Pith. sign in

REVIEW 3 major objections 6 minor 273 references

A transformer trained on SDSS quasar spectra can reconstruct masked broad emission lines to within 4–16% of the flux, and outperforms purpose-built algorithms on a blind reconstruction challenge.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 12:09 UTC pith:RP5PFWCU

load-bearing objection Solid self-supervised spectral reconstruction with a clean transformer design, but the X-Shooter win may be object leakage rather than cross-resolution transfer. the 3 major comments →

arxiv 2607.19640 v1 pith:RP5PFWCU submitted 2026-07-22 astro-ph.GA astro-ph.IM

AGNFormer I: Reconstruction of AGN spectra using a probabilistic transformer model

Benedict L. Rouse , Franz E. Bauer , Tomasz R\'o\.za\'nski , Guillermo Cabrera-Vives , Patricia Ar\'evalo , Crist\'obal Rodrigo Donoso Oliva , Pavlos Protopapas , Gordon T. Richards
show 3 more authors
Paula S\'anchez-S\'aez Ezequiel Treister Christian Wolf
This is my paper
classification astro-ph.GA astro-ph.IM
keywords active galactic nucleiquasar spectratransformermasked autoencoderspectral reconstructionbroad-line regionuncertainty calibrationself-supervised learning
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 paper argues that a single transformer trained with a masked-autoencoding objective on the full SDSS DR16Q quasar catalogue can reconstruct masked broad emission lines (Hα, Hβ, MgII, CIV, Lyα, and others) using only the surrounding continuum and weak features. On held-out spectra the reconstruction errors are 10–16% of the flux for the full test set and 4–8% for high signal-to-noise (S/N>10) spectra, with an intrinsic floor of roughly 2–6% at S/N≈40; even larger masked halves are recovered to 12–25% (5–15%). The same untuned model wins a blind challenge for the Lyα region against eleven purpose-built reconstruction algorithms on both SDSS and X-Shooter data, demonstrating resolution-invariant transfer. If correct, this shows that the continuum and weak lines encode enough information to determine the broad-line region, enabling AGN property extraction (e.g., black hole masses) even when the main lines are corrupted, absorbed, or outside the observed window.

Core claim

AGNFormer is an encoder-only transformer trained self-supervised on ~750,000 SDSS DR16Q quasar spectra to predict both mean flux and per-pixel variance of masked regions. The central claim is that the broad-line region is learnable from the rest of the spectrum: with major broad lines masked to ±10,000 km/s, it reproduces unseen spectra to 10–16% (full) and 4–8% (S/N>10) of the flux, with well-calibrated uncertainties. On a blind Lyα reconstruction challenge it beats all eleven purpose-built algorithms (MAE 0.017 X-Shooter, 0.022 SDSS) despite being trained only on SDSS at R≈2000 and tested on higher-resolution X-Shooter data without fine-tuning. The authors conclude that continuum and weak

What carries the argument

AGNFormer is a transformer encoder trained with a masked-autoencoding objective. Three design choices carry the argument: (1) a wavelength-based positional encoder substituting log10λ for token indices, making the model insensitive to pixel grid and spectral resolution; (2) an uncertainty-aware head predicting both mean flux and variance, with a heteroscedastic Gaussian negative-log-likelihood loss weighting each pixel by its observational error (input fluxes are noise-perturbed during training); (3) a masking scheme combining random pixel, random chunk, and fixed broad-line masks (35:40:25), plus a Euclidean normalization computed only over unmasked pixels to prevent leakage. Together these

Load-bearing premise

The headline accuracy is measured on a random split of the same SDSS DR16Q catalogue used for training, so training and test spectra share the survey's noise, calibration, and redshift-assignment systematics; if the model exploits those systematics rather than intrinsic AGN correlations, the reported percentages overstate performance on genuinely new data.

What would settle it

Evaluate the fully trained model on a set of ~100 quasar spectra from a different instrument covering the same rest-frame lines, without interpolating to the SDSS pixel grid, and compute the predicted/observed flux ratios over the full ±10,000 km/s windows; if the median error exceeds the quoted 4–8% for S/N>10, the generalization claim is falsified. Alternatively, retrain with noise values shuffled randomly across pixels: a sharp drop in reconstruction accuracy would indicate the model uses noise-pattern information rather than spectral physics.

