Pith. sign in

REVIEW 4 major objections 4 minor 10 cited by

DeepInflation claims that an LLM-orchestrated agent with symbolic regression can discover simple single-field slow-roll inflationary potentials matching any requested (ns, r) target, demonstrated by V(phi) = exp(-0.42214/phi), which gives n

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

2026-08-03 10:36 UTC pith:UPYAQPTL

load-bearing objection A genuine LLM+SR+RAG prototype for inflationary model discovery, with one sound worked example but an unsupported 'any given' claim and fitting-framed-as-prediction. the 4 major comments →

arxiv 2601.14288 v2 pith:UPYAQPTL submitted 2026-01-14 astro-ph.CO cs.AIcs.CEgr-qchep-th

DeepInflation: an AI agent for research and model discovery of inflation

Ze-Yu Peng , Hao-Shi Yuan , Qi Lai , Jun-Qian Jiang , Gen Ye , Jun Zhang , Yun-Song Piao This is my paper
classification astro-ph.CO cs.AIcs.CEgr-qchep-th
keywords inflationary cosmologysymbolic regressionlarge language modelsslow-roll inflationCMB observablesmodel discoveryretrieval-augmented generationtensor-to-scalar ratio
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.

This paper tries to establish that the search over inflationary potentials can be automated end to end: a user states a target for the scalar spectral index ns and the tensor-to-scalar ratio r, and an AI agent returns a short, viable single-field slow-roll potential whose predicted observables hit the target. The demonstration targets the latest CMB preference for ns ≈ 0.974 and small r, and finds V(phi) = exp(-0.42214/phi), predicting ns ≈ 0.975 and r ≈ 0.005 at N = 55. If this works generally, it gives cosmologists and non-specialists a natural-language tool for exploring model space and checking new proposals against data. The value is mostly methodological: it combines language understanding, literature grounding, and numerical physics verification into one loop.

Core claim

On the paper's own terms, the central discovery is that a multi-agent system can discover and verify inflationary potentials without a human prescribing a functional form. The agent converts the request into a symbolic-regression search using a small operator set and a physics loss that solves the slow-roll background (end of inflation at epsilon = 1, N = 55 e-folds back) and scores each candidate by chi-squared distance in the (ns, r) plane. The flagship result is the very simple plateau-like potential V(phi) = exp(-0.42214/phi), whose predictions ns ≈ 0.975 and r ≈ 0.005 sit within the new CMB constraints; at large phi it behaves as 1 - 0.42214/phi, placing it in the known inverse-power br

What carries the argument

The load-bearing mechanism is the coupling of an LLM orchestrator to a symbolic-regression engine whose fitness function is not a generic data fit but the actual inflationary equations: for each trial potential V(phi), the kernel solves epsilon(phi_end) = 1, integrates the field back N e-folds, evaluates ns = 1 - 6 epsilon + 2 eta and r = 16 epsilon, and compares to the requested target with a chi-squared loss. Simplicity is encouraged by restricting the candidate operators, while literature grounding comes from a retrieval-augmented knowledge base built from a catalogue of known inflation models. The agent is what makes the search askable: it turns natural-language constraints into regressi

Load-bearing premise

For any requested (ns, r) pair, a short formula built from the allowed operators exists whose slow-roll prediction at 55 e-folds lands on that pair; only one pair has been tested, so this coverage is assumed rather than shown.

What would settle it

Run the same discovery request on a grid of (ns, r) pairs spanning the full allowed range, for instance ns from 0.90 to 1.00 and r from 0.001 to 0.1, and record whether a simple candidate under the agent's complexity threshold is returned for every pair with predictions inside target errors. Any requested but unreturned pair, or a returned potential whose computed (ns, r) miss the target beyond the quoted uncertainty, would falsify the 'any given ns and r' claim.

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

If this is right

  • A researcher can ask for a potential matching any given ns and r and receive, in minutes, a short candidate expression plus its predicted observables at N = 50, 55, and 60.
  • Discovered expressions come with physical interpretation: the flagship potential is identified as an inverse-power plateau of the brane/pole family, connecting automated discovery to a known model class.
  • The knowledge base lets the agent answer targeted questions about obscure inflation models with the correct potential and symmetry reasoning, not just fluent text.
  • The same architecture can be pointed at new targets by swapping the loss function, and the paper names small-scale observables for primordial black holes as the next step.
  • Non-experts can run model-discovery tasks in natural language, lowering the entry barrier to exploring the inflationary landscape.

