AI in the Cosmos

N. Sahakyan

REVIEW 4 major objections 6 minor 18 references

AI in the Cosmos

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Machine learning can classify blazars and fit spectra in milliseconds, and keeping a scientist in the loop makes AI trustworthy, this review argues.

desk verdict A readable but thin review: the HG-AI thesis is asserted, not demonstrated, and the paper works better as an entry-level survey than as a scientific contribution. read the letter →

arxiv 2412.10093 v1 pith:EWHTLCEJ submitted 2024-12-13 astro-ph.HE astro-ph.GAastro-ph.IMcs.AI

N. Sahakyan This is my paper · ORCID

classification astro-ph.HEastro-ph.GAastro-ph.IMcs.AI

keywords ArtificialIntelligenceMachineLearningBlazarclassificationSpectralenergydistributionHuman-GuidedAIGenerativeAstronomicaltechniquesFermi-LAT

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This review argues that artificial intelligence and machine learning have become necessary tools in astrophysics because current surveys produce datasets too large for traditional analysis, and it demonstrates this with concrete applications: classifying gamma-ray blazars and fitting multiwavelength spectra. The paper's central recommendation is Human-Guided AI (HG-AI), in which scientists set objectives, understand their data, and validate outputs, keeping human judgment in control while letting machines do the heavy computation. If the review is right, astrophysicists can use black-box models without losing interpretability, and AI becomes a collaborator that accelerates discovery rather than a source of opaque, unchecked results.

What carries the argument

The argument runs on two machine-learning pipelines. The classification pipeline uses gradient-boosted decision trees and artificial neural networks trained on 18 gamma-ray spectral and temporal features of 2,219 labeled blazars from the Fermi-LAT 4FGL catalog, with 15-fold cross-validation and Bayesian hyperparameter tuning, to assign BL Lac or FSRQ labels to sources of uncertain type. The modeling pipeline trains a convolutional neural network on 200,000 synchrotron self-Compton and 1 million external inverse-Compton spectra generated by a radiative code that solves the Fokker-Planck equation, then couples the network with a nested-sampling optimizer to fit observed spectral energy distributions in milliseconds. Human-Guided AI is the proposed overarching mechanism: scientists understand their datasets, guide AI objectives, and validate outputs so that machine efficiency and human intuition combine.

What would settle it

A head-to-head comparison on a held-out, spectroscopically confirmed sample of blazars, pitting expert-only classification, the gradient-boosted pipeline alone, and an expert-guided HG-AI version of the same pipeline, would settle whether human guidance actually improves outcomes; if it does not, the paper's central recommendation loses its empirical support.

Watch

Extended reading notes

Core claim

The central claim, stated most directly in the discussion of challenges, is that AI systems directed by human intelligence can solve complex astrophysical problems more effectively than either humans or machines alone. The review argues for this through two working demonstrations: a gradient-boosted decision-tree model trained on Fermi-LAT gamma-ray properties that classifies blazar candidates of uncertain type with 88% recall and precision, and a convolutional neural network trained on hundreds of thousands of simulated blazar spectra that reproduces spectral energy distributions from synchrotron self-Compton and external inverse-Compton models and fits real sources like Mrk 421 and CTA 102 in milliseconds. It then proposes that generative AI will extend these gains by helping researchers navigate literature and data, provided that bias, error, and black-box opacity are controlled by human oversight, transparent reporting, and domain-specific adaptation.

Load-bearing premise

The review assumes that the AI applications it features, mostly from the author's own prior studies, are representative demonstrations of AI's value in astrophysics, and that human oversight can reliably catch AI bias, errors, and black-box opacity.

Editorial extensions

If this is right

The fraction of Fermi-LAT blazar candidates of uncertain type can be reduced by machine classification, enabling population studies of BL Lacs and FSRQs.
Neural-network surrogate models make real-time spectral energy distribution fitting practical, cutting computation from seconds or minutes to milliseconds per evaluation.
Generative AI tools can act as research assistants that summarize literature and, as the paper's planned assistant is designed to do, combine data access with modeling capabilities.
Publishing AI-driven results will require releasing data, model architectures, and optimization details so that outputs can be reproduced and audited.
Under HG-AI, researchers can adopt AI without surrendering creativity and reasoning, because the scientist remains the final validator.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Going beyond the paper, HG-AI becomes testable only when specified as an operational protocol, such as active learning or human-in-the-loop corrections; the review gives examples but no such protocol.
The CNN-surrogate strategy should transfer to any source whose emission can be simulated, not just blazars; testing it on pulsar wind nebulae or tidal disruption events would be a natural extension.
Because the flagship examples come from the author's own research group, independent replication on other catalogs is the quickest way to check whether the review's confidence in ML and HG-AI is warranted.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Dear colleague,

The short version: this is a review, not a research paper, and the only thing resembling a thesis—Human-Guided AI—is named but never operationalized or tested. If you're looking for a new result, there isn't one. If you're looking for a readable, high-level survey of ML in astrophysics with a couple of worked examples from the author's own group, it's fine.

