Pith. sign in

REVIEW 3 major objections 5 minor 43 references

A controlled study of eight physics research projects finds that mid-2025 AI assistants rarely match expert literature selections and that most AI-supplied references are real papers with corrupted metadata rather than invented ones.

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 01:48 UTC pith:CBLLWQW7

load-bearing objection Useful controlled study; the 64% metadata-mismatch headline needs a normalization sensitivity check before trusting it. the 3 major comments →

arxiv 2607.25672 v1 pith:CBLLWQW7 submitted 2026-07-28 astro-ph.IM astro-ph.COcs.CLgr-qc

AI's Capability in Assisting Scientific Research in Physics, Astrophysics, and Cosmology I: Literature Review

Anamaria Hell , Kateryna Vovk , Veena Krishnaraj , Jia Liu , Kosuke Aizawa , Adrian E. Bayer , Linda Blot , Jessica Cowell
show 13 more authors
Suyog Garg Jonathan Gr\'ee Ben Horowitz Masaya Ichikawa Kanyuni Iemoto Keigo Kondo Zacharie Lorsin Kevin McCarthy Jamie Robinson Miguel Ruiz-Granda Leander Thiele Ievgen Vovk Mingshen Zhou
This is my paper
classification astro-ph.IM astro-ph.COcs.CLgr-qc
keywords literature reviewlarge language modelshallucinationmetadata mismatchreference verificationphysics researchastrophysicscosmology
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 asks whether large language models can help physicists discover relevant literature at the research frontier. It sets up a controlled test: eight expert-designed projects in physics, astrophysics, and cosmology, each searched independently by a human expert and by three mid-2025 AI assistants. The AI produced many references, but almost none overlapped with the experts' choices, and a detailed check of the AI-only references found most were real papers carrying at least one incorrect field, not outright fabrications. A single-project test of a later 2026 model produced only perfect references, suggesting rapid improvement. The takeaway: AI can flag extra papers a human might miss, but its bibliographies need systematic verification.

Core claim

On the paper 's own terms, the central discovery is that the dominant failure mode of LLM-generated bibliographies in physics is not the invention of papers (3%) but the corruption of real references: 64% are papers that exist with at least one wrong title, author, year, journal, DOI, or link field. Among mid-2025 models, ChatGPT Deep Research was the most reliable, producing no fabrications and only 22% mismatches, while Gemini produced the most errors. The paper also reports that a single-project test of ChatGPT Pro 5.5 produced only perfect references, though the authors caution this may reflect stronger tool use rather than a change in the underlying language model alone.

What carries the argument

The method is a controlled, parallel literature search with a standardized prompt. For each of eight expert-conceived projects, a human expert and three LLMs (ChatGPT-4o, ChatGPT Deep Research, and Gemini) independently built bibliographies capped at 50 papers, and experts judged the AI suggestions for relevance. Hallucination was then measured by resolving each AI-only reference's DOI against a bibliographic metadata database, falling back to the provided link or a title search, and classifying each reference as perfect, metadata mismatch, or fabrication. This DOI-first verification chain is the mechanism that separates real-but-corrupted references from invented ones.

Load-bearing premise

The central 64% statistic relies on counting any difference from the metadata database's copy as an error, even though the prompt asked for last-name-only authors; relax that rule and the headline number could change.

What would settle it

Take the 408 references labeled metadata mismatches, normalize author fields to last-name-only, and recompute the mismatch rate while ignoring year and journal formatting differences; if the rate falls below 50%, the '64% require verification' claim would need a strong caveat.

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

If this is right

  • If the 64% mismatch rate holds, LLM-generated reference lists in physics cannot be trusted at face value; each entry must be checked field by field before citation.
  • The less-than-6% overlap with expert selection implies human and AI searches are largely complementary, so a combined human-plus-AI search should find more relevant work than either alone.
  • The strong difference between plain chat models and a tool-augmented research model shows that verification-oriented architectures dramatically reduce hallucination.
  • The single-project zero-error result from ChatGPT Pro 5.5 points to rapid improvement, but the paper itself flags that this is not yet a systematic benchmark.

Where Pith is reading between the lines

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

  • If the recent-model trend generalizes, the near-term role of LLMs in literature review will shift from producing cite-able entries to generating discovery candidates, with humans or retrieval tools verifying metadata.
  • The paper's counting rule treats any field disagreement, however trivial (such as a last-name-only author), as a mismatch; the true rate of practically harmful errors may be lower than 64%.
  • The observation that AIs are keyword-driven while humans search more broadly suggests that prompting for adjacent fields and foundational works could close part of the relevance gap.
  • Comparing against a metadata database rather than the original published version may itself introduce mismatches; a check against the publisher's own records would test that.

