Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

The paper argues that a simple MLP pre-trained on LAMOST low-resolution spectra transfers to DESI’s medium-resolution spectra so effectively that it beats the DESI pipeline even with zero shot transfer, and that modest fine-tuning improves

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pre-trained MLPs on LAMOST low-resolution spectra generalize to DESI medium-resolution spectra for [Fe/H] and [α/Fe], outperforming the DESI SP pipeline in zero-shot and improving with modest fine-tuning.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Useful cross-survey benchmark, but the zero-shot claim needs a source-disjointness check before it can be trusted. the 4 major comments →

arxiv 2602.15021 v2 pith:DFO7FX5V submitted 2026-02-16 astro-ph.SR astro-ph.GAcs.LG

Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI

Xiaosheng Zhao , Yuan-Sen Ting , Rosemary F.G. Wyse , Alexander S. Szalay , Yang Huang , L\'aszl\'o Dobos , Tam\'as Budav\'ari , Viska Wei This is my paper
classification astro-ph.SR astro-ph.GAcs.LG
keywords cross-survey generalizationstellar parameter estimationmultilayer perceptrontransfer learningspectral foundation modelsLAMOSTDESIfine-tuning strategies
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks whether simple pretrained neural networks can carry stellar labels across surveys with different resolution and wavelength coverage. It pre-trains multilayer perceptrons on roughly 90,000 LAMOST low-resolution spectra with APOGEE-based labels and applies them to DESI medium-resolution spectra, with and without fine-tuning. The central claim is that these MLPs generalize well even with zero fine-tuning, agreeing with APOGEE reference labels better than the DESI production pipeline or models trained from scratch, and that a few hundred to roughly 2,000 DESI spectra are enough to correct remaining biases. It also claims that transformer-based spectral embeddings help only for iron abundance in the metal-rich regime, underperform for metal-poor stars, and do not help alpha abundances, so simple MLPs remain competitive. If this is right, lightweight pretrained models are a practical baseline for upcoming spectroscopic programs, and the value of larger spectral foundation models for cross-survey transfer should not be assumed.

Core claim

On its own terms, the discovery is that cross-survey stellar label transfer does not require a large foundation model: an MLP pre-trained on LAMOST low-resolution spectra, applied to DESI with no fine-tuning, agrees with APOGEE labels better than the DESI SP pipeline or an MLP trained from scratch on DESI. Fine-tuning with roughly 2,000 DESI spectra (as few as 100–500 for some metrics) further improves accuracy and corrects the zero-shot metal-rich [Fe/H] bias. Transformer-based spectral embeddings help only in the metal-rich [Fe/H] regime, hurt in the metal-poor regime (79 test stars), and do not help [α/Fe]. The best fine-tuning strategy depends on the parameter: residual-head adaptation f

What carries the argument

The load-bearing object is a pre-trained multilayer perceptron that maps 1,462 continuum-normalized flux bins (400–560 nm) to [Fe/H] or [α/Fe], trained on LAMOST low-resolution spectra with APOGEE-based labels. For comparison, the same MLP is trained on 768-dimensional embeddings from SpecCLIP, a self-supervised transformer with contrastive alignment between LAMOST and Gaia XP spectra. Three adaptation mechanisms are compared: residual-head adapters (a small parallel MLP correcting the output), LoRA (low-rank weight updates), and full fine-tuning. The MLP is what carries the transfer; the embeddings and fine-tuning variants are the comparison arms.

Load-bearing premise

The regime-specific conclusions—that embeddings underperform for metal-poor stars and that residual-head fine-tuning is best for [Fe/H]—rest on just 79 metal-poor test stars, and noisy APOGEE labels there could reverse those differences.