Where Pith is reading between the lines

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

  • Because only a single (ns, r) target is demonstrated, the strongest reading of 'any given' is an extrapolation; a grid scan over the observable plane would turn this into a testable claim.
  • The loss relies on leading-order slow-roll formulas and a fixed N = 55; predictions for the same potential could shift under full numerical mode evolution or if N is treated as free, so the candidate list may shrink under exact analysis.
  • The search is confined to the allowed operator set and to single-field slow-roll inflation; models outside this class, such as multi-field or non-canonical scenarios, are outside the claimed capability.
  • If the approach generalizes, it suggests a workflow for other theory-vs-data problems with fast forward models: let the LLM translate constraints, let symbolic regression propose formulas, and let the physics kernel veto.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper presents DeepInflation, a multi-agent LLM system that combines a PySR symbolic-regression engine, a Julia-based physics kernel, and a RAG knowledge base built from the Encyclopædia Inflationaris. The agent discovers single-field slow-roll potentials by minimizing a χ² loss in the (n_s, r) plane (Eq. 3). The central demonstration is one target, n_s = 0.974 ± 0.003, r = 0 ± 0.018, for which Candidate A, V(φ) = exp(−0.42214/φ), yields n_s = 0.97526 and r = 0.00497 at N = 55. A second example shows retrieval of pseudo-natural inflation from the knowledge base. The paper claims in the abstract that DeepInflation can discover simple potentials for 'any given n_s and r', and that the discovered models are viable single-field slow-roll potentials consistent with the latest observations.

Significance. If the central claim is defensible, the paper would be a useful demonstration of an AI agent integrating LLM planning, symbolic regression, and a domain-specific physics kernel. The concrete Candidate A computation is checkable and appears sound, and the code is posted on GitHub; these are strengths. The incremental novelty over existing SR/LLM tooling lies in the inflationary physics kernel and the RAG grounding. However, the evidence is far thinner than the conclusion: only one target is demonstrated, the 'any given n_s and r' universal claim has no coverage analysis, the loss function makes the final outputs a fit rather than an out-of-sample prediction, and the representation of the ACT DR6 likelihood used as the target is an ad hoc Gaussian simplification. The paper is best viewed as a prototype presentation; the claims need to be either substantially weakened or supported with additional experiments.

major comments (4)
  1. [Abstract and Sec. III.A] The abstract claims discovery for 'any given n_s and r', but Sec. III.A demonstrates exactly one target: n_s = 0.974 ± 0.003, r = 0 ± 0.018, with N fixed at 55. No argument or experiment establishes that the allowed operator set {+, −, *, ^, /, exp, log, sqrt} together with the simplicity preference contains a sufficiently simple potential for arbitrary points in the observationally relevant n_s–r plane. The N dependence alone undercuts the universal claim: Candidate A gives n_s = 0.97271 at N = 50 and 0.97725 at N = 60, a spread comparable to the quoted σ_ns = 0.003. The authors should either remove 'any given n_s and r' or provide a systematic coverage study over the target plane, including success rates and complexity bounds, and state how N is selected or marginalized.
  2. [Sec. III.A, Eq. (3)] The loss function in Eq. (3) defines success as closeness in (n_s, r) to user-supplied targets, and the final response then quotes the same computed (n_s, r) values as 'predictions' and labels the model 'viable'. This is a fit to the input, not an out-of-sample test. The 'analyze potential' tool uses the same physics kernel as the loss, so it cannot provide independent verification. To support 'viable' or 'discovery', the authors should check observables not used in the loss — for example, the running of the spectral index, the full slow-roll trajectory, or robustness to reheating/N variation — or explicitly frame the outputs as reconstructions consistent with the input constraints.
  3. [Sec. III.A, Eq. (3)] The ACT DR6 constraints are represented as independent Gaussians n_s = 0.974 ± 0.003 and r = 0.0 ± 0.018. This is not a faithful representation of the ACT DR6 likelihood: the actual constraints are non-Gaussian and correlated, and a central value r = 0 for the tensor-to-scalar ratio is not the data. Since the paper claims consistency with 'the latest observations', this target choice needs a citation and justification, or an explicit disclaimer that the target is illustrative. Otherwise the 'viable' claim is not anchored to the stated dataset.
  4. [Sec. IV] The limitations section acknowledges that unconstrained symbolic regression is prone to combinatorial explosion, but it does not address the coverage problem for the universal 'any given n_s and r' assertion. The manuscript contains no discussion of when the search might fail — e.g., for extreme targets, for targets requiring high complexity, or for targets that are physically unreachable under single-field slow roll at the fixed N. This is a missing support for the paper's principal claim and should be added alongside the existing limitations.
minor comments (4)
  1. [Sec. III.A (transcript)] The user prompt only specifies n_s = 0.974 ± 0.003, but the SR sub-agent logs 'Targets: n_s = 0.974 ± 0.003, r = 0.0 ± 0.018, N = 55'. The r and N values appear without justification; clarify whether these were hidden in the full prompt, chosen by the main agent, or hard-coded by the SR sub-agent.
  2. [Sec. III.B] Typo: 'repect' should be 'respect'. Also, the final response for pseudo-natural inflation should include the original citations [64,65] in the answer, not only the Encyclopædia source, to make the retrieved context traceable.
  3. [Sec. II.A and Sec. III.A] The slow-roll observables are computed using the approximations of Eq. (2). For the demonstrated candidates the slow-roll parameters are small, but the paper should state the validation tolerance and, if relevant, confirm with the full mode equation or a higher-order slow-roll expansion that the quoted n_s and r values are accurate at the quoted precision.
  4. [Fig. 2] Figure 2 is described as a diagnostic plot generated by the agent, but no visual content appears in the manuscript text. In the published version, ensure the figure is actually included and readable.