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 / 5 minor

Summary. The paper reports a controlled study of LLM-assisted literature review in physics, astrophysics, and cosmology. For eight expert-conceived projects, a human expert and three mid-2025 LLMs (ChatGPT-4o, ChatGPT Deep Research, Gemini) independently produced bibliographies using a standardized prompt. The authors compare human/AI overlap, then classify the 641 AI references not found by humans as perfect, metadata mismatch, or fabrication using a DOI/link/title verification pipeline against Crossref. Main quantitative findings are <6% human-AI overlap, 3% fabrications, and 64% metadata mismatches among analyzed AI references; ChatGPT Deep Research is more reliable, and a single-project ChatGPT Pro 5.5 spot-test yields zero errors. The paper concludes that mid-2025 models require systematic verification of AI-generated references.

Significance. If the quantitative claims hold, the paper provides useful, field-specific evidence on LLM bibliographic reliability in physics: the distinction between fabricated references and real-but-corrupted references is valuable, and the explicit verification pipeline (Fig. 1), per-model tables, and standardized prompt in Appendix B are strengths. The study is non-circular: AI output is compared against an external registry (Crossref), and the definitions are not outcome-dependent. However, the headline 64% mismatch rate and the <6% overlap claim depend on comparison conventions that are not fully specified or tested, so the central quantitative result needs robustness checks before being accepted at face value.

major comments (3)
  1. [§II.E, Appendix B, Table IV] The 64% metadata-mismatch rate is load-bearing and may be inflated by prompt-conformant formatting. Appendix B instructs models to write "only the last name of the first author," while the evaluation compares the generated first author against Crossref's full formatted name. Section II.E counts any "partial or complete" disagreement as a mismatch. If the comparison is raw string equality, every "Smith" vs "Smith, John" is scored as a first-author mismatch, and Table IV reports 306 first-author mismatches among 399 resolved mismatches. The paper does not state whether author names were normalized to surnames, whether journal names were canonicalized, or whether title punctuation/capitalization was normalized. Please specify the exact matching procedure and provide a sensitivity analysis, e.g., recounting mismatches with surname-only author comparison, journal-name canonicalization, and ti
  2. [§II.D, §III.A, Table II] The human-AI overlap definition requires a match in both title and category: a paper placed by the human as "recent" and by the AI as "highly cited" is counted as two different references, and the same paper placed by one model in two categories counts twice. This convention can artificially lower the measured overlap and affects the abstract's "<6%" claim. The authors should report a title-only overlap sensitivity analysis (ignoring category) and state how often human and AI agree on the title but disagree on the category. Without this, the overlap statistic conflates bibliographic coverage with subjective categorization.
  3. [§III.B, Table III, Abstract] The 3% fabrication and 64% mismatch rates are computed on 641 AI references "not found by a human," not on all 701 AI-generated references. The abstract's phrase "of the AI-generated references" is therefore imprecise; the 60 excluded references are a systematically different subset (human-confirmed relevant papers). Please either report the rates on the full 701-reference set or consistently qualify the denominator in the abstract and Section III.B. The magnitude of the effect is likely modest, but the current wording overstates the scope of the measurement.
minor comments (5)
  1. [Table III / text] The table header calls the row "AI-generated references" while the text specifies "those not found by a human." Make the qualifier explicit in the table itself to avoid misreading.
  2. [Section III.B / Fig. 3] The text refers to "Fig. III B" in two places; this should be "Fig. 3." The figure caption also does not mention that ChatGPT Pro 5.5 is a single-project spot-test; add that to the caption.
  3. [Section II.D] The sentence "The overall results are presented in subsection III A" appears twice in slightly different forms; remove the duplicate.
  4. [Section III.C] The handling of near-duplicate references for ChatGPT-4o (5 papers differing only in link or DOI) is described only here. State in Section II how near-duplicates are treated in the main analysis, since the same issue could affect the 701 total.
  5. [Table IV] The rows "1/4 mismatch" through "4/4 mismatch" should explicitly define the denominator as the four compared fields (title, first author, year, journal).