What the paper does well: the summary of supervised/unsupervised applications is accurate and accessible. The section on the CNN surrogate for blazar SEDs (Sec. 4) is the most concrete part; it describes a real tool available through MMDC and correctly explains the speed gain and how it works. The discussion of bias, errors, and the black-box problem in Sec. 6 is sensible and appropriately cautious—it even says there is 'no straightforward solution' to the black-box problem.

The soft spots are the ones the stress-test flags, and I think they land. The HG-AI claim in Sec. 6—that human-guided AI can solve complex problems more effectively than AI alone—is asserted, not argued. No worked example in the paper isolates a human-in-the-loop step. The blazar classifier uses human-selected features and labels, but that's standard ML; it doesn't measure whether human guidance improves anything. The SED surrogate is similarly a conventional surrogate model. So the central recommendation isn't supported by the evidence presented, and as written it's unfalsifiable. The review also leans heavily on the author's own prior work (Refs 7, 10, 11, 13, 15); that's not fatal, but the paper doesn't critically evaluate those results, so it reads a bit like a summary of one group's contributions. There are also typos ('straightness' for 'strengthens', 'Feature with Generative AI') and some breathless language about the CNN approach being 'groundbreaking.'

Who's this for? Someone new to the field who wants a quick orientation, or a referee looking for a baseline review. It doesn't advance the science itself. I'd send it to peer review only if the journal treats it as a review contribution and the author is willing to substantially revise—define HG-AI concretely, add at least one example where human guidance is the variable, and tone down the novelty claims. Otherwise it's a desk-reject-or-accept-with-minor-revisions type of manuscript.

If it crosses my desk as a referee, I'd recommend major revision.

Referee Report

4 major / 6 minor

Summary. This review-style manuscript argues that AI/ML methods have become essential in astrophysics and illustrates this with a catalog of applications, two detailed worked examples (blazar classification from Fermi-LAT data, and CNN-based spectral energy distribution fitting for blazars), and a discussion of generative AI. The paper's central conceptual contribution is the proposal of Human-Guided AI (HG-AI), defined as AI systems directed by human intelligence, which the author claims can 'solve complex problems more effectively than either could alone.' The manuscript also surveys challenges (bias, errors, black-box opacity) and offers practical recommendations such as transparency of data and algorithms.

Significance. As a review, the paper provides a useful, readable overview of ML applications in high-energy astrophysics and gives two concrete worked examples that are reproducible and publicly accessible via the MMDC. The explicit discussion of bias, error, and the black-box problem is a helpful contribution, and the idea of human oversight as a guiding principle is a plausible and important research perspective. However, the central HG-AI claim is asserted rather than demonstrated: no operational definition, mechanism, or comparative test is provided, and the worked examples are conventional ML pipelines that do not isolate the effect of human guidance. If revised to frame HG-AI as a testable research hypothesis and to support it with evidence or a clear evaluative framework, the paper could be a valuable perspective piece for the astronomical community.

major comments (4)

