Automated Synthesis and Adversarial Validation of Executable Causal Research Pipelines

Dan Zeltzer; Irena Girshovitz; Ran Gilad-Bachrach

REVIEW 3 major objections 3 minor 18 references

Automated causal analysis is safer when it makes invalid assumptions visible rather than returning silent estimates.

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

2026-08-01 08:13 UTC pith:V4DNA5K4

load-bearing objection A validity-first causal-analysis pipeline that trades benchmark accuracy for visible failure modes; the safety claim is plausible but rests on self-generated validity flags. the 3 major comments →

arxiv 2607.21173 v1 pith:V4DNA5K4 submitted 2026-07-23 cs.LG

Automated Synthesis and Adversarial Validation of Executable Causal Research Pipelines

Irena Girshovitz , Dan Zeltzer , Ran Gilad-Bachrach This is my paper

classification cs.LG

keywords causal inferencelarge language modelsstructural causal modelssynthetic dataadversarial validationtest-driven developmentautomated sciencevalidity flags

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 main risk in automated causal analysis is not wrong numbers but silent wrongness, and that a pipeline can convert silent failures into visible ones. It introduces ARA, an AI-based Epidemiology Research Assistant, a multi-agent framework that first writes a causal protocol, then generates synthetic datasets from Structural Causal Models with known ground-truth effects, including deliberately violated identification assumptions, and only then produces analysis code validated against those synthetic tests. On 33 observational studies with explicit and vague prompts, ARA did not consistently beat a direct LLM baseline at matching benchmark estimates. But it changed the error profile: no sign flips at the final checkpoint, more confidence intervals that included zero or were withheld, and explicit protocol-level concerns instead of unqualified causal claims. The paper argues that validity-first automated science systems should be judged by whether they indicate when a causal answer is unwarranted, not only by whether they return the benchmark answer.

Core claim

ARA's central claim is that a test-driven validation loop can make causal-validity failures visible. Rather than treating analysis code as finished once it executes, ARA encodes the study protocol as a formal artifact, turns each identification assumption into a synthetic-data adversarial test with known ground truth, and requires the generated code to pass baseline and violation scenarios before deployment. Across 33 benchmark cases, the authors report that at the final checkpoint ARA produced no confidence-interval sign flips in either explicit or vague conditions, while the vanilla baseline produced one sign flip in the vague condition. ARA more often returned intervals including zero or

What carries the argument

The load-bearing mechanism is the test-driven validation framework. A Principal Investigator agent and a Reviewer agent iterate until a structured protocol is approved; the protocol is then translated into Structural Causal Models (SCMs) — families of structural equations with known causal effects — that generate a baseline dataset and up to three datasets with injected assumption violations. Analysis code is produced separately from the ground truth, run through syntactic, smoke, audit, and synthetic-ground-truth gates, and repaired in a loop until approval or an iteration budget. This converts untestable identification assumptions such as parallel trends, exclusion restrictions, and contin

Load-bearing premise

The paper's safety claim rests on trusting ARA's own AI-generated validity flags, which the authors state require manual review, to distinguish validity-driven withholding from ordinary failure, and those flags were never checked against independent ground truth.

What would settle it

Give the 33 terminal ARA outputs plus the vanilla outputs to independent epidemiologists who label each missing or non-causal output as a real validity concern, a technical failure, or a false alarm. If the model-generated flags disagree substantially with the expert labels, or if on the full benchmark ARA produces just as many sign flips as vanilla while withholding more often, the central safety-advantage claim would collapse.

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

If this is right

If ARA's central claim holds, automated causal systems should be evaluated on whether they decline or downgrade claims when assumptions fail, not only on whether they match benchmark numbers.
Synthetic-data validation can support protocol and code development when real-world data are confidential, because the pipeline is tested on known ground-truth effects before touching real data.
Validity-driven withholding means some missing confidence intervals are safety signals; aggregate metrics that penalize all missing intervals misread the system's behavior.
Stricter protocol auditing can improve methodological compliance while reducing short-run estimation convergence, so validation budgets and audit strictness must be balanced.
Treatment recovery under vague prompts improved with the pipeline (83% versus 75% for the baseline), suggesting protocol construction helps most when the user query is underspecified.

