pith. sign in

arxiv: 2605.22162 · v2 · pith:QIRRYZT2new · submitted 2026-05-21 · 🌌 astro-ph.IM · astro-ph.SR· cs.LG

Spectra as Language: Large Language Models for Scalable Stellar Parameter and Abundance Inference

Hai-Ling Lu , Yu-Yang Li , Yin-Bi Li , Cun-Shi Wang , A-Li Luo , Jun-Chao Liang , Shuo Li This is my paper

Pith reviewed 2026-05-25 02:49 UTC · model grok-4.3

classification 🌌 astro-ph.IM astro-ph.SRcs.LG
keywords stellar spectralarge language modelsstellar parameterschemical abundancesspectroscopic surveysscaling lawsparameter inference
0
0 comments X

The pith

A two-stage large language model framework infers stellar parameters and chemical abundances by treating spectra as sequential signals.

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

The paper shows that stellar spectra can be handled as continuous sequential data, allowing architectures developed for language modeling to estimate effective temperature, surface gravity, metallicity, and abundances for about twenty elements. This targets the challenge of processing massive spectroscopic survey data where traditional fitting methods become inefficient. Scaling analyses demonstrate that estimation accuracy improves as the volume of training spectra increases, offering a path to handle even larger future datasets.

Core claim

The central claim is that a two-stage large language model framework, by modeling stellar spectra directly as sequential signals, achieves accurate estimation of effective temperature, surface gravity, metallicity, and abundances of roughly twenty chemical elements, with performance following scaling laws that improve systematically with more data.

What carries the argument

Two-stage large language model framework that processes stellar spectra as continuous sequential signals

If this is right

  • Accurate estimates of effective temperature, surface gravity, metallicity, and abundances of ~20 elements become feasible on high-dimensional survey data.
  • Performance on parameter inference improves in a predictable way as the quantity of training spectra grows.
  • The framework supplies a scalable route for processing data from forthcoming large spectroscopic surveys.

Where Pith is reading between the lines

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

  • The same sequential modeling approach could be tested on other ordered scientific signals such as time-series photometry or radial-velocity curves.
  • If the scaling continues, the method might reduce the need for per-spectrum manual feature engineering in future surveys.
  • Cross-survey generalization could be checked by training on one instrument's spectra and evaluating on another's without retraining.

Load-bearing premise

Stellar spectra act as continuous sequential signals that can adopt language-model architectures and training methods without losing the physical links between wavelength bins and stellar parameters.

What would settle it

Training the model on successively larger spectral datasets produces no measurable gain in accuracy for effective temperature, surface gravity, metallicity, or element abundances on held-out test spectra.

Figures

Figures reproduced from arXiv: 2605.22162 by A-Li Luo, Cun-Shi Wang, Hai-Ling Lu, Jun-Chao Liang, Shuo Li, Yin-Bi Li, Yu-Yang Li.

Figure 1
Figure 1. Figure 1: Teff compared with Gaia-ESO. Ramachandra, N., Ting, Y.-S., Sun, Z., Wells, A., & Habib, S. 2025, arXiv e-prints, arXiv:2508.10075, doi: 10.48550/arXiv.2508.10075 Shao, M., Wang, H., Li, Y., et al. 2025, arXiv e-prints, arXiv:2511.08970, doi: 10.48550/arXiv.2511.08970 Shetrone, M., Beaton, R. L., Hayes, C. R., et al. 2025, arXiv e-prints, arXiv:2511.04365, doi: 10.48550/arXiv.2511.04365 Smolinski, J. P., Le… view at source ↗
Figure 2
Figure 2. Figure 2: logg compared with Gaia-ESO. Zhao, F., Li, Y., Liu, Z., et al. 2025, Machine Learning: Science and Technology, 6, 045005, doi: 10.1088/2632-2153/ae0c56 Zheng, Z.-P., Qiu, B., Luo, A. L., & Li, Y.-B. 2020, PASP, 132, 024504, doi: 10.1088/1538-3873/ab5ed7 [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: [Fe/H] compared with Gaia-ESO [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Teff compared with Galah DR4 [PITH_FULL_IMAGE:figures/full_fig_p019_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: logg compared with Galah DR4 [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: [Fe/H] compared with Galah DR4 [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Abundance compared with Gaia ESO [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Abundance compared with GALAH DR4 [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
read the original abstract