[Section 6] The central claim that 'By integrating human intuition and contextual understanding with the data-processing capabilities of AI, HG-AI can solve complex problems more effectively than either could alone' is never operationalized or tested. The manuscript does not define what constitutes a successful HG-AI system, what mechanism produces the claimed synergy, or how the reader could measure the incremental benefit of human guidance. As written, this claim is unfalsifiable. I recommend either providing a precise definition and at least one comparative demonstration, or explicitly reframing HG-AI as a research hypothesis/perspective rather than an established result.
[Sections 3.1 and 4] The two detailed worked examples do not instantiate HG-AI. The LightGBM blazar classifier uses human-selected features and human-labeled training data but has no interactive human-in-the-loop component, and the CNN-MultiNest SED fitting uses a precomputed surrogate without any measurement of whether human oversight improves accuracy, calibration, interpretability, or trust relative to the identical pipeline without human intervention. Consequently, these examples cannot support the Section 6 thesis as written. The paper needs either a demonstration in which human guidance is the manipulated variable, or a clear statement that these examples motivate HG-AI rather than test it.
[Section 6] The manuscript states that 'there is no straightforward solution to the black box problem' and then proposes HG-AI as the solution, but it does not explain the mechanism by which human oversight resolves opacity. The listed requirements—(i) thoroughly understand datasets and (ii) guide AI objectives—do not describe how a researcher can inspect, validate, or interpret the billions of internal states of a trained network. This internal tension needs to be addressed, for example by clarifying that HG-AI aims at interpretability-by-design or at rigorous output validation rather than internal transparency.
[Sections 3-4] The selection of illustrative applications draws heavily on the author's own prior work (Refs. 7, 10, and 11), and these examples are presented without critical comparison to independent implementations, alternative methods, or known failure cases. This is not inherently problematic, but it weakens the evidential basis of the review's recommendations: the reader is asked to accept that these examples are representative of AI's benefits and that their success generalizes. Adding independent examples or a critical discussion of the limitations and failure modes of these tools would substantially strengthen the paper's credibility.

minor comments (6)

[Section 3.1] There is a numerical inconsistency: the text states that the model 'classified 1,493 BCUs' and identifies 825 as BL Lac candidates and 405 as FSRQ candidates while leaving 190 unclassified; 825 + 405 + 190 = 1,420, which matches the right panel of Fig. 2 but not 1,493. Please clarify which number is correct and reconcile the text.
[Section 5] The section heading reads 'F eature with Generative AI' with an erroneous space; it should likely be 'Advances with Generative AI' or 'Features of Generative AI.'
[Section 3.1] The phrase 'showing a similar trend that straightness the robustness of the model' contains a typo; 'straightness' should be 'strengthens' (or 'confirms').
[Introduction] Several sentences are ungrammatical, for example 'advanced comprehensive tools and algorithms' and 'a significant computational resources and time is required'; the prose would benefit from a careful editing pass.
[Section 6] In the sentence 'AGIis a theoretical form of AI', there is a missing space between 'AGI' and 'is'.
[References] Reference [2] appears to have an incomplete volume number: 'AJ 9 (2019) 157' should likely read 'AJ 157 (2019) 157'; please verify all reference details for consistency.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the HG-AI claim is an asserted perspective, not a derived result, and self-cited applications are illustrative rather than load-bearing.

full rationale

This is a review/perspective paper, not a derivation. Section 6 proposes HG-AI as a framework and states that integrating human intuition with AI 'can solve complex problems more effectively than either could alone'; this is presented as an argument/opinion, not as the output of an equation, fit, or uniqueness theorem. The worked examples (Refs. 7, 10, 11) are literature illustrations of ML classification and SED surrogate modeling; they do not reduce to the paper's inputs by construction, and no fitted parameter is relabeled as a prediction. The paper's reliance on the author's own prior work is prominent, but those citations are used as examples of ML applications rather than as load-bearing proof of the HG-AI thesis. The lack of an operational test for HG-AI is an evidentiary weakness, not a circularity under the stated rules. Accordingly, no circular step is identified.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No fitted free parameters appear because the paper contains no original quantitative analysis. The review relies on domain assumptions about data growth and ML utility, plus an unproved assertion that human oversight resolves AI challenges. It does not introduce new physical entities.

assumptions (4)

domain assumption Astronomical data volumes have grown beyond what humans can process manually
Motivates the need for AI/ML; stated in Sec. 1.
domain assumption Supervised and unsupervised learning can extract meaningful patterns from astrophysical datasets
Underlies all cited examples in Sec. 3.
ad hoc to paper Human oversight can mitigate AI bias, errors, and black-box opacity
Central HG-AI thesis presented in Sec. 6 without empirical support.
domain assumption Generative AI can reliably support scientific literature search and summarization
Assumed in Sec. 5, with only qualitative caveats about hallucinations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AI in the Cosmos." pith.science (2026). https://pith.science/paper/EWHTLCEJ

@misc{pith2026241210093,
  author       = {Pith},
  title        = {Pith review of: AI in the Cosmos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWHTLCEJ}},
  note         = {Machine review of arXiv:2412.10093}
}

read the original abstract

Artificial intelligence (AI) is revolutionizing research by enabling the efficient analysis of large datasets and the discovery of hidden patterns. In astrophysics, AI has become essential, transforming the classification of celestial sources, data modeling, and the interpretation of observations. In this review, I highlight examples of AI applications in astrophysics, including source classification, spectral energy distribution modeling, and discuss the advancements achievable through generative AI. However, the use of AI introduces challenges, including biases, errors, and the "black box" nature of AI models, which must be resolved before their application. These issues can be addressed through the concept of Human-Guided AI (HG-AI), which integrates human expertise and domain-specific knowledge into AI applications. This approach aims to ensure that AI is applied in a robust, interpretable, and ethical manner, leading to deeper insights and fostering scientific excellence.