Where Pith is reading between the lines

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

I infer that the no-sign-flip result is fragile as a headline number — one event in 33 cases cannot separate the systems — but the qualitative shift toward zero-inclusive and withheld intervals is the more durable and testable claim.
A testable extension would be to measure the precision of ARA's validity flags against independent expert review; if many flags are false alarms, the safety advantage becomes 'fails loudly' rather than 'detects correctly'.
The same synthetic stress-test loop could be adapted to target trial emulation protocol development, since the paper's protocol-first design maps naturally onto eligibility criteria, treatment strategies, time zero, and estimands.
I further infer that the framework's value may concentrate in its approval gate: separating approved terminal outputs from budget-exhausted ones could show that approved outputs are both accurate and safe, making the audit loop the decisive component.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

A validity-first causal-analysis pipeline that trades benchmark accuracy for visible failure modes; the safety claim is plausible but rests on self-generated validity flags.

read the letter

The new thing here is the mechanism: ARA converts causal identification assumptions into executable synthetic-data tests, then runs a multi-agent loop that audits and repairs analysis code against those tests. That is a concrete step beyond cooperative refinement in Causal-Copilot or CausalAgent, and it is evaluated on an external benchmark rather than self-generated examples. The paper is also honest about the main negative result: protocol construction and adversarial validation do not consistently improve numerical agreement with benchmark estimates. The claimed contribution is a change in failure mode, not better accuracy.

What the paper does well is that it takes its own central claim seriously enough to build a failure taxonomy, report completion rates at three checkpoints, and acknowledge in Section 4.6 that validity flags are model-generated and need manual review. The directional-agreement analysis is the most safety-relevant part, and the zero sign flips versus vanilla's one is suggestive, if weak on its own. The code and prompts are provided, and the evaluation is not circular relative to the benchmark.

Now the soft spots. The central safety claim rests on the distinction between validity-driven withholding and technical failure. The taxonomy is produced by ARA's own flags, no independent ground truth is applied, and the paper explicitly states that manual review has not been done. Table 1 makes the worry concrete: ARA has higher hard-failure and incomplete-inference rates than vanilla at the final checkpoint (e.g., 33.3% hard + 36.4% incomplete for vague ARA versus 15.2% for vanilla). If a substantial share of those missing outputs are just crashes or partial runs being rationalized by the LLM, the 'zero sign flips' could mostly be a consequence of silence, not screening. The paper is aware of this and says the flags require review, but it does not perform that review. That is a load-bearing gap, not a minor one.

Other weaknesses are proportional: 33 cases, a single sign-flip difference, reliance on one LLM family for code generation, and the extra engineering cost. The limitations section is candid about all of these. Still, the central argument — that ARA changes the error profile from silent to visible — is plausible and worth testing harder, but it is not yet established with the data presented here.

Who this is for: people building or evaluating AI-assisted causal analysis tools, especially in medical data settings where silent failures are dangerous. The paper deserves serious peer review because it addresses an important problem with a concrete mechanism and honest evidence, even though the evidence is currently too weak to accept the safety claim as proven. I would send it to referees, with the expectation that they push for external validation of the failure taxonomy and a larger, more balanced sample.

Referee Report

3 major / 3 minor

Summary. The paper presents ARA, a multi-agent LLM pipeline for automated causal inference that interposes protocol construction, SCM-based synthetic data generation with injected assumption violations, and iterative code auditing/validation between a research question and executable analysis. The system is evaluated on 33 cases from the Automated Causal Reasoning Benchmark under explicit and vague prompt formulations, comparing ARA at three checkpoints against a vanilla LLM baseline. The headline result is not improved numerical agreement—indeed, agreement metrics are mixed—but a claimed change in failure mode: ARA produced no sign-flipped confidence intervals in the sample (0/33) versus one for vanilla on vague prompts, and it more often returned missing intervals, protocol concerns, or downgraded non-causal interpretations instead of silent causal estimates. The authors argue that validity-first systems should be evaluated by their ability to signal when causal claims are unwarranted, not only by benchmark reproduction.