Circularity Check

2 steps flagged

The reported 'predictions' of ns and r are the same quantities minimized by Eq. (3), so the claimed discovery is a fitted target restated as a prediction, with verification using the same loss function.

specific steps
  1. fitted input called prediction [Sec. III.A, Eq. (3) and Final Response]
    "3. Custom Loss Function: The fitness of the candidate potential is then evaluated using the χ2 distance in the (ns, r) plane based on the target observational constraints: 2L=... Final Response: Here are two very simple single-field slow-roll potentials that hit your target ns=0.974±0.003 and also give safely small tensors. ... Predictions: – N=55: ns=0.97526, r=0.00497"

    The SR engine selects candidate potentials by minimizing exactly this χ2 distance between their computed (ns, r) and the user-specified target. The final response then reports the computed (ns, r) as 'Predictions' and celebrates that they 'hit your target.' But a candidate that survives evolutionary search is chosen for minimizing that very distance; its reported values are the optimized objective, not an independent prediction. No out-of-sample observable (e.g., amplitude As, running αs, or a full Planck/BICEP/Keck likelihood) is checked, so the claim 'consistent with observations' reduces to restating the fitted target.

  2. self definitional [Sec. II.C, Analysis and Visualization Tools]
    "analyze potential: Computes observables for a given potential expression. This tool follows the same procedure as the loss function in the SR engine, allowing the agent to verify the predictions of discovered models or user-specified potentials."

    The verification tool is explicitly defined to use the same procedure as the loss function in the SR engine. Thus when the agent 'verifies' a discovered candidate, it re-runs the same physics kernel and re-evaluates the same (ns, r) objective that already selected the candidate. The diagnostic plot overlays the same ns–r contours used to set the target, so it cannot falsify the match. The claimed verification is tautological: the check is identical to the fitting criterion, so success is guaranteed by the selection process.

full rationale

The paper is not circular in the usual self-citation or imported-uniqueness sense: Eqs. (1)–(2) are standard slow-roll relations, the SR engine genuinely searches a space of functional expressions, and no load-bearing result is imported solely from the authors' prior work. The circularity is concentrated in the validation loop. Eq. (3) defines the loss as a χ2 distance from the target (ns, r); PySR selects potentials by minimizing this loss; the final response reports the same computed (ns, r) as 'Predictions'; and the analyzer is described as using 'the same procedure as the loss function in the SR engine.' Consequently, the demonstrated 'discovery consistent with latest observations' is the optimized fitting objective restated as a prediction, with no independent observable checked. The universal 'any given ns and r' claim is a coverage overgeneralization rather than a circularity: the paper demonstrates only one target and provides no argument that the operator set covers the full ns–r plane; that is a correctness risk, not a definitional tautology. Overall score 6: a central part of the claim reduces by construction, while the architecture, physics kernel, and RAG literature components retain independent content.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The model search relies on standard slow-roll assumptions and a hand-chosen e-fold number, plus four SR-fitted constants. No new physical entity is introduced. The main burden is that target matching is built into the loss function.

free parameters (4)
  • Candidate A coefficient a = 0.42214
    Chosen by SR to minimize the chi-squared loss against the target (ns, r) in Eq. (3); V = exp(-a/phi).
  • Candidate B coefficient b = 0.00286515
    SR-fitted coefficient in V = exp(-b * phi^{-1.50769}).
  • Candidate B exponent p = 1.50769
    SR-fitted exponent in V = exp(-b * phi^{-p}); p is non-integer and was not fixed a priori.
  • Number of e-folds N = 55
    Hand-chosen input for the search; the final table shows N = 50, 55, 60, and the match to target depends on this choice.
axioms (3)
  • domain assumption Single-field slow-roll equations of motion and observables (Eqs. 1-2)
    All inflation calculations use the slow-roll approximation; multi-field, non-slow-roll, and NEC-violating models are excluded by design.
  • domain assumption Inflation ends when epsilon(phi_end) = 1
    Standard slow-roll end condition used to set integration limits.
  • ad hoc to paper ACT DR6 constraint represented as independent Gaussian targets ns = 0.974 +/- 0.003 and r = 0 +/- 0.018
    The paper does not use the full covariance or actual likelihood; adopted in Eq. (3) for the demo.

