base+Val (no z-aug)' zlbg actually exceeds lbgNET (0.778 vs 0.759), whereas in 'Base (no z-aug)' it is lower (0.737 vs 0.789). This non-monotonic pattern suggests the comparison is noisy with a single seed and a small test set. Please report multiple seeds and error bars before drawing a firm conclusion about which training configurations are required for parity.","section":"Section 5.3, Figure 8"}],"minor_comments":[{"comment":"The label 'Flatter' in Figure 2 should read 'Flatten', and the text uses 'Replayer' where 'representation layer' or 'Rep layer' would be clearer for readers outside the DESI pipeline.","section":"Figure 2 and Section 3.1"},{"comment":"There are several typographical errors, including 'spectras', 'redshiftidentification', and 'redshiftpredictedredshift'; these should be corrected in a revision.","section":"Section 2"},{"comment":"The caption of Figure 3 says that n_aug=5 is 'the ratio between the number of redshift augmentations and SNR augmentations,' but Section 4.3 defines n_aug=5 as the total number of spectra including the original spectrum; please reconcile these statements.","section":"Figure 3 and Section 4.3"},{"comment":"The Pearson correlation coefficients restricted to QSO and ELG subsets are computed over 63 and 72 objects, respectively; the low r values may reflect small-sample noise rather than genuine information complementarity, so the joint-information claim should be phrased more cautiously.","section":"Section 5.4"},{"comment":"The sentence 'While the MLP performs well on contaminants, it performs poorly at identifying the correct subtype of LBG' is followed by 'the MLP offers sufficient classification performance'; please rephrase to avoid the apparent contradiction.","section":"Section 3.4"},{"comment":"In the sentence 'redrock is ran on the co-added spectra,' the grammar should be corrected, and it would be helpful to state explicitly that the same redrock templates from Section 4.1 are used for both pipelines so that template choice cannot bias the comparison.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for JCAP and the code-release and ablation structure are commendable. My main concern is that the redshift-comparability claim is currently validated only through the same template-based augmentation scheme that is used to train the network, and the test set is too small for the headline AUC differences to be sharply assessed. I would recommend sending the revision to a referee with experience in spectroscopic redshift pipelines or in statistical evaluation of ML classifiers, in addition to the cosmology audience."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. First, the contaminant-selection result is the real result: zlbg gets AUC 0.997 vs 0.988 for lbgNET, and that holds across training configurations, including when redshift augmentation is removed. Second, the redshift-parity claim is conditional on a template-based redshift augmentation. Remove the augmentation and zlbg's redshift AUC falls to 0.737 vs 0.789 for lbgNET; with it, they are 0.790 vs 0.799. So the headline \"comparable redshift\" rests on spectra that are partly synthetic at z > 3.8, and that regime has no independent validation in the test set.\n\nWhat is new: a supervised contrastive loss with continuous relationship weights (a CWCL-style generalisation) applied to LBG spectra, with a class-similarity matrix and a Gaussian redshift kernel. That combination is not in the cited literature. They also re-train lbgNET on the same data for a fair comparison, which is rare and commendable. Code and data are released on GitHub/Zenodo, and the paper is clearly written, with honest stress tests in the appendices.\n\nSoft spots, in order of seriousness. The test set is small (~780 LBG spectra) and the quoted AUCs come from single runs with seed-scatter quoted at ~0.2% for classification and ~2% for redshift. That's enough to see the classification gap is real, but the redshift comparison is within that scatter. Second, the Gaussian kernel width is set to 0.025, the same value as the success criterion \\ell_z = 0.025(1+z) for purity/efficiency. That is not full circularity—evaluation is on held-out spectra—but it means the training objective and the evaluation metric share the same scale, which flatters the comparison. Third, the z-augmentation templates are built from VI stacks plus lbgNET high-confidence classifications at \\tau=0.99. At z>3.8, the blue side of an augmented spectrum is mostly template, not data; the embedding and KNN redshifts there can inherit template systematics. The paper flags the need for larger samples, but it doesn't provide an independent high-z check.