Pith. sign in

REVIEW 3 major objections 5 minor 43 references

LLM-written simulators plus neural Bayes-factor scoring can jointly identify model families and fit their parameters.

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 17:40 UTC pith:MJUOP2PA

load-bearing objection Solid methods paper with real derivations and honest benchmarks; the ODE feedback loop reuses the observation for both search and scoring, and that should be fixed before publication. the 3 major comments →

arxiv 2607.17540 v1 pith:MJUOP2PA submitted 2026-07-20 cs.LG stat.ML

Program Synthesis for Simulation-Based Inference: Joint Model Selection and Parameter Estimation

Siddharth Mishra-Sharma This is my paper
classification cs.LG stat.ML
keywords program synthesissimulation-based inferencemodel selectionBayesian model comparisonlarge language modelsneural ratio estimationevidence networksdark matter substructure
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 proposes a framework that treats scientific model discovery as Bayesian inference over executable simulator programs. A large language model, given a plain-language description of a problem and a dataset, generates candidate simulator programs together with parameter priors; the candidates are then compared using neural density estimators that target log Bayes factors directly, yielding a joint posterior over both which model generated the data and what its parameters are. Iterative rounds of resampling and LLM mutation concentrate computation on promising model families, while pairwise comparisons are aggregated into a globally consistent model posterior by least squares, avoiding a subtle overconfidence bias in naive averaging. The authors show on damped-oscillator fitting, SIR-type epidemic models, and dark-matter substructure lensing that the right model family is usually recovered from open-ended prompts, and that when the data cannot distinguish candidates, the posterior spreads rather than collapsing. The claim that matters is that model selection and parameter estimation can be performed over a pool of LLM-proposed simulators rather than within a fixed model.

Core claim

The paper's central claim is that the joint posterior p(M, θ | x) over model index and parameters can be computed when the model space is itself generated by a large language model rather than fixed in advance. The pipeline factorizes this into model selection via pairwise neural ratio estimation ('evidence networks') that read log Bayes factors directly, with a least-squares projection to reconcile all pairwise comparisons into a globally consistent log-evidence vector, and parameter inference via normalizing-flow density estimators. To keep the candidate pool alive, models are resampled by their current weight and the LLM is used as a mutation kernel to propose structural variants; the pap

What carries the argument

The central machinery is the pairwise evidence network: a neural classifier trained with an exponential loss whose optimum is the log Bayes factor between two models' marginalized likelihoods, so model comparison does not require estimating either likelihood or posterior separately. A least-squares projection (equivalent to a Laplacian solve on the complete comparison graph) turns the collection of pairwise log Bayes factors into a single consistent log-evidence vector, correcting the K/(K−1) inverse-temperature bias of naive averaging and exposing cyclic disagreement as a diagnostic residual. Around this core sits an iterative loop where candidate programs are resampled by their current wei

Load-bearing premise

The load-bearing premise is that a model posterior computed from candidates that were themselves proposed and mutated using the same observation is still a valid Bayesian posterior, even though the final weights are conditional on a data-dependent pool.

What would settle it

Rerun the SIR benchmark with the final model weights computed on an independent observation generated from the same true process, never shown to the LLM during mutation; if the leading-family accuracy falls well below the paper's reported level, the feedback loop was leaking scoring information into the candidate pool. A complementary test is to apply the pipeline to a synthetic dynamical system unlikely to be represented in LLM training data and require it to recover the true family.

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

If this is right

  • A scientist could hand the pipeline a plain-language problem statement plus data and receive a ranked set of executable simulator candidates with per-model parameter posteriors, without hand-coding any simulator.
  • Because marginal-likelihood scoring automatically penalizes complexity, the pipeline prefers simpler models when the data are ambiguous; the authors argue this should be read as a diagnostic that more data are needed, not as evidence against a more complex model.
  • When the data cannot separate candidate families, the model posterior remains spread instead of collapsing, providing an explicit uncertainty signal through the effective-sample-size diagnostic.
  • The lensing benchmark shows that increasing the data per evaluation (from 4 to 20 lensing systems) removes a systematic count bias and shifts posterior weight onto the true model, implying the framework's selection accuracy scales with information content.
  • The inference stage is agnostic to how candidates are proposed, so improvements in LLM code generation should transfer directly into better model proposals without changing the statistical core.