Significance. The paper addresses a real and important problem: LLM-generated causal analyses can execute successfully while silently relying on invalid identification assumptions. ARA's design—explicit protocol artifacts, adversarial review, and synthetic-data validation—is a plausible architectural response, and the paper is admirably honest in reporting that it does not beat a vanilla baseline on numerical benchmark agreement. The central claim, if established, would be a useful contribution to evaluation methodology for AI-assisted causal inference: measuring not only answer accuracy but also visible withholding of causal claims. Strengths include code/prompt release, transparent checkpoint reporting, and explicit discussion of limitations. However, the empirical basis is a single 33-case sample, and the key safety claim rests on a taxonomy of missing outputs that is generated by the system itself and not independently validated.

major comments (3)

[§4.6, Table 1] The central claim that ARA 'changed the failure mode' depends on classifying missing or non-causal outputs as validity-driven withholding rather than hard script failure or incomplete inference. Table 1 shows ARA's final checkpoint has higher hard-failure and incomplete-inference rates than vanilla (explicit: 27.3% hard/30.3% incomplete vs. 21.2%; vague: 33.3%/36.4% vs. 15.2%). Section 4.6 concedes 'the validity flags themselves are model-generated and require manual review,' but no such manual review is reported. Without independent, expert verification that ARA's flags correspond to genuine inferential threats rather than rationalized crashes or partial outputs, the observed zero sign flips could simply reflect ARA's higher non-output rate, not genuine safety screening. Please provide case-level evidence or a manual audit distinguishing these categories, or at least report sign-flip ra
[§4.2, Figure 2; Table 2] There is a numeric inconsistency in the reported sample sizes. The text reports '0/33' sign flips for ARA in both explicit and vague conditions, but Table 2 states the ARA explicit condition had 32 runs, with the footnote 'One case was not executed due to content management policy triggering.' No reconciliation is given. If one explicit case was not run, the explicit-condition denominators in Figures 2–4 and Section 4.2 are incorrect and must be corrected. Additionally, with 33 cases and one vanilla sign flip versus zero for ARA, the difference is not statistically meaningful; no confidence interval or test is reported. The paper should quantify the uncertainty around these counts and state the exact denominator used for each rate.
[§4.4, Figure 4] The numerical agreement results are mixed—ARA improves scaled relative error on explicit prompts but is worse on vague prompts and on CI Jaccard similarity. The interpretation that this trade-off reflects 'conservative' behavior is plausible but is not formally supported. More importantly, because ARA has many more missing confidence intervals, the agreement metrics computed on evaluable outputs may be subject to selection bias: systems that return fewer intervals can look better or worse depending on which cases are omitted. A missingness-aware analysis or sensitivity bounds would strengthen the claim that the observed pattern is a genuine trade-off rather than an artifact of differential completion rates.

minor comments (3)

[§2.5.1] Typo: 'direct acyclic graph' should read 'directed acyclic graph.'
[§4.5] The definition of 'detected causal-validity concern' is broad and includes several heterogeneous phenomena. It would help to report how often each type occurred (e.g., 'unavailable instrument' vs. 'downgraded to descriptive') rather than aggregating them, especially because the claim of a changed failure mode relies on these categories.
[§4.7] The limitations paragraph is candid and welcome, but it appears after the results rather than shaping the presentation of the headline claims. Consider moving the statement about model-generated validity flags into the results section where the taxonomy is first used.

Circularity Check

1 steps flagged

ARA's safety advantage is partly self-referential: 'detected causal-validity concerns' are defined as whatever ARA flags, and the same unvalidated flags are used to reclassify missing CIs as 'validity-driven withholding'; the numerical benchmark comparisons themselves are external and not circular.

specific steps

self definitional [Section 4.5 (definition of detected concern) and Section 4.6 (taxonomy of missing CIs)]
"We define a detected causal-validity concern as a case in which the pipeline flags a violation of an identification assumption, fails a required diagnostic, downgrades a causal claim to descriptive, or states that no causal identification is possible."