\n\nThis doesn't sink the paper. The classification contribution stands on its own, and the redshift work is a reasonable proof-of-concept. Who is it for? Anyone planning DESI Run 2 LBG processing, and people working on contrastive learning for spectra. It deserves a serious referee. I would send it to review with the expectation of major revisions: add an out-of-sample or simulated high-z validation, quote bootstrap error bars on the AUCs, and discuss the kernel/evaluation alignment directly. The authors have already done the honest part; they need to close the validation gap.","headline":"Classification win is real; redshift parity is conditional on template-based augmentation that lacks independent high-z validation.","tokens_in":30225,"tokens_out":3142,"would_cite":true,"duration_ms":28955,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Supervised weighted contrastive learning separates DESI Lyman-break galaxies from contaminants better than the current line-detector network, at equal redshift accuracy.","keywords":["Lyman break galaxies","contrastive learning","redshift estimation","spectral classification","DESI","deep learning","galaxy contaminants","K-nearest neighbors"],"falsifier":"Take newly observed DESI survey-validation LBG spectra with secure visual-inspection redshifts at z > 3.8, run zlbg's KNN redshift and classification head, and compare purity and outlier fraction against the z < 3.8 test set; if high-redshift objects show a sharp drop in redshift accuracy or a rise in contaminant misclassification, the template-splicing augmentation is not representative of real spectra at the redshifts it was designed to cover.","tokens_in":28893,"feed_emoji":"🔭","tokens_out":6975,"duration_ms":60576,"temperature":0.7,"pith_summary":"This paper tries to establish that supervised weighted contrastive learning can handle both tasks needed for a faint high-redshift galaxy sample: separating Lyman break galaxies (LBGs) from quasars and low-redshift emission line galaxies, and estimating redshifts, using a small, visually inspected training set. The proposed zlbg pipeline is tested on DESI pilot spectra and reports stronger contaminant classification than the current lbgNET line-detector network (AUC 0.997 versus 0.988) with comparable redshift identification (AUC 0.790 versus 0.799). If this result holds, DESI Run 2 could obtain a cleaner LBG sample at equal redshift completeness, which matters for primordial non-Gaussianity measurements, Ly-alpha forest tomography, and void catalogs that rely on high-redshift tracers.","feed_headline":"Contrastive learning cleans DESI's high-redshift galaxy sample","feed_subtitle":"New zlbg pipeline flags quasars and emission-line galaxies while matching redshift accuracy for Run 2.","key_machinery":"The load-bearing mechanism is a supervised weighted contrastive loss. For each batch of spectra, a relationship coefficient multiplies the contrastive log-softmax: a class-similarity matrix sets pairwise weights among the five classes (LBGa, LBGme, LBGse, ELG, QSO), and for LBG pairs a Gaussian kernel in redshift, w(z_a,z_i) = exp(-(z_a-z_i)^2 / (2 $sigma_z^{2}$ (1+mean z)^2)) with sigma_z = 0.025, sharpens the embedding by redshift proximity. The encoder is the same Conv1D backbone as lbgNET; the projection head is discarded after training, and downstream tasks use the representation layer: a small MLP for classification and a 50-neighbor KNN on cosine distance for redshift, with a quality flag q_z derived from neighborhood scatter. Data augmentation splices the four LBG stacked templates into observed spectra to generate redshift coverage up to z = 4.5 and SNR variation via different exposure co-adds.","core_discovery":"The central claim is that replacing the line-finder readout of the DESI LBG network with a contrastively trained embedding changes what the network learns about spectra: instead of reporting confidence in individual spectroscopic lines, it organizes spectra by galaxy type and by redshift, so contaminants separate cleanly while redshift proximity is encoded continuously. On the same visually inspected test set, zlbg reaches AUC 0.997 for contaminant selection versus 0.988 for lbgNET, and AUC 0.790 versus 0.799 for redshift identification, a difference the paper reads as comparable. The redshift information is meant to be consumed as a prior for redrock template fitting, and the paper shows that broadening the prior to a half-width of 0.035(1+z) recovers more sources with zlbg than with lbgNET, evidence that the embedding is less prone to catastrophic line misidentification.","pith_inferences":["If the template-splicing augmentation is faithful at z > 3.8, the same contrastive recipe should transfer to other faint high-redshift populations where visual inspection coverage is sparse, such as DESI's Ly-alpha emitter targets.","The learned representation could serve as a reusable prior for other LBG science beyond redrock fitting, such as Ly-alpha forest correlations or void catalogs, since the embedding appears to carry a continuous redshift gradient.","A direct testable extension is to retrain with separate high-redshift template realizations rather than one randomly chosen LBGme template, which may reduce the LBGme subtype confusion seen in the confusion matrix."],"forward_implications":["Switching DESI LBG processing to zlbg, or combining it with