Where Pith is reading between the lines

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

  • A natural extension is to propagate the least-squares fit residual—the cyclic disagreement among pairwise classifiers—into the model posterior as an additional uncertainty component; the paper reports it as a diagnostic but leaves it out of the posterior.
  • The most decisive test of the framework would hold out the observation used for mutation feedback in the ODE benchmarks: recomputing final weights on an independent observation from the same ground truth would quantify how much of the reported model-selection accuracy comes from the feedback loop rather than from an honest Bayesian comparison.
  • The set-level evidence network recipe looks directly transferable to other population-level inference problems with intractable likelihoods, such as estimating shared parameters across many short time series or multi-patient clinical trajectories.
  • If the LLM proposal prior is the real bottleneck, the pipeline's discovery power will track the LLM's domain familiarity; a stress test on a genuinely novel dynamical system absent from LLM training data would separate the contribution of the Bayesian scoring from that of the proposal prior.

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 proposes a framework for joint model selection and parameter estimation that combines LLM-based program synthesis with neural simulation-based inference. Given a natural-language problem description, an LLM proposes candidate simulator programs with parameter priors; an iterative resample/mutate loop refines the candidate pool; pairwise 'evidence networks' estimate log Bayes factors for all candidate pairs; a least-squares projection converts the pairwise estimates into model posterior weights; and masked autoregressive flows provide per-model parameter posteriors. Experiments cover a damped oscillator, stochastic SIR/SEIR/SIS epidemic dynamics, and dark-matter subhalo population inference from gravitational-lensing images. The paper claims that the pipeline recovers plausible model families from open-ended prompts, with accuracy reflecting data information content and identifiability.

Significance. If valid, this is a useful step toward automating model discovery and Bayes-factor model comparison over an open candidate space. The least-squares aggregation of pairwise log Bayes factors in Section 3.3/Appendix C is a genuine methodological contribution: the derivation of the implicit inverse temperature of naive averaging and the closed-form Laplacian solution are correct and clearly presented. The subhalo benchmark is also honest in its information-content analysis: the N=4 vs N=20 comparison diagnoses why count posteriors are prior-dominated, and the held-out scoring split in Appendix B is the right safeguard. The main caveat is that the ODE benchmarks double-use the observed data to guide the search and then to compute the final model weights, which calls the headline model-selection accuracies into question. The paper is transparent about this setup, but the transparency does not remove the bias.

major comments (3)
  1. [§4.1 and §3.2] In the ODE benchmarks, the mutation feedback includes each candidate's RMSE against the observation and the sign of its residual trend (Section 4.1), and the final model weights are then computed from ratio estimation on the same observation x_obs (Section 3.2). This is a double use of the data: x_obs shapes which models are proposed, and the same x_obs scores those models in the final round. The paper says the posterior is 'conditional on the final evaluation pool', but conditioning on an adaptively selected, x_obs-dependent pool is not the same as conditioning on a pre-specified model space. Consequently Eq. (1) is not a posterior over a fixed model space, and the reported accuracies (10/10, 8/10, 9/10, 5/10) may overstate what a prior-based model comparison would give. The subhalo benchmark avoids this by using a held-out scoring split (Appendix B); a similar split or pre-registration
  2. [§4.2, §5, Appendix E] The paper claims joint parameter estimation as part of the central contribution, yet the SBC validation in Appendix E is performed only on the well-specified damped-cosine base model. The paper itself notes in Section 5 that the NPE stage 'can also be loose for heavily mutated candidates' and that the posterior-predictive check in Figure 1 uses exact-likelihood MCMC rather than the NPE posterior. This leaves the per-model parameter posteriors for the actual LLM-generated winners unvalidated. Please report SBC or another coverage diagnostic on the actual winning programs, or at least on deliberately misspecified LLM-generated candidates, so the parameter-estimation half of the claim is supported.
  3. [§4.4] The subhalo benchmark's residual channel uses the ground-truth smooth lens model ('Supplying the ground-truth smooth model is an oracle simplification'). This is an important caveat for the substructure-family results: in a real analysis the smooth-lens subtraction would carry additional uncertainty into the residual channel. The paper acknowledges this, but the abstract's phrasing ('dark matter substructure inference from gravitational-lensing images') is stronger than what the oracle channel supports. Please either soften the claim or add an experiment that infers the smooth model jointly or marginalizes over its nuisance parameters.
minor comments (5)
  1. [§4.1] The summary network for the ODE evidence networks is described only as '32-dimensional learned summary statistics produced by a summary network trained per round'. The architecture and training details of this summary network are omitted; please include them for reproducibility.
  2. [§4.3, Figure 2] The caption of Figure 2 says the rightmost panel shows ESS for the SIR ground truth, but the panel shows a jump from round 1 to round 2. Please clarify whether the reported median ESS is over the same 10 seeds as the left panels and whether any seed terminated early.
  3. [§5] The sentence 'Because the inference stage is agnostic to the source of candidates, it evaluates whatever models are proposed' is too strong. The inference stage is agnostic in principle, but in this pipeline the proposed pool is influenced by the observed data through the mutation feedback, so the final posterior is not independent of the proposal mechanism.
  4. [§4.1] All experiments use the proprietary Claude Opus 4.6 model. For reproducibility, please state the exact model version, sampling temperature, and whether the same model version was used across all seeds; also consider releasing the exact prompts and generated programs.
  5. [Appendix D] The example programs in Appendix D are useful, but the text says 'three damped-oscillator runs that converged in Round 1' while Figure 1 shows a run that converged after 10 rounds. Please make the relationship between the appendix examples and the main runs explicit.

