Pith. sign in

REVIEW 4 major objections 4 minor 24 references

Gated human-in-the-loop oversight makes AI-assisted economic theory more auditable, and evaluators prefer it in four of five matched tests.

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 07:57 UTC pith:PMECN5X4

load-bearing objection A serious, honest design paper whose comparative results are confounded in ways the authors themselves concede; worth treating as an auditability demonstration, not as an architecture-effect test. the 4 major comments →

arxiv 2607.21268 v1 pith:PMECN5X4 submitted 2026-07-23 cs.MA cs.AIecon.GNq-fin.EC

pAI-Econ-claude: A Gated Human-in-the-Loop Multi-Agent Architecture for AI-Assisted Economic Theory Development

Chen Zhu , Xiaolu Wang , Weilong Zhang This is my paper
classification cs.MA cs.AIecon.GNq-fin.EC
keywords human-in-the-loopmulti-agent systemseconomic theoryquality gatesauditabilitylarge language modelshuman oversightcanonical model library
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 asks how to organize multi-agent AI systems when no component can certify the final output — the situation in most social-science theory work. It argues that reliability should be carried by the oversight architecture, not by agent autonomy, and proposes a gated, human-in-the-loop pipeline in which every stage writes an inspectable record, gates diagnose failures and recommend loopbacks without certifying, and human checkpoints sit at decisions that are costly to reverse. In five matched economic-theory tasks, blinded evaluators preferred the gated workflow in four cases, mean failure severity fell from 1.58 to 1.16, and overall usefulness rose from 2.60 to 3.10. The clearest gain came when a reality check rejected a planted false monopoly premise and a proof review forced revision of a false welfare claim; the one negative case shows the scaffolding can compress an economically important mechanism too aggressively.

Core claim

The paper's central claim is that gated oversight improves the auditability of AI-assisted economic theory without substituting for formal verification, and that the allocation of irreversible human judgment is a more informative design variable than pure agent autonomy. Concretely, a staged architecture with a shared workspace of persistent intermediate records, diagnose-only gates, and human checkpoints produces final manuscripts that two blinded evaluators ranked above an ungated baseline in four of five tasks, with unanimity on all pairwise rankings. The largest observed gain occurred when an early reality check defeated a false market-structure premise and a proof gate prompted rewritin

What carries the argument

The load-bearing mechanism is a staged pipeline (Stages 0–10 plus sub-stages) in which each stage writes a named, inspectable file to a shared blackboard workspace. Nine quality gates follow targeted stages; each gate is itself a language model that emits only a PASS or REFRAME verdict with a failure reason, severity, and recommended loopback, where PASS means 'no targeted failure detected,' never a correctness certificate. Six scheduled human checkpoints plus conditional adjudication pauses place researcher judgment at decisions that are costly to reverse, with fixing the equilibrium concept as the one unconditional hard stop. A canonical model library and theory-lineage protocol enforce th

Load-bearing premise

The paper credits the gated architecture for the improvements, but the full workflow also received more expert human attention at checkpoints and gate adjudications, produced longer and differently scoped outputs, and ran on different model versions across task pairs; if these factors rather than the gate-and-checkpoint design drive the gains, the central claim does not follow.

What would settle it

An ablation with matched human-attention budgets and output lengths: if a gated run whose researcher spends the same review time as the baseline and whose final manuscript is capped at the baseline's length shows no improvement in blinded failure and usefulness scores, the architecture's claimed advantage is falsified. A second check: seed a false institutional premise in every task and measure whether the reality-check gate's interception rate exceeds the baseline's chance detection rate.

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

If this is right

  • In domains without a task-complete external verifier, auditability — a documented trail of assumptions, rejected alternatives, and unresolved obligations — is the achievable reliability standard, not certification.
  • Where human oversight is placed matters more than how much autonomy agents are given; concentrating attention at irreversible decisions is a transferable design principle.
  • Process records can show which check caught which error: Task 4's gains trace to a reality-check rejection and a proof-gate revision, while Tasks 2, 3, and 5 show a detect-and-narrow pattern with weaker gains.
  • Scaffolding is not monotone: the Task 1 negative case demonstrates that a fuller pipeline can produce a formally weaker and institutionally thinner manuscript than a plain baseline.
  • The workflow's value depends on gate coverage, the fit between task and canonical model library, and the quality of human adjudication, so the architecture is a reliability scaffold rather than a verification mechanism.