Stellar spectra encode key information on the physical properties and chemical compositions of stars. Accurate stellar parameter determination is essential for addressing major questions such as galaxy and stellar evolution. Large-scale spectroscopic surveys have accumulated unprecedented spectral data. Traditional feature extraction or model-fitting approaches struggle with high-dimensional, massive datasets, limited generalization, and computational inefficiency. Recent advances in large language models demonstrate strong generalization and feature-learning in tasks like natural language processing, DNA/RNA sequence analysis, and protein/chemical parsing. Stellar spectra are continuous sequential signals, enabling the transfer of language models to stellar spectroscopy. Here, we propose a two-stage large language model framework for stellar parameter inference, achieving accurate estimation of effective temperature, surface gravity, metallicity, and abundances of ~20 chemical elements. Scaling-law analyses show systematic performance improvements with increasing data, providing a scalable framework for forthcoming large-scale surveys.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes a two-stage large language model framework that treats stellar spectra as continuous sequential signals to infer effective temperature, surface gravity, metallicity, and abundances for ~20 chemical elements. It claims accurate estimation is achieved and that scaling-law analyses demonstrate systematic performance gains with increasing data volume, offering a scalable alternative to traditional methods for large spectroscopic surveys.

Significance. If the accuracy claims and scaling behavior are substantiated with rigorous validation, the work could provide a data-efficient, generalizable pipeline for processing the high-volume outputs of forthcoming surveys, complementing physics-based fitting approaches in stellar and galactic archaeology.

major comments (2)
  1. [Abstract] Abstract: the central claim of 'accurate estimation' of Teff, log g, [M/H] and ~20 abundances is unsupported by any reported quantitative metrics, error budgets, cross-validation procedures, or baseline comparisons, rendering the performance assertions unevaluable from the given information.
  2. [Introduction/Methods] Introduction/Methods (assumed §2): the load-bearing assumption that next-token-prediction inductive biases and positional embeddings transfer to stellar spectra without material loss of wavelength-specific physical correlations (radiative transfer, line profiles, continuum opacity) is stated but not tested; the manuscript must demonstrate generalization across resolutions or instruments, as failure here would invalidate the scaling-law results for new surveys.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help clarify the presentation of our results. We respond point-by-point to the major comments below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim of 'accurate estimation' of Teff, log g, [M/H] and ~20 abundances is unsupported by any reported quantitative metrics, error budgets, cross-validation procedures, or baseline comparisons, rendering the performance assertions unevaluable from the given information.

    Authors: The abstract provides a concise overview; the manuscript reports quantitative metrics (MAE, RMSE, and cross-validation results), error budgets, and baseline comparisons in Sections 3 and 4. To improve self-containment, we will revise the abstract to include representative numerical performance values. revision: yes

  2. Referee: [Introduction/Methods] Introduction/Methods (assumed §2): the load-bearing assumption that next-token-prediction inductive biases and positional embeddings transfer to stellar spectra without material loss of wavelength-specific physical correlations (radiative transfer, line profiles, continuum opacity) is stated but not tested; the manuscript must demonstrate generalization across resolutions or instruments, as failure here would invalidate the scaling-law results for new surveys.