pith-pipeline@v1.3.0-alltime-deepseek · 9753 in / 14697 out tokens · 136688 ms · 2026-08-03T10:36:56.127688+00:00 · methodology

0 comments
read the original abstract

We present DeepInflation, an AI agent designed for research and model discovery in inflationary cosmology. Built upon a multi-agent architecture, DeepInflation integrates Large Language Models (LLMs) with a symbolic regression (SR) engine and a retrieval-augmented generation (RAG) knowledge base. This framework enables the agent to automatically explore and verify the vast landscape of inflationary potentials while grounding its outputs in established theoretical literature. We demonstrate that DeepInflation can successfully discover simple and viable single-field slow-roll inflationary potentials consistent with the latest observations (with the ACT DR6 results taken as an example) or any given $n_s$ and $r$, and provide accurate theoretical context for obscure inflationary scenarios. DeepInflation serves as a prototype for a new generation of autonomous scientific discovery engines in cosmology, which enables researchers and non-experts alike to explore the inflationary landscape using natural language. This agent is available at https://github.com/pengzy-cosmo/DeepInflation.

Figures

Figures reproduced from arXiv: 2601.14288 by Gen Ye, Hao-Shi Yuan, Jun-Qian Jiang, Jun Zhang, Qi Lai, Yun-Song Piao, Ze-Yu Peng.

Figure 1
Figure 1. Figure 1: FIG. 1. System architecture of [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Diagnostic plot generated by the agent for the discovered best model [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 10 Pith papers

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

  1. DarkAgents

    hep-ph 2026-06 unverdicted novelty 6.0

    DarkAgents is an LLM-powered multi-agent framework for model building, pipeline computation, and assumption auditing in astroparticle physics, demonstrated on first-order phase transitions fitting NANOGrav gravitation...

  2. DarkAgents

    hep-ph 2026-06 conditional novelty 6.0

    A multi-agent LLM framework that runs and audits an astroparticle physics pipeline, reproducing human results for NANOGrav phase-transition fits and flagging limitations.

  3. Towards Verifiable and Self-Correcting AI Physicists for Quantum Many-Body Simulations

    physics.comp-ph 2026-03 unverdicted novelty 6.0

    QMP-Bench supplies a realistic test set for AI on quantum many-body problems while PhysVEC uses integrated verifiers to turn unreliable LLM generations into code that passes both syntax and physics checks, outperformi...

  4. ACT Data and Positive Running of the Spectral Index for Scalar Theory and Modified Gravity

    gr-qc 2026-05 unverdicted novelty 5.0

    Positive running of the spectral index is achievable in Einstein-Gauss-Bonnet gravity with viable inflation, unlike standard scalar field and F(R) models which face challenges.

  5. ACT-DR6 consistent inflation in generalised entropic cosmology and $f(Q)$ gravity

    gr-qc 2026-07 conditional novelty 4.0

    Reconstruction produces explicit f(Q) and generalised-entropic inflation models (and scalar-coupled versions) whose slow-roll parameters match ACT-DR6 + Planck-BAO constraints on n_s and r.

  6. Constraining supermassive primordial black hole clustering with the angular auto-correlation of $z\simeq 6$ quasars

    astro-ph.CO 2026-06 unverdicted novelty 4.0

    MCMC comparison of projected PBH correlation functions with z≈6 quasar angular auto-correlation data yields posterior constraints f_PBH∼10^{-3}, m_PBH∼10^{12}M_⊙ for Poisson models and ξ_eff≃2.1, r_cl≃76 Mpc for clust...

  7. Reconstructing ACT-compatible and GW170817-compatible Einstein-Gauss-Bonnet Inflation from the Observational Indices

    gr-qc 2026-05 unverdicted novelty 4.0

    Four Einstein-Gauss-Bonnet inflationary models are reconstructed from a chosen tensor-to-scalar ratio and shown to satisfy ACT and GW170817 constraints including scalar perturbation amplitude.

  8. ACT Data and Positive Running of the Spectral Index for Scalar Theory and Modified Gravity

    gr-qc 2026-05 unverdicted novelty 4.0

    Einstein-Gauss-Bonnet inflation can yield positive spectral-index running with a viable era under ACT-like hints, while scalar-field and F(R) models largely cannot.

  9. String-inspired Gauss-Bonnet Gravity Inflation and ACT

    gr-qc 2026-04 unverdicted novelty 4.0

    MCMC analysis of sixteen ghost-free f(R,G) inflation models shows all reproduce ns ≈ 0.97 at 60 e-folds with stable μ ≈ 0.1, preference set by Hubble parametrization.