What would settle it

Cross-match a larger DESI sample (a few hundred stars with [Fe/H] < -1.0) to APOGEE, recompute the Table 1 R2 values for zero-shot and fine-tuned models, and check whether the MLP trained directly on spectra still beats the embedding-based MLPs in that regime; if not, the paper's central regime-specific claim fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Cross-survey stellar parameter pipelines can be deployed quickly: pre-train a simple MLP on an existing low-resolution sample with high-quality labels, then apply it to a new medium-resolution survey immediately without retraining.
  • A few hundred to roughly 2,000 labeled spectra from the target survey are enough to correct resolution-driven biases; larger samples mainly tighten scatter.
  • Foundation-model embeddings are not automatically better for transfer: their advantage is limited to metal-rich [Fe/H], and they can degrade metal-poor estimates.
  • Fine-tuning strategy should be chosen per stellar parameter: residual heads preserve the metal-poor regime for [Fe/H], while LoRA best adapts [α/Fe].
  • Physical structure such as the thin–thick disk separation in the [α/Fe]–[Fe/H] diagram can be recovered by a transferred model even when the target survey's own pipeline does not show it.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the zero-shot MLP result holds up, it gives a strong null hypothesis for future claims about spectral foundation models: a cheap MLP trained on the same labels should be the baseline before attributing gains to transformer representations.
  • The 79-star metal-poor subset is too small to settle whether embeddings are genuinely worse there; a dedicated sample of a few hundred metal-poor DESI stars would test this directly.
  • Because fine-tuning shifts which population the model fits best, a natural extension is to split metal-poor and metal-rich stars and apply separate adapters to each, a testable refinement of the residual-head versus LoRA finding.
  • The same pretrained LAMOST MLP should transfer to other medium-resolution surveys with similar wavelength coverage; if it does, resolution differences matter less than label quality and continuum normalization.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper investigates cross-survey generalization from LAMOST low-resolution spectra (R≈1800) to DESI medium-resolution spectra (R≈2000–3200) using pre-trained multilayer perceptrons (MLPs). The authors pre-train MLPs on LAMOST LRS (90,106 spectra for [Fe/H], 85,400 for [α/Fe]) with labels drawn primarily from APOGEE and supplemented by metal-poor compilations, and also train MLPs on SpecCLIP transformer embeddings. They apply these models zero-shot to 5,539 DESI DR1/APOGEE cross-matched stars and fine-tune them with 2,069 DESI spectra using residual-head, LoRA, and full fine-tuning. The main claims are: (1) zero-shot MLP-LRS agrees with APOGEE better than the DESI SP pipeline and from-scratch models; (2) modest fine-tuning further improves accuracy; (3) SpecCLIP embeddings help [Fe/H] only in the metal-rich regime and underperform in the metal-poor regime; and (4) the optimal fine-tuning strategy depends on the stellar parameter. Supporting evidence includes R² and robust σ statistics, MMD comparisons of [α/Fe]–[Fe/H] diagrams, five-seed averages, ablations, loss-landscape visualizations, and a saliency analysis.

Significance. If the central claims survive a source-disjointness check, the paper would establish that lightweight pre-trained MLPs are a strong, practical baseline for cross-survey stellar parameter estimation, with implications for rapid deployment in DESI DR2, SDSS-V, and PFS. The study has clear strengths: evaluation against external APOGEE labels, five-seed averaging, systematic ablations, a clean calibrated DESI SP comparison in Appendix D, and physically interpretable saliency checks. However, the central zero-shot claim is currently unverifiable because the manuscript never states that DESI test stars were excluded from the LAMOST pre-training sample. In addition, several headline comparisons—especially those involving the 79-star metal-poor subset and the MMD differences—are reported without uncertainties. The paper is a useful empirical contribution if these issues are resolved, but the main conclusions are not yet fully established.