Where Pith is reading between the lines

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

  • The three architectural principles plausibly generalize beyond economics to any agentic pipeline whose outputs cannot be automatically verified — for instance, qualitative social science, policy analysis, or legal drafting — but the paper itself only presents this as an untested architectural hypothesis.
  • Because the full workflow also received more expert human intervention and produced differently scoped outputs, the measured advantage cannot be attributed to the gate-and-checkpoint design without ablation studies that match human-attention and output budgets; a testable extension would vary researcher involvement while holding the pipeline fixed.
  • A further testable extension: run the same five tasks with the gates' loopback recommendations automatically executed but without human adjudication; if gains persist, the human checkpoints are less responsible than claimed, and if they vanish, the human is the load-bearing element.
  • The finding that a single planted false premise was caught by the reality check suggests a concrete stress-testing protocol: deliberately seed false institutional premises into prompts and measure interception rates across tasks and model families.

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 proposes pAI-Econ-claude, a staged, human-in-the-loop multi-agent architecture for LLM-assisted economic theory development. The system decomposes theory-building into named stages, each writing persistent workspace records; quality gates diagnose targeted failure modes and recommend loopbacks without certifying correctness; and human checkpoints are placed at decisions deemed costly to reverse. The authors evaluate the full workflow against a no-skill, ungated baseline on five matched economic-theory tasks. Two blinded evaluators prefer the full workflow in four of five tasks, with mean failure severity falling from 1.58 to 1.16 and overall usefulness rising from 2.60 to 3.10. The paper presents Task 1 as an honest negative case and Task 4 as the largest positive gain, where a reality check rejected a false monopoly premise and a proof review led to revision of a false welfare claim. The authors advance a bounded claim: gated oversight improves auditability without substituting for formal verification, and the allocation of irreversible human judgment is a more informative design variable than pure agent autonomy.

Significance. If interpreted as a bounded, descriptive demonstration, the paper makes a useful contribution. It ships an open-source implementation, a concrete failure taxonomy, an inspectable workspace design, and a matched evaluation protocol with blinded endpoint ratings and detailed process traces. The honest reporting of Task 1, the explicit admission that ablations are missing, and the public release of evaluation artifacts are notable strengths. The process traces provide vivid examples of how a targeted gate can intercept a false premise (Task 4) or expose a proof gap without resolving it (Tasks 2, 3, 5). However, the causal reading of the comparative results is not justified by the design: the full workflow differs from the baseline in human-attention budget, compute usage, output scope, and bundled components, and the paper itself concedes that individual contributions cannot be isolated. The significance therefore rests on the architecture as an auditable scaffold and evaluation methodology, not on the empirical proof that gated oversight per se causes the reported improvements.