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

If this is right

  • If the broad-line region is this well constrained by the continuum and weak features, black hole masses and Eddington ratios—normally measured from broad-line widths—can be estimated from spectra where the lines are corrupted, absorbed, or outside the observed window.
  • The wavelength-based positional encoder offers a concrete architectural path to cross-instrument, cross-resolution transfer: a model trained at R≈2000 generalizes to R≈4500–7500 without retuning, as shown on the X-Shooter part of the blind challenge.
  • The self-supervised pretraining on ~750,000 spectra yields representations that the authors argue will support downstream fine-tuning tasks, including probabilistic prediction of AGN properties and variability studies.
  • For broad-absorption-line quasars, the model—trained without BALs—reconstructs an intrinsic CIV profile and overpredicts flux blueward of the line, a behavior the authors propose as a tool for recovering intrinsic line properties in absorbed sources.
  • Whole-half predictions extend the effective wavelength coverage of a spectrograph, offering a way to homogenize heterogeneous surveys and to flag strong features in unobserved windows for follow-up.

Where Pith is reading between the lines

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

  • A clear testable extension is to train on a second, independent spectroscopic survey with different instrumentation and noise properties, and measure whether the reported in-survey accuracy survives; the current external evidence rests on only 30 X-Shooter spectra compared over a 52 Å window.
  • The error floor of ~2–6% at S/N≈40 suggests the residual error is dominated by physical diversity (e.g., NLR geometry, orientation, absorption) rather than noise; improvements may require multi-epoch or multi-band conditioning, not just more spectra.
  • The per-pixel variance output could itself serve as an anomaly detector: large predicted uncertainty on a cleanly measured flux would flag genuinely ambiguous regions (e.g., absorption troughs or unusual profiles) for follow-up.
  • If the continuum-to-BLR mapping is as tight as claimed, reverberation-mapping campaigns could use the model's prediction as a prior or interpolant between sparse epochs, potentially reducing the number of epochs needed to constrain BLR geometry.

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

3 major / 6 minor

Summary. The paper introduces AGNFormer, an encoder-only transformer trained with a masked-autoencoding objective on SDSS DR16Q rest-frame quasar spectra. The inputs are wavelength, flux, and reported flux uncertainty; two projection heads output a predictive Gaussian mean and variance for each masked pixel, trained with a heteroscedastic negative log-likelihood. Fixed ±10^4 km/s masks around major broad lines, plus half-spectrum masks, are evaluated on an unseen 3% test split. The authors report 10–16% (4–8%) median predicted/observed dispersion for full (S/N>10) broad-line reconstruction, an error floor of ≈2–6% at S/N≈40, calibrated pull distributions, and QMS-resolved reconstructions including FeII complexes and Civ profiles. They also benchmark against the Greig et al. (2024) Lyα reconstruction challenge, reporting superior MAE on 30 X-Shooter and 30 SDSS spectra.

Significance. If the in-survey results hold, the paper provides a useful, well-calibrated self-supervised spectral reconstruction method. Its strengths include a held-out test split, S/N-resolved pull diagnostics, an explicit error-floor fit, an external benchmark comparison, public code and pretrained weights, a formal positional-encoder equivalence proof in Appendix A, and a clear statement that the predictive distribution is an independent Gaussian. However, the strongest external generalization claim rests on the X-Shooter comparison, which is compromised by a likely object-level overlap with the training set, and the quantitative half-spectrum claim in the abstract is not supported by any measurement in Section 5.2. The in-survey broad-line reconstruction numbers are credible, but the cross-resolution and half-spectrum headline claims need substantial revision.