Circularity Check

0 steps flagged

No significant circularity: the SBI machinery is self-contained; the disclosed pool-conditional ODE refinement is a selection-bias limitation, not a definitional reduction.

full rationale

The paper's load-bearing derivations are self-contained. Evidence networks (Section 3.3, Eqs. 5-6) derive log-Bayes factors from a defined exponential loss; least-squares aggregation (Eq. 7, Appendix C) is a self-contained correction of naive averaging; NPE training (Section 3.4) uses only prior-predictive simulations, and parameter-recovery and SBC checks use held-out synthetic data. The only data-dependent step is the ODE mutation feedback (Section 4.1: 'the mutation feedback (Section 3) additionally includes each candidate’s root-mean-square error (RMSE) against the observation and the sign of its residual trend over time'), and final ODE model weights are computed on that same observation, unlike the subhalo benchmark's held-out scoring split (Appendix B: 'final-round model weights computed on a held-out scoring split, disjoint from the images used during search'). This is a double-use/selection-bias concern for the unconditional interpretation of the reported accuracies. Crucially, the paper itself discloses this: Section 3.2 states the procedure 'is best understood as population-based evolutionary search guided by evidence estimates, not a rigorous SMC sampler targeting p(M|x)' and that the final weights are 'Bayesian posterior probabilities conditional on the final evaluation pool.' Because the claim is explicitly pool-conditional, the result does not reduce by construction to the feedback; the absence of a held-out split is a validity limitation, not circularity. Self-citations (Mishra-Sharma 2023; Brehmer et al. 2019) are contextual and not load-bearing.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The framework rests on the LLM as an implicit prior over program space and on neural ratio estimation converging to marginal likelihood ratios. The free parameters are mostly compute/architecture choices; the LLM-specified priors are a genuine source of freedom because they directly affect the Occam penalty. There are no invented physical entities.

free parameters (5)
  • Number of candidates per round K = 8
    Chosen by hand; diversity of the pool and the cost of pairwise ratio estimation (K choose 2) both depend on it.
  • Number of SMC-style rounds R = 10 (ODE), 15 (subhalo)
    Termination budget; the final posterior and family recovery rates depend on how much mutation is allowed.
  • Simulation budget per model per round = 2000 (ODE), 320 (subhalo)
    Controls the variance of the evidence-network estimates; chosen by hand.
  • LLM-specified parameter priors = per-candidate, generated by LLM
    The marginal likelihood (and thus the Occam penalty) is sensitive to prior width; these are generated, not measured, and are part of the (program, prior) pair being compared.
  • Benchmark-specific mutation feedback features = RMSE + residual sign (ODE)
    Hand-designed guidance to the LLM; different feedback would change the search trajectory and final pool.
axioms (4)
  • domain assumption The LLM proposal/mutation distribution has non-negligible probability of generating the true model family and reasonable alternatives.
    If the true model is never in the pool, the posterior cannot include it; the paper notes all benchmarks are in-distribution for LLMs (Section 5).
  • standard math The exponential-loss evidence network converges to the log Bayes factor when trained on prior-predictive simulations.
    Functional derivative Eq. (6) gives f* = log BF under sufficient capacity/statistics; this is the mathematical core of the model-comparison step.
  • standard math Pairwise log-Bayes-factor estimates are conditionally unbiased enough that the least-squares projection yields a valid log-evidence vector.
    The Gaussian-noise interpretation in Appendix C justifies the projection; if pairwise estimates are biased or have heteroskedastic noise, the posterior is distorted.
  • ad hoc to paper For the ODE benchmarks, the data used to steer mutation can also be used to compute final model weights without an invalid double-use of the data.
    No held-out scoring split is reported for ODE benchmarks (contrast Section 4.4 subhalo's held-out split); this is an unaddressed selection-bias assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 19775 in / 15953 out tokens · 130205 ms · 2026-08-01T17:40:35.577479+00:00 · methodology