The key safety outcome is defined as whatever the pipeline itself flags. The paper then uses this self-defined outcome to conclude that ARA 'surfaced protocol concerns... incomplete inference' and to classify missing confidence intervals into 'hard script failure,' 'incomplete inference,' and 'validity-driven withholding.' Because the validity flags are model-generated and the paper explicitly states they 'require manual review' without performing that review (Section 4.6), the central claim that ARA 'changed the failure mode' from silent errors to visible validity concerns is partly true by construction: ARA is credited for producing the same labels that count as evidence. The external benchmark comparisons (sign flips, numerical agreement, specification recovery) are independent and not

full rationale

Most of the paper's evaluation is external and not circular: the benchmark comparisons use the public Automated Causal Reasoning Benchmark, and the reported numerical results (agreement, sign flips, specification recovery) do not feed back into the system's design. No load-bearing self-citation or imported uniqueness theorem appears. The main circularity risk is concentrated in the interpretation of the safety advantage. The paper defines a 'detected causal-validity concern' as any case in which the pipeline itself flags a violation or downgrades a claim, then uses those flags to classify missing confidence intervals as 'validity-driven withholding' rather than technical failure. Since the flags are model-generated and the paper acknowledges they are not ground-truth labels and require manual review, the conclusion that ARA 'changed the failure mode' from silent failures to visible ones rests substantially on ARA's own self-reports. The paper is transparent about this limitation, and the raw behavioral differences (more missing CIs, zero sign flips) are observable, so the circularity is partial and moderate rather than fatal: the numerical evaluation stands, but the safety-centric interpretation is not independently validated.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The paper introduces new conceptual categories (detected causal-validity concern, validity-driven withholding) but no new physical or mathematical entities. The central claim is empirical; its main ledger entries are domain assumptions about the validity of the synthetic SCM stress tests and the trustworthiness of the model-generated validity flags.

free parameters (3)

validation iteration budget = 12
Hand-chosen cap on the validation loop; determines whether the final checkpoint is an approved script or a budget-stopped output (Section 3, Table 1).
repair attempts per iteration = 5
Hand-chosen; the fixer gets up to 5 repairs before regeneration (Section 3).
minimum review rounds = 2
Adversarial protocol loop requires at least two full reviews before approval (Section 2.4).

axioms (4)

domain assumption Synthetic SCMs constructed by the planning agent from the protocol encode the identification assumptions and their violations
Section 2.5.1: datasets are constructed as functional stress-tests; if the SCMs do not actually represent the assumptions, passing/failing synthetic tests is not informative about real-world validity.
domain assumption The CausalReasoningBenchmark ground-truth labels and estimates are correct references
Section 3: all agreement metrics compare against benchmark labels; if the benchmark itself has errors, the evaluation is skewed.
domain assumption Model-generated validity flags are meaningful signals of real causal-validity problems
Sections 4.5/4.6: the central failure-mode claim relies on treating ARA's flags (unavailable instruments, cutoffs, negative controls, etc.) as genuine concerns rather than LLM confabulation.
domain assumption The selected 33 benchmark cases are representative of the benchmark's strategy distribution
Section 3: stratified round-robin selection; the paper acknowledges small counts and that strategies may be unevenly represented.

pith-pipeline@v1.3.0-alltime-deepseek · 8972 in / 9753 out tokens · 97860 ms · 2026-08-01T08:13:26.317390+00:00 · methodology