major comments (4)
  1. [§5.1 and §6] The headline claim that the gated architecture improves reliability and usefulness is not identifiable from the reported design. In the full workflow, human experts intervene at six scheduled checkpoints plus every gate-adjudication pause and can edit workspace records, whereas the baseline is a single direct request with no such interventions. The full workflow also consumes 4.6–18× the usage allowance and produces systematically longer artifacts (e.g., Task 4: twenty-page duopoly paper vs seven-page monopoly excerpt). Section 6 explicitly concedes that 'the contribution of any individual component cannot be isolated without ablation studies and matched human-attention budgets' and that 'variation in output length and completeness may have affected the comparison.' The differences in Tables 3 and 4 and the 4-of-5 preference therefore support a comparison between two full configurations,
  2. [§1, §7, Abstract] The statement that 'the allocation of irreversible human judgment is a more informative design variable than pure agent autonomy' is not tested by the evaluation. No condition varies agent autonomy while holding human-judgment placement fixed, or vice versa; the only comparison varies many components simultaneously. As the paper itself notes in §7, the broader implications are 'architectural hypotheses derived from the economic-theory case rather than empirically established results.' This sentence should be rephrased as an untested hypothesis, or a design that independently varies autonomy and human-judgment placement should be provided.
  3. [§5.1–§5.4, Table 3] The statistical basis for 'improves' is thin: five task pairs, two evaluators (one model-based, one human economist acquainted with the authors), model versions differing across pairs, and no inferential statistics or confidence intervals. The paper honestly labels the numerical differences as descriptive, but the abstract and introduction use causal language ('improves,' 'reduces failure severity') that goes beyond what a descriptive mean over five heterogeneous pairs can support. At minimum, the causal wording should be made conditional on these limitations, or additional analysis (per-dimension agreement, sensitivity to excluding Task 4, evaluator-level tests) should be reported.
  4. [§5.1, §D.1–D.5] The process-tracing narratives are produced after unblinding and are authored by the researchers who designed and ran the workflow; they are used to attribute endpoint differences to specific gates. Because the same records are also used to classify what was 'flagged,' 'resolved,' or 'missed,' the attribution of Task 4's gain to the reality-check and proof-review gates is not independent of the designers' expectations. The paper should either present the process traces as illustrative examples rather than causal evidence, or have independent coders classify gate interventions before unblinding.
minor comments (4)
  1. [§5.1] The two evaluators' exact scores differ substantially on several dimensions (e.g., T1 failure 2.20 vs 1.40; T4 failure 0.40 vs 1.00). The paper relies on pairwise ranking agreement; reporting per-dimension inter-rater agreement would strengthen the claim that the endpoint preferences are stable.
  2. [§4.2] The two demonstrations of canonical matching are illustrative, but the claim that 'the rejections are the instructive part' is not tied to evaluation data. Consider presenting these as motivating examples rather than systematic evidence.
  3. [Table 3] The column structure is confusing: 'Failure Overall' appears to combine task-level failure and overall usefulness, while 'Gap' and 'Diff.' are defined only in the caption. Clarify the table so each column's scale and direction is immediately readable.
  4. [§5.2, §D.1] The Task 1 negative case is a strength of the paper, but the recommended manual corrections in §D.1 are substantive and could be used as a concrete check on whether the workflow's final output is 'usable.' Consider adding a short discussion of whether these corrections were subsequently applied and whether they changed the evaluators' rankings.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical matched-pair comparison with blinded endpoint ratings; self-citations are contextual and Section 6's admitted confounds are identifiability limitations, not circular reductions.

full rationale

The paper's central claim is comparative and empirical: does a gated human-in-the-loop architecture produce more reliable and useful theory manuscripts than an ungated baseline on five matched tasks? The evaluation uses blinded external endpoint ratings (failure severity F1-F5 and usefulness dimensions), not a derivation that fits a target result into input assumptions. No parameter is fitted to the outcome; no equation-level prediction is constructed from the data it is later said to predict. The failure taxonomy F1-F5 informs both the architecture's gates and the scoring rubric, but the endpoint scores are assigned by evaluators to complete manuscripts independently of the workflow's internal gate verdicts, so the design-taxonomy overlap does not make the outcome self-definitional. The process-tracing evidence (flag/resolved/missed classification) is secondary and is not used to compute the primary endpoints. Self-citations [17,18] are cited as related companion work ('In companion work we develop a related pipeline...') and are not load-bearing: the architecture's effectiveness is not justified by those citations, and the main evidence is the external evaluation and released artifacts. The paper explicitly states its own limitations in Section 6: 'the full workflow combines staged decomposition, a canonical model library, quality gates, human checkpoints, and numerical analysis, while also receiving more expert intervention than the baseline, so the contribution of any individual component cannot be isolated without ablation studies and matched human-attention budgets' and 'variation in output length and completeness may have affected the comparison.' These are genuine threats to the causal attribution of the observed gains to the gate-and-checkpoint design, but they are confounds/identifiability problems, not circularity: they do not show that the measured outcome is equivalent to the input by construction. The negative case (Task 1) further shows the evaluation can go against the authors' preferred artifact, which is inconsistent with a self-fulfilling or circular design. Overall, the derivation chain is self-contained as an empirical study; the minor self-referentiality (authors as operators and interpreters of process traces) is acknowledged and mitigated by blinded endpoint evaluation, and it does not rise to circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The paper contributes an architecture and an evaluation rather than a derived law or a fitted model, so there are no free parameters and no new physical or mathematical entities. The load-bearing inputs are domain assumptions about the absence of an oracle, the completeness of the failure taxonomy, the value of irreversible human checkpoints, library coverage, and the informativeness of the two-evaluator panel.