major comments (3)
  1. [§6.1, Table 2] The X-Shooter benchmark likely leaks object identity. The model is 'pretrained on the entire SDSS DR16Q catalogue' and tested on 30 X-Shooter spectra from XQ-100. The text excludes only 'the same 30 SDSS (not present in our training sample)' and does not state that the SDSS spectra of the X-Shooter targets were removed from training. XQ-100 is an SDSS-selected sample at z≈3.5–4.5, so those objects almost certainly have their own SDSS spectra inside DR16Q; if so, the model has seen each test object's continuum and line correlations at SDSS resolution. The Table 2 value MAE=0.017 and the 'resolution invariance' conclusion would then reflect object memorization, not cross-resolution generalization to unseen objects. Please verify object overlap explicitly, rerun with overlapping SDSS spectra excluded, or remove/qualify the X-Shooter-based generalization claim. The in-survey random-split res
  2. [Abstract; §5.2] The abstract and conclusions state that predictions for unseen spectral halves grow to 12–25% (5–15%) of the flux with distance from the cut-off. Section 5.2 shows only four individual examples and says the blue/red ends are too noisy for comparison. No aggregate P/O statistic, figure, or table analogous to Fig. 6 is provided to support this quantitative claim. Please add a quantitative evaluation of the half-mask test set, or revise the abstract/conclusions to match the qualitative evidence actually presented.
  3. [§4 vs §6.2] The BAL exclusion criterion is stated inconsistently. Section 4 says 'exclude any source with P_BAL > 0.0,' while Section 6.2 says 'aside from removing objects with P_BAL > 0.7 from training,' and the Fig. 12 caption states 'BALs were excluded from the training of AGNFormer.' These thresholds define materially different training sets (for example, whether mini-BALs with P_BAL=0.5/0.75 are included). Because Section 6.2 interprets the Civ overprediction as a consequence of BAL exclusion, please correct the threshold and specify the exact criterion in both places.
minor comments (6)
  1. [§3 vs App. B.1] The batch size is given as 16 in §3 but Appendix B.1 concludes 'we chose a batch size of 32.' With 1,000,000 training steps this is a factor-of-two difference in the number of spectra seen; please harmonize.
  2. [Fig. 8 caption] The caption refers to 'validation sources with P_BAL > 0.95,' while the text in §5.1 says the BAL subset is split out from the test sample. Please align the terminology.
  3. [§5.1 vs Conclusions] The full-sample broad-line dispersion is quoted as ≈10–15% in §5.1 but 10–16% in the abstract and conclusions; reconcile the numbers.
  4. [§5.1 vs Conclusions] A pull spread below 1 at mask edges is described in §5.1 as 'the uncertainties may be overestimated by ≈20% and the model is too conservative,' but the Conclusions say the model 'might be slightly overconfident.' If the pull width is <1, the model is underconfident, not overconfident; please fix the wording.
  5. [Abstract vs App. B.3] The abstract quotes an error floor of ≈2–6%, while Appendix B.3 reports individual floors from 1.6±0.2% (Hγ) to 6.5±0.3% (Hα). Align the stated range or describe it as approximate.
  6. [§6.1; §4.1] Please clarify the X-Shooter preprocessing: the only stated adaptation is interpolation to 4500 pixels, but §4.1 says SDSS spectra are not resampled. If this interpolation degrades X-Shooter's native resolution, the 'resolution invariance' claim should be qualified accordingly.

Circularity Check

0 steps flagged

No significant circularity: the masked-reconstruction targets are held-out pixels and the headline benchmark is externally defined.

full rationale

AGNFormer is trained with a masked-autoencoding objective in which pixels in the masked set M are zeroed and the negative log-likelihood (Eqs. 15-16) compares predictions against the observed fluxes f_i for those same pixels. The headline broad-line accuracies are measured on a disjoint 3% test split (70:27:3, Sec. 3), and Sec. 5.1 states that 'None of the spectra in our testing have been previously seen by the model.' Thus the reconstruction claim does not reduce by construction to the training objective. The external comparison uses the Greig et al. (2024) blind challenge, which the authors did not curate, and Table 2 reports parameter-free MAE values against eleven competing models. The only notable caveat is that the X-Shooter portion of that benchmark may involve objects whose SDSS spectra are inside DR16Q ('pretrained on the entire SDSS DR16Q catalogue'), which would weaken the cross-resolution generalization argument; however, that is a data-leakage/validity concern, not a circular-derivation concern: no equation or fitted parameter is being renamed as a prediction. The per-line 'error floors' in App. B.3 come from a two-parameter fit (sigma_model/f_pred = sqrt(a^2 + (b/S/N)^2)) used as a diagnostic of where prediction error dominates, not as a model output. Self-citations (e.g., Rozanski et al. 2024 and the 'in preparation' follow-ups) are incidental related-work references and are not load-bearing for the central claim. No circular step can be exhibited from the paper's own derivations.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claims rest on standard ML training on public data rather than on new physics, so the ledger is short: model/diagnostic hyperparameters and design choices (mask widths, normalization, masking mix), the domain assumptions that SDSS redshifts and P_BAL flags are correct, and the paper's own explicit Gaussian-independence simplification. No invented physical entities. The most consequential unstated premise is that a random split of SDSS DR16Q is a fair proxy for 'unseen' AGN spectra — this sets the scope of every headline percentage.