major comments (4)
  1. [§4, §5.1, Table 1] The zero-shot claim in the abstract and §5.1 requires that DESI test stars were not seen during LAMOST pre-training. Section 4 describes pre-training on 90,106 LAMOST LRS spectra and testing on 5,539 DESI DR1 stars cross-matched to APOGEE via Gaia DR3 source_id, but it never states that the DESI test set is source-disjoint from the LAMOST training set. Given the heavy sky overlap between DESI and LAMOST, many of these stars plausibly have LAMOST LRS spectra in the pre-training sample. If so, the MLP could have memorized APOGEE labels for those stars from their LAMOST spectra, inflating the reported zero-shot R² values in Table 1 and the MMD improvement in Figure 3, and also contaminating the fine-tuning experiments. This is load-bearing for the paper's central claim. Please either explicitly confirm source disjointness or rerun the analysis after removing all overlapping Gaia source_ids
  2. [Table 1, §6.1, §6.2] The nuanced conclusions—that embedding-based models underperform in the metal-poor regime and that residual-head fine-tuning is best for [Fe/H]—rest heavily on the metal-poor test subset of only 79 stars. Table 1 reports R² values without uncertainties; for example, zero-shot [Fe/H] on the metal-poor subset is 0.678 for lrs versus 0.694 for embed, and after fine-tuning it is 0.636 for residual versus 0.571 for LoRA. These differences are plausibly within sampling noise. The authors acknowledge the small sample in §6.1, but the abstract still states the metal-poor underperformance as a conclusion. Please provide bootstrap confidence intervals or significance tests for these comparisons, and qualify the claims accordingly if the differences are not significant.
  3. [§5.2, §6.2, Table 1] The choice of fine-tuning strategy appears to be made after inspecting the test-set results. Section 5.2 states that LoRA is used for [α/Fe] and residual-head for [Fe/H], and §6.2 explains that LoRA was adopted because it 'achieves slightly better performance across experiments.' Given the large number of strategy/subset combinations in Table 1 and the modest differences among several entries, this is post-hoc selection on the test set and the reported advantage may reflect noise. Please either (a) select strategies on a held-out validation set before evaluating on the test set, or (b) clearly disclose the selection procedure, report confidence intervals, and avoid presenting the selected strategy as genuinely superior without a significance test.
  4. [Figure 3, Eq. (1)] The MMD values in Figure 3 are point estimates with a fixed RBF bandwidth γ=2.0 and no uncertainty quantification. The central comparison is between MMD=0.362 (MLP-LRS zero-shot), MMD=0.356 (fine-tuned), and MMD=0.452 (MLP-Scratch), but without bootstrap confidence intervals it is unclear whether these differences are significant. Since the MMD improvement is used to support the claim that pre-trained MLPs reproduce the APOGEE [α/Fe]–[Fe/H] distribution better than the alternatives, please provide bootstrap error bars or another significance assessment for the MMD values.
minor comments (5)
  1. [§1] Duplicate wording: 'is how to to generalize' should be 'is how to generalize.' In several places 'R VS' appears with odd spacing; presumably 'RVS.'
  2. [§2] The ReLU activation is cited to Blanco-Cuaresma (2019), which is not the standard reference for the ReLU activation function; please cite the original source or a standard ML reference.
  3. [Figure 4 / Figure A1/A2] The captions note that residual-head fine-tuning is used for [Fe/H] and LoRA for [α/Fe], but these choices are justified only later in §6.2. Please add a pointer to the strategy-selection discussion so the reader understands this is not an arbitrary choice.
  4. [Table 1, Tables A1/A2] The captions say the top three entries are 'highlighted with decreasing shades of blue,' but in a monochrome print or plain-text rendering this information is lost. Consider using boldface or explicit markers.
  5. [§6.4] The limitations section does not mention source disjointness between the LAMOST pre-training set and the DESI test set, nor does it state whether the MLP training code is publicly available. Both are relevant for reproducibility and for interpreting the zero-shot result.

Circularity Check

0 steps flagged

No circular derivation: transfer claims are benchmarked against external APOGEE labels; self-citations supply model components, not conclusions.

full rationale

The derivation chain is empirical rather than definitional: published MLPs and SpecCLIP embeddings from Zhao et al. (2025b) are applied to DESI DR1/EDR spectra, and the resulting [Fe/H] and [alpha/Fe] estimates are compared with APOGEE DR17 labels for DESI cross-matched stars, with fine-tuning performed on a separate 2,069-star DESI subset and testing on a disjoint 5,539-star sample. No equation in the paper defines a predicted quantity in terms of the target metric; R^2 and MMD are computed against external APOGEE labels, and the APOGEE labels used for pre-training are the standard supervised training labels, not a fitted parameter renamed as a prediction. The self-citations (adoption of the pre-trained MLPs and SpecCLIP from Zhao et al. 2025b, and LoRA behavior from Zhao et al. 2025a) are component sources or corroborating context, not logical premises that force the reported transfer results; the central claims stand or fall on the independent DESI/APOGEE benchmark. Section 6.4 lists limitations such as blue-only wavelength coverage, model scale, restricted parameter set, and small fine-tuning sample sizes, and Section 6.1 notes the metal-poor comparison may be limited by only 79 test stars; these are robustness concerns, not circular reductions. One additional concern is that Section 4 does not explicitly state that DESI test stars were excluded from the LAMOST pre-training set, and Section 6.4 does not mention this source-disjointness requirement; that is a possible data-leakage/correctness risk for the zero-shot claim, but it is not a circular step because even with overlap the outputs are still compared with external APOGEE labels rather than derived from the inputs by construction. Overall, I find no circularity and assign the lowest score.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central empirical claims rely on the reliability of APOGEE labels and the comparability of re-gridded, continuum-normalized DESI spectra to LAMOST LRS inputs. No new physical entity is introduced; the main hand-chosen inputs are ML hyperparameters and the metallicity threshold used to partition results.