  10. Reheating in geometric Weyl-invariant Einstein-Cartan gravity

    gr-qc 2026-01 unverdicted novelty 4.0

    Reheating temperature and equation-of-state parameter assumptions in Weyl-invariant Einstein-Cartan gravity models significantly alter predicted inflationary observables.

Reference graph

Works this paper leans on

85 extracted references · 58 linked inside Pith · cited by 8 Pith papers

  1. [2]

    A. H. Guth, Phys. Rev. D23, 347 (1981)

    1981

  2. [3]

    A. D. Linde, Phys. Lett. B108, 389 (1982)

    1982

  3. [4]

    Albrecht and P

    A. Albrecht and P. J. Steinhardt, Phys. Rev. Lett.48, 1220 (1982)

    1982

  4. [5]

    A. A. Starobinsky, Phys. Lett. B91, 99 (1980)

    1980

  5. [6]

    V. F. Mukhanov and G. V. Chibisov, JETP Lett.33, 532 (1981)

    1981

  6. [7]

    Aghanimet al.(Planck), Astron

    N. Aghanimet al.(Planck), Astron. Astrophys.641, A6 (2020), [Erratum: Astron.Astrophys. 652, C4 (2021)], arXiv:1807.06209 [astro-ph.CO]

    Pith/arXiv arXiv 2020

  7. [8]

    Martin, C

    J. Martin, C. Ringeval, and V. Vennin, Phys. Dark Univ. 5-6, 75 (2014), arXiv:1303.3787 [astro-ph.CO]

    Pith/arXiv arXiv 2014

  8. [9]

    Kallosh and A

    R. Kallosh and A. Linde, Gen. Rel. Grav.57, 135 (2025), arXiv:2505.13646 [hep-th]

    arXiv 2025

  9. [10]

    Hinshawet al.(WMAP), Astrophys

    G. Hinshawet al.(WMAP), Astrophys. J. Suppl.208, 19 (2013), arXiv:1212.5226 [astro-ph.CO]

    Pith/arXiv arXiv 2013

  10. [11]

    Louiset al.(Atacama Cosmology Telescope), JCAP 11, 062 (2025), arXiv:2503.14452 [astro-ph.CO]

    T. Louiset al.(Atacama Cosmology Telescope), JCAP 11, 062 (2025), arXiv:2503.14452 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  11. [12]

    Calabreseet al.(Atacama Cosmology Telescope), JCAP11, 063 (2025), arXiv:2503.14454 [astro-ph.CO]

    E. Calabreseet al.(Atacama Cosmology Telescope), JCAP11, 063 (2025), arXiv:2503.14454 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  12. [13]

    Camphuiset al.(SPT-3G), (2025), arXiv:2506.20707 [astro-ph.CO]

    E. Camphuiset al.(SPT-3G), (2025), arXiv:2506.20707 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  13. [14]

    A. A. Starobinsky, JETP Lett.55, 489 (1992)

    1992

  14. [15]

    Linde, (2025), arXiv:2509.01675 [hep-th]

    A. Linde, (2025), arXiv:2509.01675 [hep-th]

    arXiv 2025

  15. [16]

    Kallosh, A

    R. Kallosh, A. Linde, and D. Roest, Phys. Rev. Lett. 135, 161001 (2025), arXiv:2503.21030 [hep-th]

    Pith/arXiv arXiv 2025

  16. [17]

    Berera, S

    A. Berera, S. Brahma, Z. Qiu, R. O. Ramos, and G. S. Rodrigues, JCAP11, 059 (2025), arXiv:2504.02655 [hep- th]

    arXiv 2025

  17. [18]

    Dioguardi, A

    C. Dioguardi, A. J. Iovino, and A. Racioppi, Phys. Lett. B868, 139664 (2025), arXiv:2504.02809 [gr-qc]

    Pith/arXiv arXiv 2025

  18. [19]

    Brahma and J

    S. Brahma and J. Calder´ on-Figueroa, (2025), arXiv:2504.02746 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  19. [20]

    I. D. Gialamas, A. Karam, A. Racioppi, and M. Raidal, Phys. Rev. D112, 103544 (2025), arXiv:2504.06002 [astro-ph.CO]

    arXiv 2025

  20. [21]

    L. Liu, Z. Yi, and Y. Gong, (2025), arXiv:2505.02407 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  21. [22]

    Q. Gao, Y. Gong, Z. Yi, and F. Zhang, Phys. Dark Univ. 50, 102106 (2025), arXiv:2504.15218 [astro-ph.CO]

    arXiv 2025

  22. [23]

    I. D. Gialamas, T. Katsoulas, and K. Tamvakis, JCAP 09, 060 (2025), arXiv:2505.03608 [gr-qc]

    arXiv 2025

  23. [24]

    Heidarian, M

    H. Heidarian, M. Solbi, S. Heydari, and K. Karami, Phys. Lett. B869, 139833 (2025), arXiv:2506.10547 [astro-ph.CO]

    arXiv 2025

  24. [25]