axioms (5)
  • domain assumption LLM-generated economic theory has no cheap, task-complete, machine-readable correctness signal; auditability is therefore the achievable standard.
    Stated in §1 and in Principle P2; it motivates the entire architecture and would collapse if such an oracle existed.
  • domain assumption The five failure modes F1–F5 (canonical mismatch, trivial propositions, hidden proof gaps, interpretive overreach, unreliable citations) are a useful and sufficiently complete taxonomy for evaluating LLM-assisted theory work.
    Set out in §2.1 and Table 1; the evaluation and process-tracing analyses rely on these categories being meaningful and scoreable.
  • domain assumption Human judgment is a scarce resource that should be allocated at points of highest irreversibility; fixing the equilibrium concept is the highest-irreversibility decision.
    Principle P3, §3.3; this design principle is motivated but not empirically established by the paper.
  • domain assumption The canonical model library's general layer (sixteen entries) and human-capital sub-library provide adequate coverage for the five test tasks.
    Described in §4.1; the paper itself acknowledges in §6 that model-library coverage is bounded.
  • domain assumption Two evaluators, one model-based and one human economist acquainted with the authors, provide a meaningful measure of manuscript quality for the comparative claim.
    Evaluation protocol in §5.1; the small panel and the acquaintance are acknowledged in §6, and the results are called descriptive.

pith-pipeline@v1.3.0-alltime-deepseek · 16752 in / 13289 out tokens · 145774 ms · 2026-08-01T07:57:21.816258+00:00 · methodology

0 comments
read the original abstract

In many social-science research tasks, such as economics, LLM-based agents must produce outputs for which no cheap, task-complete, machine-readable correctness signal exists. This creates a distinctive reliability problem for multi-agent systems: how should generation, critique, coordination, and human judgment be organized when no component can certify the final result? We address this problem through pAI-Econ-claude, a gated, human-in-the-loop multi-agent architecture for AI-assisted economic theory development. Agents coordinate through a shared workspace of inspectable intermediate records; specialized gates diagnose targeted failure modes and recommend loopbacks without certifying correctness; and human checkpoints retain authority over decisions that are costly to reverse. We evaluate the architecture on five matched economic-theory tasks against an ungated baseline. Two evaluators blinded to configuration agreed on all five pairwise rankings, preferring the gated architecture in four tasks and the baseline in one. Mean failure severity fell from 1.58 to 1.16, while overall usefulness rose from 2.60 to 3.10. The largest observed gain occurred when a reality check rejected a false market-structure premise and a proof review prompted revision of a false welfare claim. The negative case shows that scaffolding can also compress an economically important mechanism too aggressively. The results support a bounded claim: gated oversight improves the auditability of AI-assisted economic theory without substituting for formal verification, and the allocation of irreversible human judgment is a more informative design variable than pure agent autonomy. The workflow is publicly available at https://github.com/maxwell2732/pAI-Econ-claude.

Figures

Figures reproduced from arXiv: 2607.21268 by Chen Zhu, Weilong Zhang, Xiaolu Wang.