free parameters (5)
  • Per-region error-floor asymptote a = Oi 4.0±0.1%, Siiv 4.3±0.1%, Civ 4.6±0.1%, Ciii] 3.0±0.1%, Mgii 2.8±0.1%, Hγ 1.6±0.2%, Hβ 3.9±0.1%, Hα 6.5±0.3%
    Fitted in App. B.3 via σ_model/f_pred = sqrt(a² + (b/S/N)²); a diagnostic summary of the model noise floor, not used to derive the headline reconstruction claims.
  • Error-floor nuisance parameter b = not quoted
    Second parameter of the same B.3 fit; absorbs S/N-dependent noise. Diagnostic only.
  • Architecture hyperparameters (d_model=256, layers=8, heads=4, batch=32, 1M steps, 1 register token)
    Chosen via validation-loss experiments (App. B), standard ML practice; not fit to the benchmark data.
  • Mask design: ±10,000 km/s fixed mask width; 35:40:25 random-pixel:random-chunk:fixed masking split; Feii mask 4500-5300
    Design choices. The fixed-mask width defines what 'unseen broad line' means in every reported percentage; the masking mix shapes the learned representations.
  • Selection cuts: log10(L5100)>43.5; P_BAL exclusion threshold (0.0 or 0.7 — stated inconsistently) = 43.5; P_BAL > 0.0 (Sec. 4) vs > 0.7 (Sec. 6.2)
    Luminosity cut removes host-dominated sources; the P_BAL cut removes BALs from training. The inconsistent threshold statement affects interpretation of the BAL test in Sec. 6.2.
axioms (4)
  • domain assumption SDSS DR16Q rest-frame wavelengths and quoted redshifts are accurate enough that spectral realignment and the ±10^4 km/s masks are correctly placed.
    Sec. 4.1: spectra are shifted using Wu & Shen (2022) redshifts, ~90% from the SDSS pipeline. Miscentered masks would bias P/O ratios near line peaks.
  • domain assumption P_BAL classifications (Lyke et al. 2020) reliably identify the BAL QSOs that were excluded from training.
    Sec. 4 footnote 2 notes P_BAL is tuned to HiBALs and less robust for LoBALs; the paper's own BAL overprediction residuals (Sec. 6.2, Fig. 8) show this assumption is imperfect.
  • ad hoc to paper The predictive distribution for each masked pixel is an independent Gaussian.
    Sec. 2.1.6 states this explicitly, noting it cannot capture multimodality (BAL presence/absence) or correlated uncertainty across a line profile; the pull-distribution calibration claims (Fig. 7) depend on it.
  • domain assumption A random 70:27:3 split of SDSS DR16Q is representative of 'unseen' AGN spectra.
    Sec. 3: test spectra share the survey's selection, noise, and calibration systematics with training, so the headline percentages are in-distribution metrics; the X-Shooter test is the only external check.

pith-pipeline@v1.3.0-alltime-deepseek · 26308 in / 21306 out tokens · 226697 ms · 2026-08-01T12:09:13.907696+00:00 · methodology

0 comments
read the original abstract

We explore how an uncertainty-aware transformer-based architecture can leverage information embedded across the entire observed optical spectra of AGN, focusing on the algorithm's ability to predict unseen or masked parts of luminous AGN spectra. This provides a direct probe of the learnable correlations between AGN continua and broad lines. We introduce AGNFormer, a transformer model trained to predict the mean expected flux and variance in masked spectral regions (major broad lines to ${\pm}10^{4}$kms$^{-1}$; missing halves), inputting rest-frame spectral fluxes and uncertainties across the entire redshift range of the SDSS DR16 Quasar Catalogue. We evaluate the performance of the model on both full (no S/N limit) and high-quality (S/N > 10) spectral samples using the negative-log likelihood, and via comparisons with existing C IV and ly-a reconstruction algorithms. The model successfully reconstructs unseen AGN broad lines to better than 10-16% (4-8%) of the flux for the full (S/N > 10) test sets, up to an error floor of $\approx$2-6% of the flux at S/N $\approx$ 40, while predictions for larger unseen halves grow to 12-25% (5-15%) of the flux the further away they are from the cut-off wavelength of the seen input spectrum. Predictions faithfully reproduce the broad AGN spectral diversity across the entire optical and UV QSO main sequence parameter spaces, including both Gaussian and Lorentzian profile regimes, Feii complexes, and narrow emission lines. Performance is similar or better compared to previous spectral reconstruction algorithms. The high precision of the broad-line region reconstruction demonstrates that the method successfully aggregates information across the spectrum and highlights how the AGN continuum and weaker lines/complexes have the potential to assist astronomers in the extraction of the entire wealth of information embedded in AGN spectra.