    Authors: We agree that explicit tests of generalization across resolutions and instruments are required to substantiate transferability. The present manuscript uses a single homogeneous dataset to establish scaling behavior. We will add experiments evaluating performance on spectra at varied resolutions and, where data permit, across instruments. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical framework with independent performance claims

full rationale

The manuscript proposes a two-stage LLM architecture for inferring stellar parameters and abundances by treating spectra as sequential signals, then reports empirical accuracy and scaling-law improvements with data volume. No equations, parameter fits, or self-citations are shown that would make any reported prediction equivalent to its inputs by construction. The central claims rest on observed model performance rather than definitional or self-referential reductions, satisfying the default expectation of non-circularity for an empirical methods paper.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The central claim implicitly rests on the unstated premise that spectra behave sufficiently like natural-language sequences for LLM transfer to succeed.

pith-pipeline@v0.9.0 · 5703 in / 1120 out tokens · 20224 ms · 2026-05-25T02:49:31.640560+00:00 · methodology

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages · 1 internal anchor

  1. [1]

    , keywords =

    Abdurro'uf , Accetta , K., Aerts , C., et al. 2022, , 259, 35, 10.3847/1538-4365/ac4414

    work page doi:10.3847/1538-4365/ac4414 2022

  2. [2]

    A., et al

    Bialek , S., Fabbro , S., Venn , K. A., et al. 2020, , 498, 3817, 10.1093/mnras/staa2582

    work page doi:10.1093/mnras/staa2582 2020

  3. [3]

    , keywords =

    Buder , S., Kos , J., Wang , X. E., et al. 2025, , 42, e051, 10.1017/pasa.2025.26

    work page doi:10.1017/pasa.2025.26 2025

  4. [4]

    Chaini, S., & Kumar, S. S. 2020, Astronomical Classification of Light Curves with an Ensemble of Gated Recurrent Units. 2006.12333

    work page arXiv 2020

  5. [5]

    2016, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16 (ACM), 785–794, 10.1145/2939672.2939785

    Chen, T., & Guestrin, C. 2016, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16 (ACM), 785–794, 10.1145/2939672.2939785

    work page doi:10.1145/2939672.2939785 2016

  6. [6]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. 2014, Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling. 1412.3555

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  7. [7]

    Research in Astronomy and Astrophysics , year = 2012, month = sep, volume =

    Cui , X.-Q., Zhao , Y.-H., Chu , Y.-Q., et al. 2012, Research in Astronomy and Astrophysics, 12, 1197, 10.1088/1674-4527/12/9/003

    work page doi:10.1088/1674-4527/12/9/003 2012

  8. [8]

    E., Allende Prieto , C., Holtzman , J

    Garc \' a P \'e rez , A. E., Allende Prieto , C., Holtzman , J. A., et al. 2016, , 151, 144, 10.3847/0004-6256/151/6/144

    work page doi:10.3847/0004-6256/151/6/144 2016

  9. [9]

    2025, RAS Techniques and Instruments, 4, rzaf048, 10.1093/rasti/rzaf048

    Gilda , S. 2025, RAS Techniques and Instruments, 4, rzaf048, 10.1093/rasti/rzaf048

    work page doi:10.1093/rasti/rzaf048 2025

  10. [10]

    Group, N. L. C. 2017, in Proceedings of ... https://www.microsoft.com/en-us/research/publication/mcr/

    2017

  11. [11]

    2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770--778, 10.1109/CVPR.2016.90

    He, K., Zhang, X., Ren, S., & Sun, J. 2016, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770--778, 10.1109/CVPR.2016.90

    work page doi:10.1109/cvpr.2016.90 2016

  12. [12]

    1997, Neural Computation, 9, 1735, 10.1162/neco.1997.9.8.1735

    Hochreiter, S., & Schmidhuber, J. 1997, Neural Computation, 9, 1735, 10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  13. [13]

    2024, arXiv e-prints, arXiv:2402.03182, 10.48550/arXiv.2402.03182

    Jiang , Y., Pan , Z., Zhang , X., et al. 2024, arXiv e-prints, arXiv:2402.03182, 10.48550/arXiv.2402.03182

    work page doi:10.48550/arxiv.2402.03182 2024

  14. [14]

    2017, in Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17 (Red Hook, NY, USA: Curran Associates Inc.), 3149–3157

    Ke, G., Meng, Q., Finley, T., et al. 2017, in Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17 (Red Hook, NY, USA: Curran Associates Inc.), 3149–3157

    2017

  15. [15]