lbgNET, would raise LBG sample purity while keeping redshift completeness, since the two confidence scores are weakly correlated.","Because zlbg outputs per-class probabilities rather than a single line confidence, it adds information about LBG subtype (LBGa, LBGme, LBGse) and about the nature of contaminants, which is useful for the separate LAE program.","The redshift-aligned embedding should be less sensitive to catastrophic outliers caused by misidentified emission lines, and the paper shows performance improves relative to lbgNET as the redrock prior width is increased to 0.035(1+z).","The same architecture can be retrained on as-yet-unobserved DESI Run 2 survey validation data; the paper explicitly leaves larger datasets for future work."],"supporting_citations":[{"why":"Defines the lbgNET baseline, the VI-0 visual inspection dataset, and the four LBG templates whose construction zlbg's redshift augmentation reuses.","marker":"[32]"},{"why":"Provides the original QuasarNET architecture whose Conv1D backbone both lbgNET and zlbg share, and the line-finder readout that zlbg replaces.","marker":"[34]"},{"why":"Introduces the supervised contrastive loss that the paper generalizes to continuous relationship weights.","marker":"[39]"},{"why":"Supplies the continuously weighted contrastive loss formulation behind the soft relationship matrix.","marker":"[55]"},{"why":"Establishes the contrastive learning framework and the practice of discarding the projection head after training.","marker":"[38]"},{"why":"Provides the redrock template-fitting code that consumes zlbg's redshift priors in the end-to-end pipeline.","marker":"[35]"},{"why":"Documents DESI quasar target selection and the visual inspection procedure used as ground truth for testing.","marker":"[16]"},{"why":"Describes the color-dropout target selection whose redshift upper limit (z ~ 4.5) and contaminants motivate the decontamination task.","marker":"[33]"}],"fun_headline_variants":["Contrastive learning cuts quasar contamination in DESI LBG sample","Contrastive embedding cleans DESI LBG sample, keeps redshift accuracy","zlbg: contrastive learning rejects impostors in DESI's LBG survey","Contrastive learning improves LBG decontamination for DESI Run 2","Better contaminant rejection with contrastive learning in DESI LBG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach rests on the assumption that redshift-augmented training spectra, built by splicing stacked LBG templates into observed spectra to reach z up to 4.5, faithfully represent real LBG spectra at redshifts where visual inspection has little coverage (above about z = 3.8), and that the templates themselves, derived from visually inspected spectra plus lbgNET's high-confidence classifications, carry no systematic error that the contrastive loss will learn.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive learning cuts quasar contamination in DESI LBG sample","Contrastive embedding cleans DESI LBG sample, keeps redshift accuracy","zlbg: contrastive learning rejects impostors in DESI's LBG survey","Contrastive learning improves LBG decontamination for DESI Run 2","Better contaminant rejection with contrastive learning in DESI LBG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000895,"raw_usage":{"total_tokens":3845,"prompt_tokens":923,"completion_tokens":2922,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":2822}},"tokens_in":539,"tokens_out":2922,"duration_ms":19103,"temperature":1.0,"reasoning_tokens":2822,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:14:11.284490+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take newly observed DESI survey-validation LBG spectra with secure visual-inspection redshifts at z > 3.8, run zlbg's KNN redshift and classification head, and compare purity and outlier fraction against the z < 3.8 test set; if high-redshift objects show a sharp drop in redshift accuracy or a rise in contaminant misclassification, the template-splicing augmentation is not representative of real spectra at the redshifts it was designed to cover.","supporting_citations":[{"cited_title":"Ruhlmann-Kleider, C","cited_arxiv_id":null,"evidence_quote":"Defines the lbgNET baseline, the VI-0 visual inspection dataset, and the four LBG templates whose construction zlbg's redshift augmentation reuses."},{"cited_title":"Busca and C","cited_arxiv_id":null,"evidence_quote":"Provides the original QuasarNET architecture whose Conv1D backbone both lbgNET and zlbg share, and the line-finder readout that zlbg replaces."},{"cited_title":"Srinivasa, J","cited_arxiv_id":null,"evidence_quote":"Supplies the continuously weighted contrastive loss formulation behind the soft relationship matrix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the contrastive learning framework and the practice of discarding the projection head after training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the redrock template-fitting code that consumes zlbg's redshift priors in the end-to-end pipeline."}],"review_version":1}