0 comments
read the original abstract

Neural simulation-based inference enables parameter estimation for complex models, but typically requires the user to specify a simulator encoding a fixed model structure. We present a framework for joint model selection and parameter estimation that combines large language models for program synthesis with neural simulation-based inference. Given a natural language description of the system and data under investigation, an LLM proposes candidate simulator programs which are iteratively refined via feedback-driven mutation and evaluated using neural density estimation. The approach enables simulation-based inference over a pool of models, not just parameters within a fixed model. On benchmarks spanning deterministic dynamics, stochastic epidemic models, and dark matter substructure inference from gravitational-lensing images, the method identifies plausible model families from open-ended prompts, with accuracy that reflects the information content of the data and identifiability of candidate models.

Figures

Figures reproduced from arXiv: 2607.17540 by Siddharth Mishra-Sharma.

Figure 1
Figure 1. Figure 1: Damped oscillator benchmark: the method recovers a damped, oscillating model and [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: SIR epidemic model selection: the method recovers the correct compartmental family when [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Subhalo benchmark (4-lens set-level inference, 3 seeds per ground truth): the profile family [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The CDM count bias is a limit of data information content: adding lensing systems removes [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Aggregating pairwise log Bayes factors: least squares recovers the true posterior where [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Coverage calibration for the damped oscillator benchmark (200 SBC test cases): the NPE [PITH_FULL_IMAGE:figures/full_fig_p021_6.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 · 12 linked inside Pith

  1. [1]

    Claude system card

    Anthropic . Claude system card. https://www.anthropic.com/system-cards, 2025. Accessed: 2025

    2025

  2. [2]

    Program synthesis with large language models

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021

    Pith/arXiv arXiv 2021

  3. [3]

    Mining for dark matter substructure: Inferring subhalo population properties from strong lenses with machine learning

    Johann Brehmer, Siddharth Mishra-Sharma, Joeri Hermans, Gilles Louppe, and Kyle Cranmer. Mining for dark matter substructure: Inferring subhalo population properties from strong lenses with machine learning. The Astrophysical Journal, 886 0 (1): 0 49, 2019

    2019

  4. [4]

    Evaluating large language models trained on code

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021

    Pith/arXiv arXiv 2021

  5. [5]

    Approximating likelihood ratios with calibrated discriminative classifiers

    Kyle Cranmer, Juan Pavez, and Gilles Louppe. Approximating likelihood ratios with calibrated discriminative classifiers. arXiv preprint arXiv:1506.02169, 2015

    Pith/arXiv arXiv 2015

  6. [6]

    The frontier of simulation-based inference

    Kyle Cranmer, Johann Brehmer, and Gilles Louppe. The frontier of simulation-based inference. Proceedings of the National Academy of Sciences, 117 0 (48): 0 30055--30062, 2020

    2020

  7. [7]

    Truncated proposals for scalable and hassle-free simulation-based inference

    Michael Deistler, Pedro J Gon c alves, and Jakob H Macke. Truncated proposals for scalable and hassle-free simulation-based inference. Advances in Neural Information Processing Systems, 35: 0 23135--23149, 2022

    2022

  8. [8]

    Simulation-based inference: A practical guide

    Michael Deistler, Jan Boelts, Peter Steinbach, Guy Moss, Thomas Moreau, Manuel Gloeckler, Pedro LC Rodrigues, Julia Linhart, Janne K Lappalainen, Benjamin Kurt Miller, et al. Simulation-based inference: A practical guide. arXiv preprint arXiv:2508.12939, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Large language Bayes

    Justin Domke. Large language Bayes . arXiv preprint arXiv:2504.14025, 2025

    arXiv 2025

  10. [10]

    Gemini 3 Pro model card

    Google DeepMind . Gemini 3 Pro model card. Technical report, Google DeepMind, 2025. November 2025

    2025

  11. [11]

    Automatic posterior transformation for likelihood-free inference

    David Greenberg, Marcel Nonnenmacher, and Jakob Macke. Automatic posterior transformation for likelihood-free inference. International Conference on Machine Learning, pages 2404--2414, 2019