free parameters (6)
  • LoRA rank r and scaling alpha = r=128, alpha=256
    Chosen by hand (Section 3.1); affects fine-tuning capacity and all LoRA results. No sensitivity analysis is reported.
  • Residual-head hidden size = 384
    Chosen so the residual head has ~0.6M parameters, comparable to LoRA (Sections 3.2 and A.2); affects [Fe/H] metal-poor performance.
  • MMD kernel bandwidth gamma = 2.0
    Arbitrary choice in Eq. 1; MMD rankings could change with bandwidth, and no sensitivity analysis is provided (Section 5.1).
  • Metallicity split threshold = [Fe/H] = -1.0
    Defines 'metal-poor' vs 'metal-rich' subsets in Table 1; affects the interpretation of where embeddings help.
  • Fine-tuning sample composition = 2,069 stars (all 469 with [Fe/H] < -0.7 plus 1,600 random metal-rich)
    Hand-selected to oversample metal-poor stars while limiting size (Section 4); affects all fine-tuning results.
  • Learning rate / weight decay / batch size = 1e-5 / 1e-4 / 32
    Standard hyperparameters chosen for pre-training and fine-tuning; conventional, but they influence all reported metrics.
axioms (4)
  • domain assumption APOGEE DR17 labels are accurate ground truth for [Fe/H] and [α/Fe].
    All training and test labels come from APOGEE DR17 cross-matches; any systematic label error transfers to the reported R2/MMD values (Sections 4-5).
  • domain assumption Re-gridding and continuum normalization of DESI spectra to LAMOST bins preserves the information needed for parameter estimation.
    DESI spectra are linearly interpolated to 1,462 LAMOST wavelength bins and continuum-normalized; if this erases resolution-dependent information, the comparison could be biased (Section 4).
  • domain assumption SpecCLIP embeddings trained on LAMOST LRS and Gaia XP are valid inputs for DESI spectra.
    The foundation-model embeddings are generated by a model that never saw DESI; the paper assumes the shared embedding space transfers to DESI spectra (Section 2.1).
  • ad hoc to paper The 79-star metal-poor test subset is representative enough to support the reported R2 differences.
    R2 differences in the metal-poor regime (e.g., 0.678 vs 0.694) rely on only 79 test stars; the authors themselves note the small sample (Section 6.1).

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI." pith.science (2026). https://pith.science/paper/DFO7FX5V

@misc{pith2026260215021,
  author       = {Pith},
  title        = {Pith review of: Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DFO7FX5V}},
  note         = {Machine review of arXiv:2602.15021}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Cross-survey generalization is a critical challenge in stellar spectral analysis, particularly in cases such as transferring from low- to moderate-resolution surveys. We investigate this problem using pre-trained models, focusing on simple neural networks such as multilayer perceptrons (MLPs), with a case study transferring from LAMOST low-resolution spectra (LRS) to DESI medium-resolution spectra (MRS). Specifically, we pre-train MLPs on either LRS or their embeddings and fine-tune them for application to DESI stellar spectra. We compare MLPs trained directly on spectra with those trained on embeddings derived from transformer-based models (self-supervised foundation models pre-trained for multiple downstream tasks). We also evaluate different fine-tuning strategies, including residual-head fine-tuning, LoRA, and full fine-tuning. We find that MLPs pre-trained on LAMOST LRS achieve strong performance, even without fine-tuning, and that modest fine-tuning with DESI spectra further improves the results. For iron abundance, embeddings from a transformer-based model yield advantages in the metal-rich ([Fe/H] > -1.0) regime, but underperform in the metal-poor regime compared to MLPs trained directly on LRS. We also show that the optimal fine-tuning strategy depends on the specific stellar parameter under consideration. These results highlight that simple pre-trained MLPs can provide competitive cross-survey generalization, while the role of spectral foundation models for cross-survey stellar parameter estimation requires further exploration.

Figures

Figures reproduced from arXiv: 2602.15021 by Alexander S. Szalay, L\'aszl\'o Dobos, Rosemary F.G. Wyse, Tam\'as Budav\'ari, Viska Wei, Xiaosheng Zhao, Yang Huang, Yuan-Sen Ting.