Figures

Figures reproduced from arXiv: 2607.19640 by Benedict L. Rouse, Christian Wolf, Crist\'obal Rodrigo Donoso Oliva, Ezequiel Treister, Franz E. Bauer, Gordon T. Richards, Guillermo Cabrera-Vives, Patricia Ar\'evalo, Paula S\'anchez-S\'aez, Pavlos Protopapas, Tomasz R\'o\.za\'nski.

Figure 1
Figure 1. Figure 1: Architecture of the probabilistic transformer used in this work. We apply a shared non-linear embedding to the fluxes and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Distribution of redshift vs bolometric luminosity, [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Top: Median spectrum (per wavelength bin) from the en￾tire SDSS DR16Q, with the broad lines analysed in this work highlighted in grey. Bottom: Distribution of the rest-frame wave￾length coverage vs median S/N of the spectrum. Each of the ver￾tical lines denote interesting broad lines that we mask in testing. Sources with lower S/N (<10) and within the rest-frame wave￾length range of 1500–4000Å dominate amo… view at source ↗
Figure 4
Figure 4. Figure 4: Top: Optical-QSO Main Sequence (QMS; also known as EV1) showing RFe ii (EWFe ii/EWHβ) vs. FWHMHβ, adopting the same grid spacing as defined by Sulentic et al. (2002) to split the population. The line profiles of the sources that make up this plot can broadly be modeled by a Lorenztian profile when FWHM < 8000 kms −1 and a Gaussian when FWHM > 8000 kms −1 . Bottom: UV-QMS showing C iv velocity offset (vC iv… view at source ↗
Figure 5
Figure 5. Figure 5: Example DR16Q spectra (observed in black) and spectral reconstructions (mean prediction [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Pull residual (r) distribution plots for each of the broad lines shown in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The plot is identical to Fig [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Median predicted/observed ratios (denoted P/O) on the same scale as Fig.6 for the extended Hβ region for our test sam￾ple, split across bins of Hβ FWHM and RFe ii as highlighted in the QSO MS (top panel of [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of the mean (top) and absolute error (bot￾tom) of the predicted/observed flux ratios from our transformer model (black) vs. 11 models (colours, as indicated among the four legends) from a challenge to predict 30 X-Shooter (left) and 30 SDSS (right) QSO spectra. This figure is a reproduction of [PITH_FULL_IMAGE:figures/full_fig_p012_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Similar to Fig [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Example predictions around the C iv line region from our transformer (red) and the MFICA model of Rankine et al. (2020) (green) for different levels and locations of absorption. when masking broad ranges (1300Å-1650Å), effectively draw￾ing on fainter BAL features present in other lines that were seen by the model. Since C iv BALs are known to correlate with absorption features across a broad range of ioni… view at source ↗

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

273 extracted references · 13 canonical work pages · 8 internal anchors

  1. [1]

    D., Allende Prieto , C., et al

    Alam , S., Albareti , F. D., Allende Prieto , C., et al. 2015, , 219, 12

    2015

  2. [2]

    Alexander , D. M. & Hickox , R. C. 2012, , 56, 93

    2012

  3. [3]

    T., Hewett , P

    Allen , J. T., Hewett , P. C., Richardson , C. T., Ferland , G. J., & Baldwin , J. A. 2013, , 430, 3510

    2013

  4. [4]

    & Auli , M

    Baevski , A. & Auli , M. 2018, arXiv e-prints, arXiv:1809.10853

    Pith/arXiv arXiv 2018

  5. [5]

    2014, arXiv e-prints, arXiv:1409.0473

    Bahdanau , D., Cho , K., & Bengio , Y. 2014, arXiv e-prints, arXiv:1409.0473

    Pith/arXiv arXiv 2014

  6. [6]

    Baldwin , J. A. 1977, , 214, 679

    1977

  7. [7]

    Baldwin , J. A. 1997, in Astronomical Society of the Pacific Conference Series, Vol. 113, IAU Colloq. 159: Emission Lines in Active Galaxies: New Methods and Techniques, ed. B. M. Peterson , F.-Z. Cheng , & A. S. Wilson , 80