    S., Beers , T

    Lee , Y. S., Beers , T. C., Sivarani , T., et al. 2008, , 136, 2022, 10.1088/0004-6256/136/5/2022

    work page doi:10.1088/0004-6256/136/5/2022 2008

  16. [16]

    2025, , 281, 58, 10.3847/1538-4365/ae1586

    Li , S., Li , Y.-B., Luo , A.-L., et al. 2025, , 281, 58, 10.3847/1538-4365/ae1586

    work page doi:10.3847/1538-4365/ae1586 2025

  17. [17]

    2024, arXiv e-prints, arXiv:2404.10757, 10.48550/arXiv.2404.10757

    Li , Y.-Y., Bai , Y., Wang , C., et al. 2024, arXiv e-prints, arXiv:2404.10757, 10.48550/arXiv.2404.10757

    work page doi:10.48550/arxiv.2404.10757 2024

  18. [18]

    2022, , 517, 4875, 10.1093/mnras/stac1959

    Li , Z., Zhao , G., Chen , Y., Liang , X., & Zhao , J. 2022, , 517, 4875, 10.1093/mnras/stac1959

    work page doi:10.1093/mnras/stac1959 2022

  19. [19]

    2022, The Astronomical Journal, 163, 153, 10.3847/1538-3881/ac4d97

    Liang, J., Bu, Y., Tan, K., et al. 2022, The Astronomical Journal, 163, 153, 10.3847/1538-3881/ac4d97

    work page doi:10.3847/1538-3881/ac4d97 2022

  20. [20]

    L., Zhao, Y

    Luo , A. L., Zhao , Y.-H., Zhao , G., et al. 2015, Research in Astronomy and Astrophysics, 15, 1095, 10.1088/1674-4527/15/8/002

    work page doi:10.1088/1674-4527/15/8/002 2015

  21. [21]

    2016, in Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS'16 (Red Hook, NY, USA: Curran Associates Inc.), 1279–1287

    Meng, Q., Ke, G., Wang, T., et al. 2016, in Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS'16 (Red Hook, NY, USA: Curran Associates Inc.), 1279–1287

    2016

  22. [22]

    Q., & Wilson, A

    Nate Gruver, Marc Finzi, S. Q., & Wilson, A. G. 2023, in Advances in Neural Information Processing Systems

    2023

  23. [23]

    W., Rix , H.-W., Ho , A

    Ness , M., Hogg , D. W., Rix , H.-W., Ho , A. Y. Q., & Zasowski , G. 2015, , 808, 16, 10.1088/0004-637X/808/1/16

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

  24. [24]

    2021, , 906, 130, 10.3847/1538-4357/abca96

    O'Briain , T., Ting , Y.-S., Fabbro , S., et al. 2021, , 906, 130, 10.3847/1538-4357/abca96

    work page doi:10.3847/1538-4357/abca96 2021

  25. [25]

    2025, arXiv e-prints, arXiv:2510.17960, 10.48550/arXiv.2510.17960

    Parker , L., Lanusse , F., Shen , J., et al. 2025, arXiv e-prints, arXiv:2510.17960, 10.48550/arXiv.2510.17960

    work page doi:10.48550/arxiv.2510.17960 2025

  26. [26]

    2025, arXiv e-prints, arXiv:2508.10075, 10.48550/arXiv.2508.10075

    Ramachandra , N., Ting , Y.-S., Sun , Z., Wells , A., & Habib , S. 2025, arXiv e-prints, arXiv:2508.10075, 10.48550/arXiv.2508.10075

    work page doi:10.48550/arxiv.2508.10075 2025

  27. [27]

    2025, arXiv e-prints, arXiv:2511.08970, 10.48550/arXiv.2511.08970

    Shao , M., Wang , H., Li , Y., et al. 2025, arXiv e-prints, arXiv:2511.08970, 10.48550/arXiv.2511.08970

    work page doi:10.48550/arxiv.2511.08970 2025

  28. [28]