Figure 1
Figure 1. Figure 1: Sketch of the pre-training and fine-tuning workflow. Pre-training uses normalized spectra or foundation-model spectral embeddings to predict labels: [Fe/H] from APOGEE (> –2.0), supplemented at lower metallicities by PASTEL, SAGA, and other VMP/UMP datasets; [α/Fe] from APOGEE. The foundation model is trained on large unlabeled datasets. Fine￾tuning then adapts the network using fewer labeled spectra, eith… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of [Fe/H] and [α/Fe] estimates between the DESI SP pipeline and MLP-based models, referenced against APOGEE DR17 labels. From left to right: DESI SP–DESI SP pipeline, MLP-Scratch–MLP trained from scratch on the same number of DESI spectra used for fine-tuning, MLP-LRS–MLP pre-trained on LAMOST LRS (zero-shot application), and MLP-LRS (fine-tuned)–MLP pre-trained on LAMOST LRS and fine-tuned with… view at source ↗
Figure 3
Figure 3. Figure 3: [α/Fe]–[Fe/H] diagrams for DESI EDR and APOGEE reference sample. From left to right: (1) DESI SP: DESI SP pipeline, (2) MLP-Scratch: MLP trained from scratch on the same number of DESI spectra used for fine-tuning, (3) MLP-LRS: MLP pre-trained on LAMOST LRS (zero-shot application), (4) MLP-LRS (fine-tuned): MLP pre-trained on LAMOST LRS and fine-tuned with DESI spectra, and (5) APOGEE: DESI DR1–APOGEE DR17… view at source ↗
Figure 4
Figure 4. Figure 4: Effect of fine-tuning sample size (0, 100, 200, 500, 1000, and 2069, where 0 denotes zero-shot performance), evaluated on the testing sample. Shown are R 2 and the ro￾bustly estimated standard deviation of the residuals (σ) for [Fe/H] and [α/Fe], obtained after fine-tuning (residual-head fine-tuning for [Fe/H] and LoRA fine-tuning for [α/Fe]). For both metrics, we plot the mean and 1σ error bars from five … view at source ↗
Figure 5
Figure 5. Figure 5: Loss landscapes for different fine-tuning strategies, evaluated on the metal-poor regime of [Fe/H] (79 stars). The first row shows 3D surfaces, and the second row shows 2D contours of the logarithmic MSE loss, plotted as a function of perturbations around the pre-trained model parameters along two random directions. The minimum losses are indicated in the colorbar of the 2D contours. rameters for the LAMOS… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Foundation Models for Astrophysics

    astro-ph.IM 2026-08 conditional novelty 3.0

    Astronomical 'foundation models' largely reuse transformers and self-supervised pretraining, but evidence of transfer to new instruments, populations, or tasks remains rare; the paper argues such evidence, not archite...

Reference graph

Works this paper leans on

46 extracted references · 4 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    2022, ApJS, 259, 35, doi: 10.3847/1538-4365/ac4414

    Abdurro’uf, Accetta, K., Aerts, C., et al. 2022, ApJS, 259, 35, doi: 10.3847/1538-4365/ac4414

    doi:10.3847/1538-4365/ac4414 2022

  2. [2]

    2023, FERRE: Match physical models to measurements, Astrophysics Source Code Library, record ascl:2301.016

    Allende-Prieto, C., & Apogee Team. 2023, FERRE: Match physical models to measurements, Astrophysics Source Code Library, record ascl:2301.016. http://ascl.net/2301.016 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068 Astropy Collaboration, Price-Whelan, A. M., Sip˝ ocz, B. M., et al. 2...

    doi:10.1051/0004-6361/201322068 2023

  3. [3]

    2019, MNRAS, 486, 2075, doi: 10.1093/mnras/stz549

    Blanco-Cuaresma, S. 2019, MNRAS, 486, 2075, doi: 10.1093/mnras/stz549

    doi:10.1093/mnras/stz549 2019

  4. [4]

    2020, in Advances in Neural Information Processing Systems, ed

    Brown, T., Mann, B., Ryder, N., et al. 2020, in Advances in Neural Information Processing Systems, ed. H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, & H. Lin, Vol. 33 (Curran Associates, Inc.), 1877–1901

    2020

  5. [5]

    2024, arXiv e-prints, arXiv:2410.16081, doi: 10.48550/arXiv.2410.16081 De Angeli, F., Weiler, M., Montegriffo, P., et al

    Buck, T., & Schwarz, C. 2024, arXiv e-prints, arXiv:2410.16081, doi: 10.48550/arXiv.2410.16081 De Angeli, F., Weiler, M., Montegriffo, P., et al. 2023, A&A, 674, A2, doi: 10.1051/0004-6361/202243680 de Jong, J. T. A., Yanny, B., Rix, H.-W., et al. 2010, ApJ, 714, 663, doi: 10.1088/0004-637X/714/1/663 De Silva, G. M., Freeman, K. C., Bland-Hawthorn, J., et...