Circularity Check

0 steps flagged

No significant circularity: the central measurements are externally grounded (human expert lists and Crossref) and the only self-citations are contextual.

full rationale

The paper's derivation chain is a controlled measurement rather than a derivation from an assumed conclusion. Eight expert-written project backgrounds are turned into a standardized prompt (Appendix B); human experts and three LLMs independently produce reference lists; the human-vs-AI overlap is scored by title-plus-category matching; AI hallucinations are judged by resolving DOIs/links and comparing title, first author, year, and journal against Crossref entries (Sec. II.E). The 33% perfect / 3% fabrication / 64% metadata-mismatch breakdown is an empirical tally under an explicit classification rule, with no fitted parameter later renamed as a prediction and no equation reducing the result to the input. The only self-citations — [26] (an earlier LLM-physics study by two of the authors, cited as one supporting example) and [27] (the companion paper) — are contextual and do not carry the central claim. The acknowledged limitations (eight projects; single-project Pro 5.5 test) are scope caveats, not circular dependencies. A separate methodological concern is that the mismatch rule may count prompt-conformant forms ('write only the last name of the first author', Appendix B) as partial disagreements against Crossref's full author and journal strings, potentially inflating the 64% figure; however, that is a measurement-validity/sensitivity issue, not circularity, since the mismatch statistic is not constructed from the headline conclusion. Therefore no circular steps are flagged.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central claims rest on these measurement conventions. No numerical parameters were fit; the load-bearing assumptions are about what counts as the reference standard (human lists and Crossref fields) and about the stability of single LLM runs. The 64% mismatch number is the most sensitive to these choices.

axioms (5)
  • domain assumption Human-expert lists of up to 50 'most important' references are a valid reference standard for a competent literature search.
    Used throughout §III.A to define 'found by human', 'missed by AI', and the <6% overlap; the paper notes the selection is subjective (§II.D) but does not test sensitivity to the cap or to expert bias.
  • domain assumption Crossref metadata is the ground truth for judging AI field accuracy, and any partial disagreement in title, first author, year, or journal counts as a mismatch.
    The hallucination audit (§II.E, Table IV) uses this rule; the standardized prompt (Appendix B) asked for last name of the first author only, so strict string comparison may inflate the 64% mismatch rate.
  • ad hoc to paper The same paper placed in two different categories counts as two references in the overlap statistics.
    This convention (§II.D, §III.A) changes union sizes and overlap counts; no sensitivity check is reported.
  • domain assumption Each LLM's output in mid-2025 is deterministic enough that running it once represents that model's performance.
    Single runs per project/model (Table II); LLMs are stochastic and the temperature/seed are not reported, so the counts have run-to-run variability.
  • domain assumption The eight expert-conceived projects span representative frontier physics/astro/cosmology work.
    The paper itself calls the result 'a snapshot in time' (§IV); generalizability beyond these eight projects is assumed, not established.

pith-pipeline@v1.3.0-alltime-deepseek · 14573 in / 17358 out tokens · 168749 ms · 2026-08-01T01:48:36.015823+00:00 · methodology

0 comments
read the original abstract

We investigate how well large language models (LLMs) can assist with literature reviews for scientific research. We perform a controlled study of eight expert-conceived research projects across the areas of physics, astrophysics, and cosmology. Each project has a defined background and goal, and human experts and AI prompters are asked to perform identical literature review tasks in parallel. We compare the relevant literature selected by humans with that selected by mid-2025 LLMs (ChatGPT-4o, ChatGPT Deep Research, and Gemini). We find the overlap between human- and AI-selected references to be small ($<$6\%), indicating that AI models do not yet reproduce a competent expert search on their own, though they have the potential to complement literature searches by humans. We then assess the reliability and completeness of AI-generated candidate references, distinguishing two types of hallucination: fabrications (references to nonexistent papers) and metadata mismatches (real papers with one or more incorrect fields). We find that while fabricated references make up 3\% of the AI-generated references, 64\% are real papers with at least one incorrect field (title, author, year, journal, DOI, or link), indicating that the mid-2025 models require systematic verification. However, the performance is significantly improved for the 2026 model ChatGPT Pro 5.5, with a single-project test showing zero fabrication or metadata mismatches.

Figures

Figures reproduced from arXiv: 2607.25672 by Adrian E. Bayer, Anamaria Hell, Ben Horowitz, Ievgen Vovk, Jamie Robinson, Jessica Cowell, Jia Liu, Jonathan Gr\'ee, Kanyuni Iemoto, Kateryna Vovk, Keigo Kondo, Kevin McCarthy, Kosuke Aizawa, Leander Thiele, Linda Blot, Masaya Ichikawa, Miguel Ruiz-Granda, Mingshen Zhou, Suyog Garg, Veena Krishnaraj, Zacharie Lorsin.