0 comments

read the original abstract

While automated research systems promise to accelerate empirical analysis, they are prone to silent failures: instances in which analysis code executes successfully yet relies on invalid causal assumptions. We present the Artificial Intelligence (AI)-based Epidemiology Research Assistant (ARA), a framework that makes these failures visible by explicitly encoding causal design principles, study-specific assumptions, and methodological constraints. ARA integrates protocol construction, synthetic data generation, and adversarial validation into a unified pipeline. The framework translates natural language research questions into structured causal protocols and executable analysis code by first constructing a protocol and then generating synthetic datasets using Structural Causal Models (SCMs) with known ground-truth effects. This synthetic-data step can also support pipeline development when access to confidential data, such as medical data, is restricted. The generated analysis is then evaluated under controlled violations of identification assumptions. We evaluate ARA on the Automated Causal Reasoning Benchmark, assessing recovery of identification strategies, causal quantities, treatment and outcome variables, and consistency between generated code and approved protocol. Protocol construction and adversarial validation did not consistently improve numerical agreement with benchmark estimates compared with standard LLM-based generation. However, they changed the failure mode: instead of silently returning causal estimates, ARA often surfaced protocol concerns, diagnostic failures, incomplete inference, or downgraded non-causal interpretations. These findings suggest that validity-first automated science systems should be evaluated not only by answer accuracy, but also by whether they indicate when causal claims are unwarranted.

Figures

Figures reproduced from arXiv: 2607.21173 by Dan Zeltzer, Irena Girshovitz, Ran Gilad-Bachrach.

**Figure 1.** Figure 1: Schematic overview of the multi-agent test-driven causal analysis framework. The framework comprises three stages: study design, SCM-based synthetic data generation, and code generation with adversarial validation. The approved protocol is converted into a Structural Causal Model (SCM) that produces baseline and failure-scenario datasets. Generated analysis code is iteratively reviewed and validated agains… view at source ↗

**Figure 2.** Figure 2: Directional agreement between generated and benchmark confidence intervals. Rows show the generated confidence-interval relation to zero, and columns show the benchmark relation. Panels compare vanilla generation and the ARA final checkpoint for explicit and vague prompts. Green cells indicate matching direction, yellow cells indicate zero/sign mismatches, red cells indicate sign flips, and gray cells indi… view at source ↗

**Figure 3.** Figure 3: Causal specification recovery across prompt types and pipeline checkpoints. Panels show recovery of the benchmark identification strategy, causal quantity, treatment variable, and outcome variable for explicit and vague prompts. ARA is evaluated at the pre-audit, post-audit, and final-checkpoint stages; vanilla is evaluated as a direct single-script baseline. Bars report the fraction of evaluable outputs t… view at source ↗

**Figure 4.** Figure 4: Statistical agreement with benchmark estimates. Panels report scaled absolute relative error, confidence-interval Jaccard similarity, and binary confidence-interval overlap for explicit and vague prompts. Scaled absolute relative error measures point-estimate agreement, CI Jaccard measures interval similarity, and CI overlap indicates whether the generated and benchmark intervals overlap at all. These metr… view at source ↗

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 2 canonical work pages

[1]

Cambridge university press, 2009

Judea Pearl.Causality. Cambridge university press, 2009

2009
[2]

Ice cream doesn’t cause drowning: Benchmarking llms against statistical pitfalls in causal inference, 2026

Jin Du, Li Chen, Xun Xian, An Luo, Fangqiao Tian, Ganghua Wang, Charles Doss, Xiaotong Shen, and Jie Ding. Ice cream doesn’t cause drowning: Benchmarking llms against statistical pitfalls in causal inference, 2026. URLhttps://arxiv.org/abs/2505.13770

Pith/arXiv arXiv 2026
[3]

A systematic literature review on llm-based information retrieval: The issue of contents classification.KDIR, pages 135–146, 2024

Diogo Cosme, António Galvão, and Fernando Brito e Abreu. A systematic literature review on llm-based information retrieval: The issue of contents classification.KDIR, pages 135–146, 2024

2024
[4]

The virtual lab of ai agents designs new sars-cov-2 nanobodies.Nature, 646(8085):716–723, 2025

Kyle Swanson, Wesley Wu, Nash L Bulaong, John E Pak, and James Zou. The virtual lab of ai agents designs new sars-cov-2 nanobodies.Nature, 646(8085):716–723, 2025