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.16081 2024

  6. [6]

    2018, arXiv e-prints, arXiv:1810.04805, doi: 10.48550/arXiv.1810.04805 Devon Hjelm, R., Fedorov, A., Lavoie-Marchildon, S., et al

    Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. 2018, arXiv e-prints, arXiv:1810.04805, doi: 10.48550/arXiv.1810.04805 Devon Hjelm, R., Fedorov, A., Lavoie-Marchildon, S., et al. 2018, arXiv e-prints, arXiv:1808.06670, doi: 10.48550/arXiv.1808.06670 Euclid Collaboration, Siudek, M., Huertas-Company, M., et al. 2025, arXiv e-prints, arXiv:2503.15312, do...

    Pith/arXiv arXiv doi:10.48550/arxiv.1810.04805 2018

  7. [7]

    J., Olsen, K., Economou, F., et al

    Fitzpatrick, M. J., Olsen, K., Economou, F., et al. 2014, in Observatory Operations: Strategies, Processes, and Systems V, ed. A. B. Peck, C. R. Benn, & R. L. Seaman, Vol. 9149, International Society for Optics and Photonics (SPIE), 91491T, doi: 10.1117/12.2057445 Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al. 2023, A&A, 674, A1, doi: 10.1051/...

    doi:10.1117/12.2057445 2014

  8. [8]

    M., Rasch, M

    Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch¨ olkopf, B., & Smola, A. 2012, Journal of Machine Learning Research, 13, 723

    2012

  9. [9]

    J., Shen, Y., Wallis, P., et al

    Hu, E. J., Shen, Y., Wallis, P., et al. 2021, arXiv e-prints, arXiv:2106.09685, doi: 10.48550/arXiv.2106.09685

    Pith/arXiv arXiv doi:10.48550/arxiv.2106.09685 2021

  10. [10]

    C., Xiao, K., et al

    Huang, Y., Beers, T. C., Xiao, K., et al. 2024, ApJ, 974, 192, doi: 10.3847/1538-4357/ad6b94

    doi:10.3847/1538-4357/ad6b94 2024

  11. [11]

    2021, Computing in Science & Engineering, 23, 15, doi: 10.1109/MCSE.2021.3057097

    Juneau, S., Olsen, K., Nikutta, R., Jacques, A., & Bailey, S. 2021, Computing in Science & Engineering, 23, 15, doi: 10.1109/MCSE.2021.3057097

    arXiv 2021

  12. [12]

    SPARCL: SPectra Analysis and Retrievable Catalog Lab

    Juneau, S., Jacques, A., Pothier, S., et al. 2024, arXiv e-prints, arXiv:2401.05576, doi: 10.48550/arXiv.2401.05576

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2401.05576 2024

  13. [13]

    2024, arXiv e-prints, arXiv:2411.04750, doi: 10.48550/arXiv.2411.04750

    Koblischke, N., & Bovy, J. 2024, arXiv e-prints, arXiv:2411.04750, doi: 10.48550/arXiv.2411.04750

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.04750 2024

  14. [14]

    E., Allende Prieto, C., Cooper, A

    Koposov, S. E., Allende Prieto, C., Cooper, A. P., et al. 2024, MNRAS, 533, 1012, doi: 10.1093/mnras/stae1842

    doi:10.1093/mnras/stae1842 2024

  15. [15]

    E., Li, T

    Koposov, S. E., Li, T. S., Allende Prieto, C., et al. 2025, arXiv e-prints, arXiv:2505.14787, doi: 10.48550/arXiv.2505.14787

    doi:10.48550/arxiv.2505.14787 2025

  16. [16]

    2024, NIST Atomic Spectra Database, https://physics.nist.gov/asd, Gaithersburg, MD: National Institute of Standards and Technology, doi: 10.18434/T4W30F

    Kramida, A., Ralchenko, Y., Reader, J., & NIST ASD Team. 2024, NIST Atomic Spectra Database, https://physics.nist.gov/asd, Gaithersburg, MD: National Institute of Standards and Technology, doi: 10.18434/T4W30F

    doi:10.18434/t4w30f 2024

  17. [17]

    W., & Bovy, J

    Leung, H. W., & Bovy, J. 2024, MNRAS, 527, 1494, doi: 10.1093/mnras/stad3015

    doi:10.1093/mnras/stad3015 2024

  18. [18]