Figure 1
Figure 1. Figure 1: FIG. 1. The hallucination-evaluation procedure. Each AI [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Composition of the combined human+AI reference set for each model, split by literature type. Within each literature [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Hallucination breakdown per model: number of AI [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] 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

43 extracted references · 14 linked inside Pith

  1. [1]

    Both AI and human experts generated a large set of relevant references that were essentially not over- lapping, indicating that a collaboration between humans and AI may yield a more complete refer- ence overview

  2. [2]

    For the mid-2025 models, 3% of the references are fabrications (nonexistent papers), while 64% are real papers with at least one incorrect field (a meta- data mismatch); both are forms of hallucination. This is one of the main weaknesses of AI, which has persisted in recent years in scientific literature searches, and suggests that one should carefully ch...

    2025

  3. [3]

    Our findings also come with two important caveats

    AIs appear to prefer more specialized search guided by the project keywords, focused narrowly on the project topic, while human experts appear to search more broadly, taking into account also re- lated fields or more fundamental papers. Our findings also come with two important caveats. First, we have limited our investigation to only eight projects, whil...

    2020

  4. [4]

    Woesle, L

    C. Woesle, L. Fischer-Brandies, and R. Buettner, IEEE Access13, 148231 (2025)

    2025

  5. [5]

    W. H. Walters and E. I. Wilder, Sci Rep13, 14045 (2023)

    2023

  6. [6]

    Scherbakov, N

    D. Scherbakov, N. Hubig, V. Jansari, A. Bakumenko, and L. A. Lenert, Journal of the American Medical Informat- ics Association32, 1071–1086 (2025)

    2025

  7. [7]

    Liang, Y

    W. Liang, Y. Zhang, H. Cao, B. Wang, D. Y. Ding, X. Yang, K. Vodrahalli, S. He, D. S. Smith, Y. Yin, D. A. McFarland, and J. Zou, NEJM AI1, 10.1056/AIoa2400196 (2024), arXiv:2310.01783

    Pith/arXiv arXiv 2024

  8. [8]

    M. Wang, R. Lin, K. Hu, J. Jiao, N. Chowd- hury, E. Chang, and T. Patwardhan, arXiv e-prints , arXiv:2601.21165 (2026), arXiv:2601.21165 [cs.AI]

    arXiv 2026

  9. [9]

    Villaescusa-Navarro, D

    F. Villaescusa-Navarro, D. Angl´ es-Alc´ azar, S. Genel, D. N. Spergel, R. S. Somerville, R. Dave, A. Pillepich, L. Hernquist, D. Nelson, P. Torrey,et al., The Astro- physical Journal915, 71 (2021)

    2021

  10. [10]

    T. D. Nguyen, Y.-S. Ting, I. Ciuc˘ a, C. O’Neill, Z.-C. Sun, M. Jab lo´ nska, S. Kruk, E. Perkowski, J. Miller, J. Li,et al., inProceedings of the Second Work- shop on Information Extraction from Scientific Publica- tions (WIESP), IJCNLP-AACL 2023(2023) pp. 49–55, arXiv:2309.06126

    Pith/arXiv arXiv 2023

  11. [11]

    A. L. Franzoni Vel´ azquez, E. Huerta, and S. Jensen, Dis- cov Educ 3226, 226 (2024)

    2024

  12. [12]

    M. e. a. Pastucha, Medical science monitor : interna- tional medical journal of experimental and clinical re- search32, e950916 (2026)

    2026

  13. [13]

    Y. D. Hezaveh, L. Perreault Levasseur, and P. J. Mar- shall, Nature548, 555 (2017)

    2017

  14. [14]

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, and D. Ha, arXiv e-prints (2024), arXiv:2408.06292

    Pith/arXiv arXiv 2024

  15. [15]

    Shcherbiak, H

    A. Shcherbiak, H. Habibnia, R. B¨ ohm, and S. Fiedler, Judgment and Decision Making19, e21 (2024)

    2024

  16. [16]

    H. Zhou, H. Huang, Y. Long, B. Xu, C. Zhu, H. Cao, M. Yang, and T. Zhao, inProceedings of the 23rd Chi- nese National Conference on Computational Linguistics (CCL)(2024) pp. 1310–1319, arXiv:2409.16788

    Pith/arXiv arXiv 2024

  17. [17]