    2019

  12. [12]

    Program synthesis

    Sumit Gulwani, Oleksandr Polozov, and Rishabh Singh. Program synthesis. Foundations and Trends in Programming Languages, 4 0 (1-2): 0 1--119, 2017

    2017

  13. [13]

    DeepSeek-R1 : Incentivizing reasoning capability in LLMs via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. DeepSeek-R1 : Incentivizing reasoning capability in LLMs via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Harris, K

    Charles R. Harris, K. Jarrod Millman, St \' e fan J. van der Walt, Ralf Gommers, Pauli Virtanen, David Cournapeau, Eric Wieser, Julian Taylor, Sebastian Berg, Nathaniel J. Smith, Robert Kern, Matti Picus, Stephan Hoyer, Marten H. van Kerkwijk, Matthew Brett, Allan Haldane, Jaime Fern \' a ndez del R \' i o, Mark Wiebe, Pearu Peterson, Pierre G \' e rard-M...

    2020

  15. [15]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770--778, 2016

    2016

  16. [16]

    Likelihood-free MCMC with amortized approximate ratio estimators

    Joeri Hermans, Volodimir Begy, and Gilles Louppe. Likelihood-free MCMC with amortized approximate ratio estimators. International Conference on Machine Learning, pages 4239--4248, 2020

    2020

  17. [17]

    J. D. Hunter. Matplotlib: A 2D graphics environment. Computing in Science & Engineering, 9 0 (3): 0 90--95, 2007. doi:10.1109/MCSE.2007.55

    doi:10.1109/mcse.2007.55 2007

  18. [18]

    Evidence networks: simple losses for fast, amortized, neural Bayesian model comparison

    Niall Jeffrey and Benjamin D Wandelt. Evidence networks: simple losses for fast, amortized, neural Bayesian model comparison. Machine Learning: Science and Technology, 5 0 (1): 0 015008, 2024

    2024

  19. [19]

    Statistical ranking and combinatorial Hodge theory

    Xiaoye Jiang, Lek-Heng Lim, Yuan Yao, and Yinyu Ye. Statistical ranking and combinatorial Hodge theory. Mathematical Programming, 127 0 (1): 0 203--244, 2011

    2011

  20. [20]

    Kimi K2 : Open agentic intelligence

    Kimi Team , Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, et al. Kimi K2 : Open agentic intelligence. arXiv preprint arXiv:2507.20534, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam : A method for stochastic optimization. In International Conference on Learning Representations ( ICLR ) , 2015. URL http://arxiv.org/abs/1412.6980

    Pith/arXiv arXiv 2015

  22. [22]

    Flexible statistical inference for mechanistic models of neural dynamics

    Jan-Matthis Lueckmann, Pedro J Gon c alves, Giacomo Bassetto, Kaan \"O cal, Marcel Nonnenmacher, and Jakob H Macke. Flexible statistical inference for mechanistic models of neural dynamics. Advances in Neural Information Processing Systems, 30, 2017

    2017

  23. [23]

    Truncated marginal neural ratio estimation

    Benjamin Kurt Miller, Alex Cole, Patrick Forr \'e , Gilles Louppe, and Christoph Weniger. Truncated marginal neural ratio estimation. Advances in Neural Information Processing Systems, 34: 0 129--143, 2021

    2021

  24. [24]

    Awesome Neural SBI , January 2023

    Siddharth Mishra-Sharma. Awesome Neural SBI , January 2023. URL https://github.com/smsharma/awesome-neural-sbi

    2023

  25. [25]

    A universal density profile from hierarchical clustering

    Julio F Navarro, Carlos S Frenk, and Simon DM White. A universal density profile from hierarchical clustering. The Astrophysical Journal, 490 0 (2): 0 493--508, 1997

    1997

  26. [26]

    AlphaEvolve : A coding agent for scientific and algorithmic discovery

    Alexander Novikov, Ng \^a n V \ u , Marvin Eisenberger, et al. AlphaEvolve : A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131, 2025

    Pith/arXiv arXiv 2025

  27. [27]

    Fast -free inference of simulation models with Bayesian conditional density estimation

    George Papamakarios and Iain Murray. Fast -free inference of simulation models with Bayesian conditional density estimation. Advances in Neural Information Processing Systems, 29, 2016

    2016

  28. [28]

    Masked autoregressive flow for density estimation

    George Papamakarios, Theo Pavlakou, and Iain Murray. Masked autoregressive flow for density estimation. Advances in Neural Information Processing Systems, 30, 2017

    2017

  29. [29]

    Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows

    George Papamakarios, David Sterratt, and Iain Murray. Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows. The 22nd International Conference on Artificial Intelligence and Statistics, pages 837--848, 2019

    2019

  30. [30]

    PyTorch : An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K \"o pf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch : An imperative style, high-...