    2017, arXiv e-prints, arXiv:1712.09913, doi: 10.48550/arXiv.1712.09913

    Li, H., Xu, Z., Taylor, G., Studer, C., & Goldstein, T. 2017, arXiv e-prints, arXiv:1712.09913, doi: 10.48550/arXiv.1712.09913

    Pith/arXiv arXiv doi:10.48550/arxiv.1712.09913 2017

  19. [19]

    2022, ApJ, 931, 147, doi: 10.3847/1538-4357/ac6514

    Li, H., Aoki, W., Matsuno, T., et al. 2022, ApJ, 931, 147, doi: 10.3847/1538-4357/ac6514

    doi:10.3847/1538-4357/ac6514 2022

  20. [20]

    2017, arXiv e-prints, arXiv:1711.05101, doi: 10.48550/arXiv.1711.05101

    Loshchilov, I., & Hutter, F. 2017, arXiv e-prints, arXiv:1711.05101, doi: 10.48550/arXiv.1711.05101

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101 2017

  21. [21]

    R., Schiavon, R

    Majewski, S. R., Schiavon, R. P., Frinchaboy, P. M., et al. 2017, AJ, 154, 94, doi: 10.3847/1538-3881/aa784d

    doi:10.3847/1538-3881/aa784d 2017

  22. [22]

    2015, ApJ, 808, 16, doi: 10.1088/0004-637X/808/1/16

    Zasowski, G. 2015, ApJ, 808, 16, doi: 10.1088/0004-637X/808/1/16

    doi:10.1088/0004-637x/808/1/16 2015

  23. [23]

    2020, Astronomy and Computing, 33, 100411, doi: https://doi.org/10.1016/j.ascom.2020.100411

    Nikutta, R., Fitzpatrick, M., Scott, A., & Weaver, B. 2020, Astronomy and Computing, 33, 100411, doi: https://doi.org/10.1016/j.ascom.2020.100411

    arXiv 2020

  24. [24]

    2024, MNRAS, 531, 4990, doi: 10.1093/mnras/stae1450

    Parker, L., Lanusse, F., Golkar, S., et al. 2024, MNRAS, 531, 4990, doi: 10.1093/mnras/stae1450

    doi:10.1093/mnras/stae1450 2024

  25. [25]

    2019, in Advances in Neural Information Processing Systems 32 (Curran

    Paszke, A., Gross, S., Massa, F., et al. 2019, in Advances in Neural Information Processing Systems 32 (Curran

    2019

  26. [26]

    SpecPT (Spectroscopy Pre-trained Transformer) Model for Extragalactic Spectroscopy: I. Architecture and Automated Redshift Measurement

    Pattnaik, R., Kartaltepe, J. S., & Binu, C. 2025, arXiv e-prints, arXiv:2501.01070, doi: 10.48550/arXiv.2501.01070

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2501.01070 2025

  27. [27]

    2009, WSEAS Transactions on Circuits and Systems, 8, 579

    Mastorakis, N. 2009, WSEAS Transactions on Circuits and Systems, 8, 579

    2009

  28. [28]

    Radford, A., Narasimhan, K., Salimans, T., & Sutskever, I. 2018

    2018

  29. [29]

    2019, Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., et al. 2019, Language models are unsupervised multitask learners

    2019

  30. [30]

    Rizhko, M., & Bloom, J. S. 2024, arXiv e-prints, arXiv:2411.08842, doi: 10.48550/arXiv.2411.08842 R´ o˙ za´ nski, T., Ting, Y.-S., & Jab lo´ nska, M. 2025, ApJ, 980, 66, doi: 10.3847/1538-4357/ad9b99

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.08842 2024

  31. [31]

    F., et al

    Sestito, F., Longeard, N., Martin, N. F., et al. 2019, MNRAS, 484, 2166, doi: 10.1093/mnras/stz043

    doi:10.1093/mnras/stz043 2019

  32. [32]

    2013, arXiv e-prints, arXiv:1312.6034

    Simonyan, K., Vedaldi, A., & Zisserman, A. 2013, arXiv e-prints, arXiv:1312.6034. https://arxiv.org/abs/1312.6034

    Pith/arXiv arXiv 2013

  33. [33]

    2024, arXiv e-prints, arXiv:2405.14930, doi: 10.48550/arXiv.2405.14930

    Huertas-Company, M. 2024, arXiv e-prints, arXiv:2405.14930, doi: 10.48550/arXiv.2405.14930