    L., Hayes , C

    Shetrone , M., Beaton , R. L., Hayes , C. R., et al. 2025, arXiv e-prints, arXiv:2511.04365, 10.48550/arXiv.2511.04365

    work page doi:10.48550/arxiv.2511.04365 2025

  29. [29]

    P., Lee , Y

    Smolinski , J. P., Lee , Y. S., Beers , T. C., et al. 2011, , 141, 89, 10.1088/0004-6256/141/3/89

    work page doi:10.1088/0004-6256/141/3/89 2011

  30. [30]

    J., et al

    Steinmetz , M., Guiglion , G., McMillan , P. J., et al. 2020, , 160, 83, 10.3847/1538-3881/ab9ab8

    work page doi:10.3847/1538-3881/ab9ab8 2020

  31. [31]

    2025, Nature Astronomy, 9, 1869, 10.1038/s41550-025-02670-z

    Stoppa , F., Bulmus , T., Bloemen , S., et al. 2025, Nature Astronomy, 9, 1869, 10.1038/s41550-025-02670-z

    work page doi:10.1038/s41550-025-02670-z 2025

  32. [32]

    2019, , 879, 69, 10.3847/1538-4357/ab2331

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

    work page doi:10.3847/1538-4357/ab2331 2019

  33. [33]

    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.). https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf

    2017

  34. [34]

    2024, arXiv e-prints, arXiv:2404.10019, 10.48550/arXiv.2404.10019

    Wang , Y., Zhang , S.-R., Momtaz , A., et al. 2024, arXiv e-prints, arXiv:2404.10019, 10.48550/arXiv.2404.10019

    work page doi:10.48550/arxiv.2404.10019 2024

  35. [35]

    2019, , 245, 34, 10.3847/1538-4365/ab5364

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

    work page doi:10.3847/1538-4365/ab5364 2019

  36. [36]

    2017, , 464, 3657, 10.1093/mnras/stw2523

    Xiang , M.-S., Liu , X.-W., Shi , J.-R., et al. 2017, , 464, 3657, 10.1093/mnras/stw2523

    work page doi:10.1093/mnras/stw2523 2017

  37. [37]

    2023, arXiv e-prints, arXiv:2308.13565, 10.48550/arXiv.2308.13565

    Xie , T., Wan , Y., Huang , W., et al. 2023, arXiv e-prints, arXiv:2308.13565, 10.48550/arXiv.2308.13565

    work page doi:10.48550/arxiv.2308.13565 2023

  38. [38]

    2024 a , arXiv e-prints, arXiv:2401.14656, 10.48550/arXiv.2401.14656

    Zhang , Q., Ding , K., Lyv , T., et al. 2024 a , arXiv e-prints, arXiv:2401.14656, 10.48550/arXiv.2401.14656

    work page doi:10.48550/arxiv.2401.14656 2024

  39. [39]

    R., Gupta , R

    Zhang , X., Chowdhury , R. R., Gupta , R. K., & Shang , J. 2024 b , arXiv e-prints, arXiv:2402.01801, 10.48550/arXiv.2402.01801

    work page doi:10.48550/arxiv.2402.01801 2024

  40. [40]

    2025, Machine Learning: Science and Technology, 6, 045005, 10.1088/2632-2153/ae0c56

    Zhao , F., Li , Y., Liu , Z., et al. 2025, Machine Learning: Science and Technology, 6, 045005, 10.1088/2632-2153/ae0c56

    work page doi:10.1088/2632-2153/ae0c56 2025

  41. [41]

    2026, , 998, 189, 10.3847/1538-4357/ae2c7e

    Zhao , X., Huang , Y., Xue , G., et al. 2026, , 998, 189, 10.3847/1538-4357/ae2c7e

    work page doi:10.3847/1538-4357/ae2c7e 2026

  42. [42]

    L., & Li , Y.-B

    Zheng , Z.-P., Qiu , B., Luo , A. L., & Li , Y.-B. 2020, , 132, 024504, 10.1088/1538-3873/ab5ed7

    work page doi:10.1088/1538-3873/ab5ed7 2020