    1997

  8. [8]

    A., Burke , W

    Baldwin , J. A., Burke , W. L., Gaskell , C. M., & Wampler , E. J. 1978, , 273, 431

    1978

  9. [9]

    A., Phillips , M

    Baldwin , J. A., Phillips , M. M., & Terlevich , R. 1981, , 93, 5

    1981

  10. [10]

    & Laor , A

    Baskin , A. & Laor , A. 2004, , 350, L31

    2004

  11. [11]

    & Laor , A

    Baskin , A. & Laor , A. 2005, , 358, 1043

    2005

  12. [12]

    2013, , 432, 1525

    Baskin , A., Laor , A., & Hamann , F. 2013, , 432, 1525

    2013

  13. [13]

    2015, , 449, 1593

    Baskin , A., Laor , A., & Hamann , F. 2015, , 449, 1593

    2015

  14. [14]

    H., Fan , X., White , R

    Becker , R. H., Fan , X., White , R. L., et al. 2001, , 122, 2850

    2001

  15. [15]

    S., Schlegel , D

    Bolton , A. S., Schlegel , D. J., Aubourg , \'E ., et al. 2012, , 144, 144

    2012

  16. [16]

    Boroson , T. A. & Green , R. F. 1992, , 80, 109

    1992

  17. [17]

    2024, , 689, A289

    Cabrera-Vives , G., Moreno-Cartagena , D., Astorga , N., et al. 2024, , 689, A289

    2024

  18. [18]

    2025, , 699, A168

    C \'a diz-Leyton , M., Cabrera-Vives , G., Protopapas , P., et al. 2025, , 699, A168

    2025

  19. [19]

    2024, arXiv e-prints, arXiv:2410.18857

    Chun , S., Kim , W., Park , S., & Yun , S. 2024, arXiv e-prints, arXiv:2410.18857

    arXiv 2024

  20. [20]

    C., Banerji , M., & Richards , G

    Coatman , L., Hewett , P. C., Banerji , M., & Richards , G. T. 2016, , 461, 647

    2016

  21. [21]

    2023, arXiv e-prints, arXiv:2309.16588

    Darcet , T., Oquab , M., Mairal , J., & Bojanowski , P. 2023, arXiv e-prints, arXiv:2309.16588

    Pith/arXiv arXiv 2023

  22. [22]

    2018, arXiv e-prints, arXiv:1810.04805

    Devlin , J., Chang , M.-W., Lee , K., & Toutanova , K. 2018, arXiv e-prints, arXiv:1810.04805

    Pith/arXiv arXiv 2018

  23. [23]

    C., et al

    Dietrich , M., Hamann , F., Shields , J. C., et al. 2002, , 581, 912

    2002

  24. [24]

    2023, , 670, A54

    Donoso-Oliva , C., Becker , I., Protopapas , P., et al. 2023, , 670, A54

    2023

  25. [25]

    2020, arXiv e-prints, arXiv:2010.11929

    Dosovitskiy , A., Beyer , L., Kolesnikov , A., et al. 2020, arXiv e-prints, arXiv:2010.11929

    Pith/arXiv arXiv 2020

  26. [26]

    N., Hall , P

    Filiz Ak , N., Brandt , W. N., Hall , P. B., et al. 2014, , 791, 88

    2014

  27. [27]

    B., Protopapas , P., Muthukrishna , D., & Brockmeier , A

    Fox Fortino , W., Bianco , F. B., Protopapas , P., Muthukrishna , D., & Brockmeier , A. 2025, arXiv e-prints, arXiv:2507.22106

    Pith/arXiv arXiv 2025

  28. [28]

    J., Hewett , P

    Francis , P. J., Hewett , P. C., Foltz , C. B., & Chaffee , F. H. 1992, , 398, 476

    1992

  29. [29]

    R., Jiang , L., Brandt , W

    Gibson , R. R., Jiang , L., Brandt , W. N., et al. 2009, , 692, 758

    2009

  30. [30]

    F., Schmidt , M., & Liebert , J

    Green , R. F., Schmidt , M., & Liebert , J. 1986, , 61, 305

    1986

  31. [31]

    Greig , B., Bosman , S. E. I., Davies , F. B., et al. 2024, arXiv e-prints, arXiv:2404.01556

    Pith/arXiv arXiv 2024

  32. [32]

    Gunn , J. E. & Peterson , B. A. 1965, , 142, 1633

    1965

  33. [33]