    S. D. Odintsov and V. K. Oikonomou, Phys. Lett. B870, 139907 (2025), arXiv:2509.06251 [gr-qc]

    arXiv 2025

  25. [26]

    Bezerra-Sobrinho and L

    J. Bezerra-Sobrinho and L. G. Medeiros, (2025), arXiv:2511.06640 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  26. [27]

    C. Fu, D. Lu, and S.-J. Wang, (2025), arXiv:2510.24682 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  27. [28]

    Qiu, Y.-H

    Z.-C. Qiu, Y.-H. Pang, and Q.-G. Huang, (2025), arXiv:2510.18320 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  28. [29]

    G. Ye, B. Hu, and Y.-S. Piao, Phys. Rev. D104, 063510 (2021), arXiv:2103.09729 [astro-ph.CO]

    Pith/arXiv arXiv 2021

  29. [30]

    Jiang, G

    J.-Q. Jiang, G. Ye, and Y.-S. Piao, Phys. Lett. B851, 138588 (2024), arXiv:2303.12345 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  30. [31]

    E. G. M. Ferreira, E. McDonough, L. Balkenhol, R. Kallosh, L. Knox, and A. Linde, (2025), arXiv:2507.12459 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  31. [32]

    McDonough and E

    E. McDonough and E. G. M. Ferreira, (2025), arXiv:2512.05108 [astro-ph.CO]

    arXiv 2025

  32. [33]

    Ye and Y.-S

    G. Ye and Y.-S. Piao, Phys. Rev. D101, 083507 (2020), arXiv:2001.02451 [astro-ph.CO]

    Pith/arXiv arXiv 2020

  33. [34]

    Jiang and Y.-S

    J.-Q. Jiang and Y.-S. Piao, Phys. Rev. D105, 103514 (2022), arXiv:2202.13379 [astro-ph.CO]

    Pith/arXiv arXiv 2022

  34. [35]

    Jiang, G

    J.-Q. Jiang, G. Ye, and Y.-S. Piao, Mon. Not. Roy. Astron. Soc.527, L54 (2023), arXiv:2210.06125 [astro- ph.CO]

    Pith/arXiv arXiv 2023

  35. [36]

    T. L. Smith, M. Lucca, V. Poulin, G. F. Abellan, L. Balkenhol, K. Benabed, S. Galli, and R. Murgia, Phys. Rev. D106, 043526 (2022), arXiv:2202.09379 [astro- ph.CO]

    Pith/arXiv arXiv 2022

  36. [37]

    Peng and Y.-S

    Z.-Y. Peng and Y.-S. Piao, Phys. Rev. D109, 023519 (2024), arXiv:2308.01012 [astro-ph.CO]

    Pith/arXiv arXiv 2024

  37. [38]

    H. Wang, G. Ye, J.-Q. Jiang, and Y.-S. Piao, Phys. Rev. D111, 123505 (2025), arXiv:2409.17879 [astro-ph.CO]

    Pith/arXiv arXiv 2025

  38. [39]

    Peng, J.-Q

    Z.-Y. Peng, J.-Q. Jiang, H. Wang, and Y.-S. Piao, Phys. Rev. D112, 123519 (2025), arXiv:2509.11902 [astro- ph.CO]

    arXiv 2025

  39. [40]

    A survey of large language models,

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y. Hou, Y. Min, B. Zhang, J. Zhang, Z. Dong, Y. Du, C. Yang, Y. Chen, Z. Chen, J. Jiang, R. Ren, Y. Li, X. Tang, Z. Liu, P. Liu, J.-Y. Nie, and J.-R. Wen, “A survey of large language models,” (2025), arXiv:2303.18223 [cs.CL]

    Pith/arXiv arXiv 2025

  40. [41]

    Llm4sr: A survey on large language models for scientific research,

    Z. Luo, Z. Yang, Z. Xu, W. Yang, and X. Du, “Llm4sr: A survey on large language models for scientific research,” (2025), arXiv:2501.04306 [cs.CL]

    Pith/arXiv arXiv 2025

  41. [42]

    A survey of ai scientists,

    G. Tie, P. Zhou, and L. Sun, “A survey of ai scientists,” (2025), arXiv:2510.23045 [cs.AI]

    arXiv 2025

  42. [43]

    K. G. Barman, S. Caron, E. Sullivan, H. W. de Regt, R. R. de Austri, M. Boon, M. F¨ arber, S. Fr¨ ose, T. Golling, L. G. Lopez, F. Hasibi, L. Heinrich, A. Ipp, R. Kapoor, G. Kasieczka, D. Kosti´ c, M. Kr¨ amer, J. Marco, S. Otten, P. Pawlowski, P. Vischia, E. We- ber, and C. Weniger, The European Physical Journal C 85(2025), 10.1140/epjc/s10052-025-14707-8

    doi:10.1140/epjc/s10052-025-14707-8 2025

  43. [44]