    2019

  31. [31]

    Mathematical discoveries from program search with large language models

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models. Nature, 625 0 (7995): 0 468--475, 2024

    2024

  32. [32]

    Bayesianly justifiable and relevant frequency calculations for the applied statistician

    Donald B Rubin. Bayesianly justifiable and relevant frequency calculations for the applied statistician. The Annals of Statistics, 12 0 (4): 0 1151--1172, 1984

    1984

  33. [33]

    LLM-SR : Scientific equation discovery via programming with large language models

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K Reddy. LLM-SR : Scientific equation discovery via programming with large language models. arXiv preprint arXiv:2404.18400, 2024

    Pith/arXiv arXiv 2024

  34. [34]

    OpenAI GPT-5 system card

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. OpenAI GPT-5 system card. arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    Handbook of approximate Bayesian computation

    Scott A Sisson, Yanan Fan, and Mark Beaumont. Handbook of approximate Bayesian computation . CRC Press, 2018

    2018

  36. [36]

    Validating Bayesian inference algorithms with simulation-based calibration

    Sean Talts, Michael Betancourt, Daniel Simpson, Aki Vehtari, and Andrew Gelman. Validating Bayesian inference algorithms with simulation-based calibration. arXiv preprint arXiv:1804.06788, 2018

    Pith/arXiv arXiv 2018

  37. [37]

    sbi : A toolkit for simulation-based inference

    Alvaro Tejero-Cantero, Jan Boelts, Michael Deistler, Jan-Matthis Lueckmann, Conor Durkan, Pedro J Gon c alves, David S Greenberg, and Jakob H Macke. sbi : A toolkit for simulation-based inference. Journal of Open Source Software, 5 0 (52): 0 2505, 2020

    2020

  38. [38]

    Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems

    Tina Toni, David Welch, Natascha Strelkowa, Andreas Ipsen, and Michael PH Stumpf. Approximate Bayesian computation scheme for parameter inference and model selection in dynamical systems. Journal of the Royal Society Interface, 6 0 (31): 0 187--202, 2009

    2009

  39. [39]

    Strong gravitational lensing as a probe of dark matter

    Simona Vegetti, Simon Birrer, Giulia Despali, Christopher D Fassnacht, Daniel Gilman, Yashar Hezaveh, Laurence Perreault Levasseur, John P McKean, Devon M Powell, Conor M O'Riordan, and Georgios Vernardos. Strong gravitational lensing as a probe of dark matter. Space Science Reviews, 220 0 (5): 0 58, 2024

    2024

  40. [40]

    Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St \'e fan J

    Pauli Virtanen, Ralf Gommers, Travis E. Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, St \'e fan J. van der Walt , Matthew Brett, Joshua Wilson, K. Jarrod Millman, Nikolay Mayorov, Andrew R. J. Nelson, Eric Jones, Robert Kern, Eric Larson, C J Carey, \.I lhan Polat, Yu Feng, Er...

    2020

  41. [41]

    A strong gravitational lens is worth a thousand dark matter halos: Inference on small-scale structure using sequential methods

    Sebastian Wagner-Carena, Jaehoon Lee, Jeffrey Pennington, Jelle Aalbers, Simon Birrer, and Risa H Wechsler. A strong gravitational lens is worth a thousand dark matter halos: Inference on small-scale structure using sequential methods. The Astrophysical Journal, 975 0 (2): 0 297, 2024

    2024

  42. [42]

    Macke, and Daniel Gedon

    Stefan Wahl, Raphaela Schenk, Ali Farnoud, Jakob H. Macke, and Daniel Gedon. A probabilistic framework for LLM -based model discovery. arXiv preprint arXiv:2602.18266, 2026

    Pith/arXiv arXiv 2026

  43. [43]

    Manzil Zaheer, Satwik Kottur, Siamak Ravanbakhsh, Barnab \'a s P \'o czos, Ruslan Salakhutdinov, and Alexander J. Smola. Deep sets. Advances in Neural Information Processing Systems, 30, 2017

    2017