    Pith/arXiv arXiv doi:10.48550/arxiv.2405.14930 2024

  34. [34]

    2016, A&A, 591, A118, doi: 10.1051/0004-6361/201628497

    Soubiran, C., Le Campion, J.-F., Brouillet, N., & Chemin, L. 2016, A&A, 591, A118, doi: 10.1051/0004-6361/201628497

    doi:10.1051/0004-6361/201628497 2016

  35. [35]

    2006, AJ, 132, 1645, doi: 10.1086/506564

    Steinmetz, M., Zwitter, T., Siebert, A., et al. 2006, AJ, 132, 1645, doi: 10.1086/506564

    doi:10.1086/506564 2006

  36. [36]

    2008, PASJ, 60, 1159, doi: 10.1093/pasj/60.5.1159

    Suda, T., Katsuta, Y., Yamada, S., et al. 2008, PASJ, 60, 1159, doi: 10.1093/pasj/60.5.1159

    doi:10.1093/pasj/60.5.1159 2008

  37. [37]

    2019, ApJ, 879, 69, doi: 10.3847/1538-4357/ab2331

    Ting, Y.-S., Conroy, C., Rix, H.-W., & Cargile, P. 2019, ApJ, 879, 69, doi: 10.3847/1538-4357/ab2331

    doi:10.3847/1538-4357/ab2331 2019

  38. [38]

    Ting, Y.-S., Rix, H.-W., Conroy, C., Ho, A. Y. Q., & Lin, J. 2017, ApJL, 849, L9, doi: 10.3847/2041-8213/aa921c

    doi:10.3847/2041-8213/aa921c 2017

  39. [39]

    2017, in Advances in Neural Information Processing Systems, ed

    Vaswani, A., Shazeer, N., Parmar, N., et al. 2017, in Advances in Neural Information Processing Systems, ed. I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, Vol. 30 (Curran Associates, Inc.)

    2017

  40. [40]

    2019, ApJS, 245, 34, doi: 10.3847/1538-4365/ab5364

    Xiang, M., Ting, Y.-S., Rix, H.-W., et al. 2019, ApJS, 245, 34, doi: 10.3847/1538-4365/ab5364

    doi:10.3847/1538-4365/ab5364 2019

  41. [41]

    2024, ApJS, 273, 19, doi: 10.3847/1538-4365/ad51dd

    Zhang, M., Xiang, M., Ting, Y.-S., et al. 2024, ApJS, 273, 19, doi: 10.3847/1538-4365/ad51dd

    doi:10.3847/1538-4365/ad51dd 2024

  42. [42]

    2012, Research in Astronomy and Astrophysics, 12, 723, doi: 10.1088/1674-4527/12/7/002

    Zhao, G., Zhao, Y.-H., Chu, Y.-Q., Jing, Y.-P., & Deng, L.-C. 2012, Research in Astronomy and Astrophysics, 12, 723, doi: 10.1088/1674-4527/12/7/002

    doi:10.1088/1674-4527/12/7/002 2012

  43. [43]

    2025, SpecCLIP v1.0.0: Aligning and Translating Spectroscopic Measurements for Stars, v1.0.0, Zenodo, doi: 10.5281/zenodo.17824840

    Zhao, X., & Huang, Y. 2025, SpecCLIP v1.0.0: Aligning and Translating Spectroscopic Measurements for Stars, v1.0.0, Zenodo, doi: 10.5281/zenodo.17824840

    work page doi:10.5281/zenodo.17824840 2025

  44. [44]

    S., & Huang, Y

    Zhao, X., Ting, Y.-S., Szalay, A. S., & Huang, Y. 2025a, arXiv e-prints, arXiv:2507.20972. https://arxiv.org/abs/2507.20972

    Pith/arXiv arXiv

  45. [45]

    2025b, arXiv e-prints, arXiv:2507.01939, doi: 10.48550/arXiv.2507.01939

    Zhao, X., Huang, Y., Xue, G., et al. 2025b, arXiv e-prints, arXiv:2507.01939, doi: 10.48550/arXiv.2507.01939

    work page doi:10.48550/arxiv.2507.01939

  46. [46]

    R., Heneka, C., et al

    Zhong, F., Napolitano, N. R., Heneka, C., et al. 2024, arXiv e-prints, arXiv:2412.21130, doi: 10.48550/arXiv.2412.21130

    doi:10.48550/arxiv.2412.21130 2024

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.