    Mitchell, Y

    E. Mitchell, Y. Lee, A. Khazatsky, C. D. Manning, and C. Finn, inProceedings of the 40th International Con- ference on Machine Learning (ICML), PMLR, Vol. 202 (2023) pp. 24950–24962, arXiv:2301.11305

    Pith/arXiv arXiv 2023

  18. [18]

    C. A. Gao, F. M. Howard, N. S. Markov, E. C. Dyer, S. Ramesh, Y. Luo, and A. T. Pearson, npj Digital Medicine6, 75 (2023)

    2023

  19. [19]

    V. S. Sadasivan, A. Kumar, S. Balasubramanian, W. Wang, and S. Feizi, arXiv e-prints (2023), arXiv:2303.11156

    Pith/arXiv arXiv 2023

  20. [20]

    Ren and W

    J. Ren and W. Wang, arXiv e-prints (2025), arXiv:2509.09709

    Pith/arXiv arXiv 2025

  21. [21]

    Panickssery, S

    A. Panickssery, S. R. Bowman, and S. Feng, inAdvances in Neural Information Processing Systems (NeurIPS), Vol. 37 (2024) arXiv:2404.13076

    Pith/arXiv arXiv 2024

  22. [22]

    C. Si, D. Yang, and T. Hashimoto, inInternational Conference on Learning Representations (ICLR)(2025) 10 arXiv:2409.04109

    Pith/arXiv arXiv 2025

  23. [23]

    Wataoka, T

    K. Wataoka, T. Takahashi, and R. Ri, arXiv e-prints (2024), arXiv:2410.21819. Presented at the NeurIPS 2024 Safe Generative AI Workshop

    Pith/arXiv arXiv 2024

  24. [24]

    Villaescusa-Navarro, B

    F. Villaescusa-Navarro, B. Bolliet, P. Villanueva- Domingo, A. E. Bayer, A. Acquah, C. Amancharla, A. Barzilay-Siegal, P. Bermejo, C. Bilodeau, P. C. Ram ´ ırez, M. Cranmer, U. L. Fran¸ ca, C. Hahn, Y.- F. Jiang, R. Jimenez, J.-Y. Lee, A. Lerario, O. Ma- mun, T. Meier, A. A. Ojha, P. Protopapas, S. Roy, D. N. Spergel, P. Taranc´ on-´Alvarez, U. Tiwari, M....

    arXiv 2025

  25. [25]

    Miaoet al., PRL-Bench: A Comprehensive Bench- mark Evaluating LLMs’ Capabilities in Frontier Physics Research (2026), arXiv:2604.15411 [cs.LG]

    T. Miaoet al., PRL-Bench: A Comprehensive Bench- mark Evaluating LLMs’ Capabilities in Frontier Physics Research (2026), arXiv:2604.15411 [cs.LG]

    Pith/arXiv arXiv 2026

  26. [26]

    Sikimi´ c, Synthese206, 282 (2025)

    V. Sikimi´ c, Synthese206, 282 (2025)

    2025

  27. [27]

    Sandstr¨ om and M

    U. Sandstr¨ om and M. Thelwall, arXiv e-prints (2026), arXiv:2603.14565

    arXiv 2026

  28. [28]

    Thorne, J

    W. Thorne, J. James, and Y. Wang, arXiv e-prints (2026), arXiv:2603.08281

    arXiv 2026

  29. [29]

    Hell, JHEP03, 167, arXiv:2111.00017 [hep-th]

    A. Hell, JHEP03, 167, arXiv:2111.00017 [hep-th]. Appendix A: Background and goals of the eight research projects The following project titles, backgrounds, and goals were written by the human experts without any AI as- sistance, and were the common input provided to the human planners and the AI prompters for each project (Sec. II A)

    Pith/arXiv arXiv

  30. [31]

    Hell and L

    A. Hell and L. Thiele, LLMs with in-context learn- ing for Algorithmic Theoretical Physics (2026), arXiv:2605.08212 [cs.LG]

    Pith/arXiv arXiv 2026

  31. [32]

    J. Liu, V. Krishnaraj, K. Vovk, , K. Aizawa, A. E. Bayer, L. Blot, J. Cowell, S. Garg, J. Gr´ ee, A. Hell, B. Horowitz, M. Ichikawa, K. Iemoto, K. Kondo, Z. Lorsin, K. McCarthy, J. Robinson, M. Ruiz-Granda, L. Thiele, I. Vovk, and M. Zhou, AI’s Capability in As- sisting Scientific Research in Physics, Astrophysics, and Cosmology II: Project Planning and P...