    Guo , C., Pleiss , G., Sun , Y., & Weinberger , K. Q. 2017, arXiv e-prints, arXiv:1706.04599

    Pith/arXiv arXiv 2017

  34. [34]

    1994, , 432, L95

    Haardt , F., Maraschi , L., & Ghisellini , G. 1994, , 432, L95

    1994

  35. [35]

    2021, arXiv e-prints, arXiv:2111.06377

    He , K., Chen , X., Xie , S., et al. 2021, arXiv e-prints, arXiv:2111.06377

    Pith/arXiv arXiv 2021

  36. [36]

    Heckman , T. M. 1980, , 88, 311

    1980

  37. [37]

    C., Filippenko , A

    Ho , L. C., Filippenko , A. V., & Sargent , W. L. 1995, , 98, 477

    1995

  38. [38]

    C., Filippenko , A

    Ho , L. C., Filippenko , A. V., & Sargent , W. L. W. 1997 a , , 112, 315

    1997

  39. [39]

    C., Filippenko , A

    Ho , L. C., Filippenko , A. V., & Sargent , W. L. W. 1997 b , , 487, 568

    1997

  40. [40]

    2024, arXiv e-prints, arXiv:2404.06395

    Hu , S., Tu , Y., Han , X., et al. 2024, arXiv e-prints, arXiv:2404.06395

    Pith/arXiv arXiv 2024

  41. [41]

    Jacobs, R. A. 1988, Neural Networks, 1, 295

    1988

  42. [42]

    M., Tremonti , C., et al

    Kauffmann , G., Heckman , T. M., Tremonti , C., et al. 2003, , 346, 1055

    2003

  43. [43]

    C., Bechtold , J., Trump , J

    Kelly , B. C., Bechtold , J., Trump , J. R., Vestergaard , M., & Siemiginowska , A. 2008, , 176, 355

    2008

  44. [44]

    J., Dopita , M

    Kewley , L. J., Dopita , M. A., Sutherland , R. S., Heisler , C. A., & Trevena , J. 2001, , 556, 121

    2001

  45. [45]

    J., Groves , B., Kauffmann , G., & Heckman , T

    Kewley , L. J., Groves , B., Kauffmann , G., & Heckman , T. 2006, , 372, 961

    2006

  46. [46]

    & Zetzl , M

    Kollatschny , W. & Zetzl , M. 2013, , 549, A100

    2013

  47. [47]

    2020, arXiv e-prints, arXiv:2008.10546

    Kong , L., Sun , J., & Zhang , C. 2020, arXiv e-prints, arXiv:2008.10546

    Pith/arXiv arXiv 2020

  48. [48]

    2020, arXiv e-prints, arXiv:2002.10118

    Kristiadi , A., Hein , M., & Hennig , P. 2020, arXiv e-prints, arXiv:2002.10118

    Pith/arXiv arXiv 2020

  49. [49]

    Krolik , J. H. & Begelman , M. C. 1988, , 329, 702

    1988

  50. [50]

    2016, arXiv e-prints, arXiv:1612.01474

    Lakshminarayanan , B., Pritzel , A., & Blundell , C. 2016, arXiv e-prints, arXiv:1612.01474

    Pith/arXiv arXiv 2016

  51. [51]

    2023, arXiv e-prints, arXiv:2310.03024

    Lanusse , F., Parker , L., Golkar , S., et al. 2023, arXiv e-prints, arXiv:2310.03024

    Pith/arXiv arXiv 2023

  52. [52]

    Leung , H. W. & Bovy , J. 2024, , 527, 1494

    2024

  53. [53]

    2019, arXiv e-prints, arXiv:1910.13461

    Lewis , M., Liu , Y., Goyal , N., et al. 2019, arXiv e-prints, arXiv:1910.13461

    Pith/arXiv arXiv 2019

  54. [54]

    2021, arXiv e-prints, arXiv:2103.13629

    Li , W., Huang , X., Lu , J., Feng , J., & Zhou , J. 2021, arXiv e-prints, arXiv:2103.13629

    Pith/arXiv arXiv 2021

  55. [55]

    L., et al

    L \'o pez , S., D'Odorico , V., Ellison , S. L., et al. 2016, , 594, A91

    2016

  56. [56]

    & Hutter , F

    Loshchilov , I. & Hutter , F. 2016, arXiv e-prints, arXiv:1608.03983

    Pith/arXiv arXiv 2016

  57. [57]