2025
[5]

Agent laboratory: Using llm agents as research assistants.Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, 2025

Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants.Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, 2025

2025
[6]

Harnessing the power of synthetic data in healthcare: innovation, application, and privacy.NPJ digital medicine, 6(1):186, 2023

Mauro Giuffrè and Dennis L Shung. Harnessing the power of synthetic data in healthcare: innovation, application, and privacy.NPJ digital medicine, 6(1):186, 2023

2023
[7]

Causal machine learning for predicting treatment outcomes.Nature Medicine, 30(4):958–968, 2024

Stefan Feuerriegel, Dennis Frauen, Valentyn Melnychuk, Jonas Schweisthal, Konstantin Hess, Alicia Curth, Stefan Bauer, Niki Kilbertus, Isaac S Kohane, and Mihaela van der Schaar. Causal machine learning for predicting treatment outcomes.Nature Medicine, 30(4):958–968, 2024

2024
[8]

Target trial emulation: a framework for causal inference from observational data.Jama, 328(24):2446–2447, 2022

Miguel A Hernán, Wei Wang, and David E Leaf. Target trial emulation: a framework for causal inference from observational data.Jama, 328(24):2446–2447, 2022

2022
[9]

Causal mas: A survey of large language model architectures for discovery and effect estimation, 2025

Adib Bazgir, Amir Habibdoust, Yuwen Zhang, and Xing Song. Causal mas: A survey of large language model architectures for discovery and effect estimation, 2025. URL https: //arxiv.org/abs/2509.00987

Pith/arXiv arXiv 2025
[10]

Elahe Khatibi, Mahyar Abbasian, Zhongqi Yang, Iman Azimi, and Amir M. Rahmani. Alcm: Autonomous llm-augmented causal discovery framework, 2025. URL https://arxiv.org/ abs/2405.01744

Pith/arXiv arXiv 2025
[11]

Integrating large language model for improved causal discovery, 2025

Taiyu Ban, Lyuzhou Chen, Derui Lyu, Xiangyu Wang, Qinrui Zhu, Qiang Tu, and Huanhuan Chen. Integrating large language model for improved causal discovery, 2025. URL https: //arxiv.org/abs/2306.16902

Pith/arXiv arXiv 2025
[12]

An ai agent for automated causal inference in epidemiology.medRxiv, 2026

Haochen Liu, Ke Shi, Ang Li, Xinyu Li, Jiali Chu, Yingqian Xue, Sihui Cen, Yuting Wang, and Tianxiao Zhang. An ai agent for automated causal inference in epidemiology.medRxiv, 2026. doi: 10.64898/2026.02.06.26345723

work page doi:10.64898/2026.02.06.26345723 2026
[13]

Causal-copilot: An autonomous causal analysis agent.arXiv preprint arXiv:2504.13263, 2025

Xinyue Wang, Kun Zhou, Wenyi Wu, Har Simrat Singh, Fang Nan, Songyao Jin, Aryan Philip, Saloni Patnaik, Hou Zhu, Shivam Singh, et al. Causal-copilot: An autonomous causal analysis agent.arXiv preprint arXiv:2504.13263, 2025. 10

Pith/arXiv arXiv 2025
[14]

Causalagent: A conversational multi-agent system for end-to-end causal inference.arXiv preprint arXiv:2602.11527, 2026

Jiawei Zhu, Wei Chen, and Ruichu Cai. Causalagent: A conversational multi-agent system for end-to-end causal inference.arXiv preprint arXiv:2602.11527, 2026

arXiv 2026
[15]

Causal AI scientist: Facilitating causal data science with large language models

Vishal Verma, Sawal Acharya, Devansh Bhardwaj, Samuel Simko, Yongjin Yang, Anahita Haghighat, Dominik Janzing, Mrinmaya Sachan, Bernhard Schölkopf, and Zhijing Jin. Causal AI scientist: Facilitating causal data science with large language models. InNeurIPS 2025 Workshop on CauScien: Uncovering Causality in Science, 2025. URL https://openreview. net/forum?...