Figures

Figures reproduced from arXiv: 2412.10093 by the authors.

**Figure 2.** Likelihood distributions for sources in the test sample (left panel) and BCUs (right panel), [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Workflow of the method for training a CNN for blazar SED modeling. A subset of param [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: The broadband SEDs of Mrk 421 and CTA 102 are modeled under the SSC and EIC [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

18 extracted references · 17 canonical work pages

[1]

D. G. York, J. Adelman, et al., AJ 120 (2000) 1579. December 16, 2024 1:36 WSPC/INSTRUCTION FILE AI˙in˙Cosmos 14 N. Sahakyan

work page 2000
[2]

Y. Bai, J. Liu, et al., AJ 9 (2019) 157

work page 2019
[3]

Huertas-Company, J

M. Huertas-Company, J. R. Primack, et al., ApJ, 858 (2018) 114

work page 2018
[4]

Pasquet, E

J. Pasquet, E. Bertin, et al., A&A, 621 (2019) A26

work page 2019
[5]

C. J. Shallue, A. Vanderburg, AJ, 155 (2018) 94

work page 2018
[6]

George, E

D. George, E. A. Huerta, PhLB, 778 (2018) 64

work page 2018
[7]

Sahakyan, V

N. Sahakyan, V. Vardanyan, M. Khachatryan, MNRAS, 519 (2023) 3000

work page 2023
[8]

Glauch, T

T. Glauch, T. Kerscher, P. Giommi, A&C, 41 (2022) 100646

work page 2022

Show all 18 references

[9]

Abbasi, et al., Sci, 380 (2023) 1338

Icecube Collaboration, R. Abbasi, et al., Sci, 380 (2023) 1338

2023
[10]

B´ egu´ e, N

D. B´ egu´ e, N. Sahakyan,et al., ApJ, 963 (2024) 71

2024
[11]

Sahakyan, D

N. Sahakyan, D. B´ egu´ e,et al., ApJ, 971 (2024) 70

2024
[12]

Abdollahi, F

S. Abdollahi, F. Acero, et al., ApJS, 260 (2022) 53

2022
[13]

Gasparyan, D

S. Gasparyan, D. B´ egu´ e, N. Sahakyan,MNRAS, 509 (2022) 2102

2022
[14]

Feroz, M

F. Feroz, M. P. Hobson, M. Bridges, MNRAS, 398 (2009) 1601

2009
[15]

Sahakyan, V

N. Sahakyan, V. Vardanyan, et al., AJ, 168, 289 (2024)

2024
[16]

Grezes, S

F. Grezes, S. Blanco-Cuaresma, et al., ASPC, 535 (2024) 119

2024
[17]

Dung Nguyen, Y.-S

T. Dung Nguyen, Y.-S. Ting, et al., arXiv, 2309.06126 (2023)

2023 arXiv
[18]

K. G. Iyer, M. Yunus, et al., arXiv, 2408.01556 (2024)

2024 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

D. G. York, J. Adelman, et al., AJ 120 (2000) 1579. December 16, 2024 1:36 WSPC/INSTRUCTION FILE AI˙in˙Cosmos 14 N. Sahakyan

work page 2000

[2] [2]

Y. Bai, J. Liu, et al., AJ 9 (2019) 157

work page 2019

[3] [3]

Huertas-Company, J

M. Huertas-Company, J. R. Primack, et al., ApJ, 858 (2018) 114

work page 2018

[4] [4]

Pasquet, E

J. Pasquet, E. Bertin, et al., A&A, 621 (2019) A26

work page 2019

[5] [5]

C. J. Shallue, A. Vanderburg, AJ, 155 (2018) 94

work page 2018

[6] [6]

George, E

D. George, E. A. Huerta, PhLB, 778 (2018) 64

work page 2018

[7] [7]

Sahakyan, V

N. Sahakyan, V. Vardanyan, M. Khachatryan, MNRAS, 519 (2023) 3000

work page 2023

[8] [8]

Glauch, T

T. Glauch, T. Kerscher, P. Giommi, A&C, 41 (2022) 100646

work page 2022