    & Hutter , F

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

    Pith/arXiv arXiv 2017

  58. [58]

    W., Higley , A

    Lyke , B. W., Higley , A. N., McLane , J. N., et al. 2020, , 250, 8

    2020

  59. [59]

    H., & Rieke , M

    Maiolino , R., Rieke , G. H., & Rieke , M. J. 1996, , 111, 537

    1996

  60. [60]

    2003, The Messenger, 114, 20

    Mayor , M., Pepe , F., Queloz , D., et al. 2003, The Messenger, 114, 20

    2003

  61. [61]

    2018, arXiv e-prints, arXiv:1812.06162

    McCandlish , S., Kaplan , J., Amodei , D., & OpenAI Dota Team . 2018, arXiv e-prints, arXiv:1812.06162

    Pith/arXiv arXiv 2018

  62. [62]

    Mittaz , J. P. D., Penston , M. V., & Snijders , M. A. J. 1990, , 242, 370

    1990

  63. [63]

    2014, arXiv e-prints, arXiv:1406.6247

    Mnih , V., Heess , N., Graves , A., & Kavukcuoglu , K. 2014, arXiv e-prints, arXiv:1406.6247

    Pith/arXiv arXiv 2014

  64. [64]

    & Ferland , G

    Mushotzky , R. & Ferland , G. J. 1984, , 278, 558

    1984

  65. [65]

    M., Nikutta , R., Ivezi \'c , Z ., & Elitzur , M

    Nenkova , M., Sirocky , M. M., Nikutta , R., Ivezi \'c , Z ., & Elitzur , M. 2008, , 685, 160

    2008

  66. [66]

    2014, arXiv e-prints, arXiv:1412.1897

    Nguyen , A., Yosinski , J., & Clune , J. 2014, arXiv e-prints, arXiv:1412.1897

    Pith/arXiv arXiv 2014

  67. [67]

    N., Tavker , S., & Efimov , D

    Oreshkin , B. N., Tavker , S., & Efimov , D. 2025, arXiv e-prints, arXiv:2508.16355

    Pith/arXiv arXiv 2025

  68. [68]

    2018, , 613, A51

    P \^a ris , I., Petitjean , P., Aubourg , \'E ., et al. 2018, , 613, A51

    2018

  69. [69]

    P., et al

    P \^a ris , I., Petitjean , P., Ross , N. P., et al. 2017, , 597, A79

    2017

  70. [70]

    Peterson , B. M. & Wandel , A. 1999, , 521, L95

    1999

  71. [71]

    Portillo , S. K. N., Parejko , J. K., Vergara , J. R., & Connolly , A. J. 2020, , 160, 45

    2020

  72. [72]

    & Ricci , C

    Ramos Almeida , C. & Ricci , C. 2017, Nature Astronomy, 1, 679

    2017

  73. [73]

    L., Hewett , P

    Rankine , A. L., Hewett , P. C., Banerji , M., & Richards , G. T. 2020, , 492, 4553

    2020

  74. [74]

    2026, arXiv e-prints, arXiv:2601.03753

    Rauba , P., Cikojevic , V., Bartolic , F., et al. 2026, arXiv e-prints, arXiv:2601.03753

    arXiv 2026

  75. [75]

    1998, , 36, 267

    Rauch , M. 1998, , 36, 267

    1998

  76. [76]

    2020, , 898, L1

    Ricci , C., Kara , E., Loewenstein , M., et al. 2020, , 898, L1

    2020

  77. [77]

    T., Kruczek , N

    Richards , G. T., Kruczek , N. E., Gallagher , S. C., et al. 2011, , 141, 167

    2011

  78. [78]

    2024, arXiv e-prints, arXiv:2407.05751

    R \'o \.z a \'n ski , T., Ting , Y.-S., & Jab o \'n ska , M. 2024, arXiv e-prints, arXiv:2407.05751

    Pith/arXiv arXiv 2024

  79. [79]

    M., Schneider , P

    Rudolf , N., G \"u nther , H. M., Schneider , P. C., & Schmitt , J. H. M. M. 2016, , 585, A113

    2016

  80. [80]

    2025, arXiv e-prints, arXiv:2501.18965

    Schaipp , F., H \"a gele , A., Taylor , A., Simsekli , U., & Bach , F. 2025, arXiv e-prints, arXiv:2501.18965

    Pith/arXiv arXiv 2025

Showing first 80 references.