2025
[16]

Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915, 2021

Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M Mooij. Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915, 2021

2021
[17]

Using natural experiments to evaluate population health and health system interventions: new framework for producers and users of evidence.BMJ, 388, 2025

Peter Craig, Mhairi Campbell, Manuela Deidda, Ruth Dundas, Judith Green, Srinivasa Vittal Katikireddi, Jim Lewsey, David Ogilvie, Frank de V ocht, and Martin White. Using natural experiments to evaluate population health and health system interventions: new framework for producers and users of evidence.BMJ, 388, 2025. doi: 10.1136/bmj-2024-080505. URL htt...

work page doi:10.1136/bmj-2024-080505 2025
[18]

Causalreasoningbenchmark: A real-world benchmark for disentangled evaluation of causal identification and estimation, 2026

Ayush Sawarni, Jiyuan Tan, and Vasilis Syrgkanis. Causalreasoningbenchmark: A real-world benchmark for disentangled evaluation of causal identification and estimation, 2026. A Technical appendices and supplementary material System Specifications Design Pipeline The study protocol is produced by a two-agent LangGraph workflow in which a Principal Investiga...

2026

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.

[1] [1]

Cambridge university press, 2009

Judea Pearl.Causality. Cambridge university press, 2009

2009

[2] [2]

Ice cream doesn’t cause drowning: Benchmarking llms against statistical pitfalls in causal inference, 2026

Jin Du, Li Chen, Xun Xian, An Luo, Fangqiao Tian, Ganghua Wang, Charles Doss, Xiaotong Shen, and Jie Ding. Ice cream doesn’t cause drowning: Benchmarking llms against statistical pitfalls in causal inference, 2026. URLhttps://arxiv.org/abs/2505.13770

Pith/arXiv arXiv 2026

[3] [3]

A systematic literature review on llm-based information retrieval: The issue of contents classification.KDIR, pages 135–146, 2024

Diogo Cosme, António Galvão, and Fernando Brito e Abreu. A systematic literature review on llm-based information retrieval: The issue of contents classification.KDIR, pages 135–146, 2024

2024

[4] [4]

The virtual lab of ai agents designs new sars-cov-2 nanobodies.Nature, 646(8085):716–723, 2025

Kyle Swanson, Wesley Wu, Nash L Bulaong, John E Pak, and James Zou. The virtual lab of ai agents designs new sars-cov-2 nanobodies.Nature, 646(8085):716–723, 2025

2025

[5] [5]

Agent laboratory: Using llm agents as research assistants.Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, 2025

Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants.Findings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, 2025

2025

[6] [6]

Harnessing the power of synthetic data in healthcare: innovation, application, and privacy.NPJ digital medicine, 6(1):186, 2023

Mauro Giuffrè and Dennis L Shung. Harnessing the power of synthetic data in healthcare: innovation, application, and privacy.NPJ digital medicine, 6(1):186, 2023

2023

[7] [7]

Causal machine learning for predicting treatment outcomes.Nature Medicine, 30(4):958–968, 2024

Stefan Feuerriegel, Dennis Frauen, Valentyn Melnychuk, Jonas Schweisthal, Konstantin Hess, Alicia Curth, Stefan Bauer, Niki Kilbertus, Isaac S Kohane, and Mihaela van der Schaar. Causal machine learning for predicting treatment outcomes.Nature Medicine, 30(4):958–968, 2024

2024

[8] [8]

Target trial emulation: a framework for causal inference from observational data.Jama, 328(24):2446–2447, 2022

Miguel A Hernán, Wei Wang, and David E Leaf. Target trial emulation: a framework for causal inference from observational data.Jama, 328(24):2446–2447, 2022

2022

[9] [9]

Causal mas: A survey of large language model architectures for discovery and effect estimation, 2025

Adib Bazgir, Amir Habibdoust, Yuwen Zhang, and Xing Song. Causal mas: A survey of large language model architectures for discovery and effect estimation, 2025. URL https: //arxiv.org/abs/2509.00987