    Huang, W

    L. Huang, W. Yu, W. Ma, W. Zhong, Z. Feng, H. Wang, Q. Chen, W. Peng, X. Feng, B. Qin, and T. Liu, ACM Transactions on Information Systems43, 1–55 (2025)

    2025

  44. [45]

    Explor- ing the limits of fine-grained llm-based physics in- ference via premise removal interventions,

    J. Meadows, T. James, and A. Freitas, “Explor- ing the limits of fine-grained llm-based physics in- ference via premise removal interventions,” (2024), arXiv:2404.18384 [cs.CL]

    Pith/arXiv arXiv 2024

  45. [46]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao, “React: Synergizing reasoning and acting in language models,” (2023), arXiv:2210.03629 [cs.CL]

    Pith/arXiv arXiv 2023

  46. [47]

    Laverick, K

    A. Laverick, K. Surrao, I. Zubeldia, B. Bolliet, M. Cran- mer, A. Lewis, B. Sherwin, and J. Lesgourgues, (2024), arXiv:2412.00431 [astro-ph.IM]

    Pith/arXiv arXiv 2024

  47. [48]

    Open source planning & control system with language agents for autonomous scientific discov- ery,

    L. Xu, M. Sarkar, A. I. Lonappan, ´I˜ nigo Zubeldia, P. Villanueva-Domingo, S. Casas, C. Fidler, C. Aman- charla, U. Tiwari, A. Bayer, C. A. Ekioui, M. Cranmer, A. Dimitrov, J. Fergusson, K. Gandhi, S. Krippendorf, A. Laverick, J. Lesgourgues, A. Lewis, T. Meier, B. Sher- win, K. Surrao, F. Villaescusa-Navarro, C. Wang, X. Xu, and B. Bolliet, “Open source...

    Pith/arXiv arXiv 2025

  48. [49]

    Casas, C

    S. Casas, C. Fidler, B. Bolliet, F. Villaescusa-Navarro, and J. Lesgourgues, (2025), arXiv:2508.05728 [astro- 7 ph.IM]

    Pith/arXiv arXiv 2025

  49. [50]

    Zhang, Z

    X. Zhang, Z. Bi, P. Lachance, X. Wang, T. Di Matteo, and R. A. C. Croft, (2025), arXiv:2507.08958 [astro- ph.IM]

    Pith/arXiv arXiv 2025

  50. [51]

    Schmidt and H

    M. Schmidt and H. Lipson, Science324, 81 (2009)

    2009

  51. [52]

    Interpretable machine learning for sci- ence with pysr and symbolicregression.jl,

    M. Cranmer, “Interpretable machine learning for sci- ence with pysr and symbolicregression.jl,” (2023), arXiv:2305.01582 [astro-ph.IM]

    Pith/arXiv arXiv 2023

  52. [53]

    Introduction to symbolic regression in the phys- ical sciences,

    D. J. Bartlett, H. Desmond, P. G. Ferreira, and G. Kron- berger, “Introduction to symbolic regression in the phys- ical sciences,” (2025), arXiv:2512.15920 [cs.LG]

    Pith/arXiv arXiv 2025

  53. [54]

    Retrieval- augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. K¨ uttler, M. Lewis, W. tau Yih, T. Rockt¨ aschel, S. Riedel, and D. Kiela, “Retrieval- augmented generation for knowledge-intensive nlp tasks,” (2021), arXiv:2005.11401 [cs.CL]

    Pith/arXiv arXiv 2021

  54. [55]

    Baumann,Cosmology(Cambridge University Press, 2022)

    D. Baumann,Cosmology(Cambridge University Press, 2022)

    2022

  55. [56]

    P. A. R. Adeet al.(BICEP, Keck), Phys. Rev. Lett.127, 151301 (2021), arXiv:2110.00483 [astro-ph.CO]

    arXiv 2021

  56. [57]

    Kallosh and A

    R. Kallosh and A. Linde, Phys. Rev. D100, 123523 (2019), arXiv:1909.04687 [hep-th]

    Pith/arXiv arXiv 2019

  57. [58]

    Kallosh, A

    R. Kallosh, A. Linde, and Y. Yamada, JHEP01, 008 (2019), arXiv:1811.01023 [hep-th]

    Pith/arXiv arXiv 2019

  58. [59]

    G. R. Dvali and S. H. H. Tye, Phys. Lett. B450, 72 (1999), arXiv:hep-ph/9812483

    Pith/arXiv arXiv 1999

  59. [60]