  32. [33]

    A. E. Bayer, Y. Zhong, Z. Li, J. DeRose, Y. Feng, and J. Liu, JCAP05, 016, arXiv:2407.17462 [astro-ph.CO]

    Pith/arXiv arXiv

  33. [35]

    AGN – MaNGA: AGN Duty Cycle Background:The time a galaxy spends in the AGN phase, from both general arguments and ensemble studies such as quasar clustering and black-hole mass-function studies and Heiiproximity-zone analysis, is suggested to last∼10 6–109 yr. Ionization studies of AGN host galax- ies and their surroundings indicate that active nuclei can...

  34. [36]

    drop-out

    LBG – The galaxy–dark matter halo connection of Lyman-break galaxies Background:A Lyman-break galaxy (LBG) is a galaxy whose broadband photometry shows a “drop-out” in the bluest bands as features in its spectrum move from blue to red through the filter set due to cosmic expansion; the reduction in flux blueward of the Lyman-αand Lyman- limit frequencies ...

  35. [37]

    Many studies have therefore investigated the characteristics of 11 IA in order to eliminate it from the data

    IA – Intrinsic alignments in varying environments Background:Weak-lensing surveys are one of the most powerful probes in cosmology; however, the intrinsic alignment (IA) of galaxies contaminates the signal. Many studies have therefore investigated the characteristics of 11 IA in order to eliminate it from the data. So far, re- searchers believe IA is rela...

  36. [38]

    AR – Prediction of debris emergence on laser-ablated sub-wavelength shapes Background:We have been developing methods to fabricate sub-wavelength structures (SWS) for anti- reflective coating in the millimeter-wave region on hard materials such as ceramics, using ultra-short-pulse laser ablation, which is crucial for machining materials with relatively wi...

  37. [39]

    RG – Radio Galaxies with HalfDome Background:At low CMB frequencies (around 100 GHz), high-energy radio galaxies act as bright point- source contaminants to CMB maps. The locations of these galaxies are likely correlated with features in the underlying large-scale structure as well as with galaxy properties (e.g., the CIB, radio continuum, X-ray). Goal:Ad...

  38. [40]

    However, the origins of these binary black holes and the environments they reside in remain unknown

    GW – Environment of gravitational-wave black hole binaries with weak-lensing maps Background:The first detection of a gravitational wave (GW) by LIGO opened a new era of multi- messenger astronomy, and around 300 GW events from binary black hole (BBH) mergers have now been ob- served. However, the origins of these binary black holes and the environments t...

  39. [41]

    PT A – F orecasting pulsar timing array sensitivity to deviations from general relativity Background:The Pulsar Timing Array (PTA) is a measurement method relying on the observation of pul- sars, fast-rotating neutron stars with well-known timing models. By measuring slight perturbations in the times of arrival (ToAs) of each pulse, computing residuals, a...

    2023

  40. [42]

    For instance, introducing a Chern–Simons cou- pling between a pseudo-scalar field and a non-Abelian gauge field can lead to slow-roll inflation, as in Chromo- 12 Natural Inflation

    SU2 – Massive Y ang–Mills theory Background:When exploring mechanisms that drive inflation, non-Abelian gauge fields – such as SU(2) Yang– Mills fields – have been proposed as alternatives to scalar fields. For instance, introducing a Chern–Simons cou- pling between a pseudo-scalar field and a non-Abelian gauge field can lead to slow-roll inflation, as in...

  41. [43]

    Author” write only the last name of the first author of the paper. “Year

    Recent papers/results on this topic (past 10 years) 4. Other papers that may be relevant The total number of papers you find (all 4 categories combined) should be no more than 50 papers. Limit to refereed papers only. As for the number of papers in each individual category it is up to you to determine what is appropriate. I want you to for each of the 4 c...

  42. [641]

    Human & AI

    One should note that this number is smaller than in the previous subsection because here we do not dis- tinguish between the different categories into which the same paper may have been placed. Following the verifica- tion procedure of subsection II E (summarized in Fig. 1), each reference is classified as aperfectreference (all fields match the true pape...

    2025

  43. [2025]

    drop-out

    – supplemented with project-specific context. The relevance of the AI-generated candidate references was then evaluated by an expert. The overall material was also compared to human output and further analyzed to assess reliability and possible hallucinations. Finally, the literature search was repeated for one of the projects us- ing ChatGPT Pro 5.5 to s...

    2025