Pith/arXiv arXiv 2025

[10] [10]

Elahe Khatibi, Mahyar Abbasian, Zhongqi Yang, Iman Azimi, and Amir M. Rahmani. Alcm: Autonomous llm-augmented causal discovery framework, 2025. URL https://arxiv.org/ abs/2405.01744

Pith/arXiv arXiv 2025

[11] [11]

Integrating large language model for improved causal discovery, 2025

Taiyu Ban, Lyuzhou Chen, Derui Lyu, Xiangyu Wang, Qinrui Zhu, Qiang Tu, and Huanhuan Chen. Integrating large language model for improved causal discovery, 2025. URL https: //arxiv.org/abs/2306.16902

Pith/arXiv arXiv 2025

[12] [12]

An ai agent for automated causal inference in epidemiology.medRxiv, 2026

Haochen Liu, Ke Shi, Ang Li, Xinyu Li, Jiali Chu, Yingqian Xue, Sihui Cen, Yuting Wang, and Tianxiao Zhang. An ai agent for automated causal inference in epidemiology.medRxiv, 2026. doi: 10.64898/2026.02.06.26345723

work page doi:10.64898/2026.02.06.26345723 2026

[13] [13]

Causal-copilot: An autonomous causal analysis agent.arXiv preprint arXiv:2504.13263, 2025

Xinyue Wang, Kun Zhou, Wenyi Wu, Har Simrat Singh, Fang Nan, Songyao Jin, Aryan Philip, Saloni Patnaik, Hou Zhu, Shivam Singh, et al. Causal-copilot: An autonomous causal analysis agent.arXiv preprint arXiv:2504.13263, 2025. 10

Pith/arXiv arXiv 2025

[14] [14]

Causalagent: A conversational multi-agent system for end-to-end causal inference.arXiv preprint arXiv:2602.11527, 2026

Jiawei Zhu, Wei Chen, and Ruichu Cai. Causalagent: A conversational multi-agent system for end-to-end causal inference.arXiv preprint arXiv:2602.11527, 2026

arXiv 2026

[15] [15]

Causal AI scientist: Facilitating causal data science with large language models

Vishal Verma, Sawal Acharya, Devansh Bhardwaj, Samuel Simko, Yongjin Yang, Anahita Haghighat, Dominik Janzing, Mrinmaya Sachan, Bernhard Schölkopf, and Zhijing Jin. Causal AI scientist: Facilitating causal data science with large language models. InNeurIPS 2025 Workshop on CauScien: Uncovering Causality in Science, 2025. URL https://openreview. net/forum?...

2025

[16] [16]

Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915, 2021

Stephan Bongers, Patrick Forré, Jonas Peters, and Joris M Mooij. Foundations of structural causal models with cycles and latent variables.The Annals of Statistics, 49(5):2885–2915, 2021

2021

[17] [17]

Using natural experiments to evaluate population health and health system interventions: new framework for producers and users of evidence.BMJ, 388, 2025

Peter Craig, Mhairi Campbell, Manuela Deidda, Ruth Dundas, Judith Green, Srinivasa Vittal Katikireddi, Jim Lewsey, David Ogilvie, Frank de V ocht, and Martin White. Using natural experiments to evaluate population health and health system interventions: new framework for producers and users of evidence.BMJ, 388, 2025. doi: 10.1136/bmj-2024-080505. URL htt...

work page doi:10.1136/bmj-2024-080505 2025

[18] [18]

Causalreasoningbenchmark: A real-world benchmark for disentangled evaluation of causal identification and estimation, 2026

Ayush Sawarni, Jiyuan Tan, and Vasilis Syrgkanis. Causalreasoningbenchmark: A real-world benchmark for disentangled evaluation of causal identification and estimation, 2026. A Technical appendices and supplementary material System Specifications Design Pipeline The study protocol is produced by a two-agent LangGraph workflow in which a Principal Investiga...

2026