    C. P. Burgess, M. Majumdar, D. Nolte, F. Quevedo, G. Rajesh, and R.-J. Zhang, JHEP07, 047 (2001), arXiv:hep-th/0105204

    Pith/arXiv arXiv 2001

  60. [61]

    Kachru, R

    S. Kachru, R. Kallosh, A. D. Linde, J. M. Maldacena, L. P. McAllister, and S. P. Trivedi, JCAP10, 013 (2003), arXiv:hep-th/0308055

    Pith/arXiv arXiv 2003

  61. [62]

    Kallosh and A

    R. Kallosh and A. Linde, JCAP12, 008 (2021), arXiv:2110.10902 [astro-ph.CO]

    Pith/arXiv arXiv 2021

  62. [63]

    Galante, R

    M. Galante, R. Kallosh, A. Linde, and D. Roest, Phys. Rev. Lett.114, 141302 (2015), arXiv:1412.3797 [hep-th]

    Pith/arXiv arXiv 2015

  63. [64]

    Terada, Phys

    T. Terada, Phys. Lett. B760, 674 (2016), arXiv:1602.07867 [hep-th]

    Pith/arXiv arXiv 2016

  64. [65]

    Arkani-Hamed, H.-C

    N. Arkani-Hamed, H.-C. Cheng, P. Creminelli, and L. Randall, JCAP07, 003 (2003), arXiv:hep-th/0302034

    Pith/arXiv arXiv 2003

  65. [66]

    Arkani-Hamed, H.-C

    N. Arkani-Hamed, H.-C. Cheng, P. Creminelli, and L. Randall, Phys. Rev. Lett.90, 221302 (2003), arXiv:hep-th/0301218

    Pith/arXiv arXiv 2003

  66. [67]

    Freese, J

    K. Freese, J. A. Frieman, and A. V. Olinto, Phys. Rev. Lett.65, 3233 (1990)

    1990

  67. [68]

    F. C. Adams, J. R. Bond, K. Freese, J. A. Frieman, and A. V. Olinto, Phys. Rev. D47, 426 (1993), arXiv:hep- ph/9207245

    arXiv 1993

  68. [69]

    A. D. Linde, Phys. Lett. B259, 38 (1991)

    1991

  69. [70]

    A. D. Linde, Phys. Rev. D49, 748 (1994), arXiv:astro- ph/9307002

    arXiv 1994

  70. [71]

    Kallosh and A

    R. Kallosh and A. Linde, Phys. Rev. D106, 023522 (2022), arXiv:2204.02425 [hep-th]

    Pith/arXiv arXiv 2022

  71. [72]

    Ye, J.-Q

    G. Ye, J.-Q. Jiang, and Y.-S. Piao, Phys. Rev. D106, 103528 (2022), arXiv:2205.02478 [astro-ph.CO]

    Pith/arXiv arXiv 2022

  72. [73]

    Braglia, A

    M. Braglia, A. Linde, R. Kallosh, and F. Finelli, JCAP 04, 033 (2023), arXiv:2211.14262 [astro-ph.CO]

    Pith/arXiv arXiv 2023

  73. [74]

    Piao and Y.-Z

    Y.-S. Piao and Y.-Z. Zhang, Phys. Rev. D70, 063513 (2004), arXiv:astro-ph/0401231

    Pith/arXiv arXiv 2004

  74. [75]

    Z.-G. Liu, J. Zhang, and Y.-S. Piao, Phys. Rev. D84, 063508 (2011), arXiv:1105.5713 [astro-ph.CO]

    Pith/arXiv arXiv 2011

  75. [76]

    Cai and Y.-S

    Y. Cai and Y.-S. Piao, Phys. Rev. D103, 083521 (2021), arXiv:2012.11304 [gr-qc]

    Pith/arXiv arXiv 2021

  76. [77]

    Y. Cai, M. Zhu, and Y.-S. Piao, Phys. Rev. Lett.133, 021001 (2024), arXiv:2305.10933 [gr-qc]

    Pith/arXiv arXiv 2024

  77. [78]

    Garcia-Bellido and E

    J. Garcia-Bellido and E. Ruiz Morales, Phys. Dark Univ. 18, 47 (2017), arXiv:1702.03901 [astro-ph.CO]

    Pith/arXiv arXiv 2017

  78. [79]

    Gao and Z.-K

    T.-J. Gao and Z.-K. Guo, Phys. Rev. D98, 063526 (2018), arXiv:1806.09320 [hep-ph]

    Pith/arXiv arXiv 2018

  79. [80]

    S. S. Mishra and V. Sahni, JCAP04, 007 (2020), arXiv:1911.00057 [gr-qc]

    Pith/arXiv arXiv 2020

  80. [81]

    A. J. Iovino and A. Riotto, (2025), arXiv:2512.19668 [astro-ph.CO]

    Pith/arXiv arXiv 2025

Showing first 80 references.