Figure 1
Figure 1. Figure 1: Architecture of the pAI-Econ-claude workflow, in three phases: grounding the research puzzle, constructing [PITH_FULL_IMAGE:figures/full_fig_p005_1.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

24 extracted references · 2 linked inside Pith

  1. [1]

    The AI scientist: Towards fully automated open-ended scientific discovery, 2024

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automated open-ended scientific discovery, 2024

    2024

  2. [2]

    Agent Laboratory: Using LLM agents as research assistants

    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. InFindings of the Association for Computational Linguistics: EMNLP 2025, pages 5977–6043, Suzhou, China, 2025. Association for Computational Linguistics

    2025

  3. [3]

    Towards an AI co-scientist, 2025

    Juraj Gottweis, Wei-Hung Weng, Alexander Daryin, Tao Tu, Anil Palepu, Petar Sirkovic, et al. Towards an AI co-scientist, 2025

    2025

  4. [4]

    Matthew Gentzkow and Jesse M. Shapiro. Code and data for the social sciences: A practitioner’s guide. University of Chicago mimeo, 2014

    2014

  5. [5]

    Christensen and Edward Miguel

    Garret S. Christensen and Edward Miguel. Transparency, reproducibility, and the credibility of economics research. Journal of Economic Literature, 56(3):920–980, 2018

    2018

  6. [6]

    The use of structural models in econometrics.Journal of Economic Perspectives, 31(2):33–58, 2017

    Hamish Low and Costas Meghir. The use of structural models in econometrics.Journal of Economic Perspectives, 31(2):33–58, 2017

    2017

  7. [7]

    Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, 2023

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. Survey of hallucination in natural language generation.ACM Computing Surveys, 55(12):1–38, 2023. 10 GATEDMULTI-AGENTTHEORYDEVELOPMENTAPREPRINT- JULY24, 2026

    2023

  8. [8]

    Autonomous LLM-driven research from data to human-verifiable research papers.NEJM AI, 2(1), 2025

    Tal Ifargan, Lukas Hafner, Maor Kern, Ori Alcalay, and Roy Kishony. Autonomous LLM-driven research from data to human-verifiable research papers.NEJM AI, 2(1), 2025

    2025

  9. [9]

    From intention to implementation: Automating biomedical research via LLMs.Science China Information Sciences, 68(7):170105, 2025

    Yi Luo, Linghang Shi, Yihao Li, Aobo Zhuang, Yeyun Gong, Ling Liu, and Chen Lin. From intention to implementation: Automating biomedical research via LLMs.Science China Information Sciences, 68(7):170105, 2025

    2025

  10. [10]

    MetaGPT: Meta programming for a multi-agent collaborative framework

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. MetaGPT: Meta programming for a multi-agent collaborative framework. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  11. [11]

    ChatDev: Communicative agents for software development

    Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. ChatDev: Communicative agents for software development. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), 2024

    2024

  12. [12]

    CAMEL: Communicative agents for “mind” exploration of large language model society

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. CAMEL: Communicative agents for “mind” exploration of large language model society. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  13. [13]

    Chawla, Olaf Wiest, and Xiangliang Zhang

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V . Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges. InProceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI), 2024

    2024

  14. [14]

    AutoGen: Enabling next-gen LLM applications via multi-agent conversations

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. AutoGen: Enabling next-gen LLM applications via multi-agent conversations. InFirst Conference on Language Modeling (COLM), 2024

    2024

  15. [15]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  16. [16]

    Self-Refine: Iterative refinement with self-feedback.Advances in Neural Information Processing Systems, 36:46534–46594, 2023

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-Refine: Iterative refinement with self-feedback.Advances in Neural Information Processing Systems, 36:46534–46594, 2023

    2023

  17. [17]

    HLER: Human-in-the-loop economic research via multi-agent pipelines for empirical discovery, 2026

    Chen Zhu and Xiaolu Wang. HLER: Human-in-the-loop economic research via multi-agent pipelines for empirical discovery, 2026. arXiv:2603.07444

    arXiv 2026

  18. [18]

    (human) attention is (still) all you need: Human oversight makes AI-assisted social science reliable, 2026

    Chen Zhu, Xiaolu Wang, and Weilong Zhang. (human) attention is (still) all you need: Human oversight makes AI-assisted social science reliable, 2026. arXiv:2606.12848

    Pith/arXiv arXiv 2026

  19. [19]

    pAI/MSc: ML theory research with humans on the loop, 2026

    Mahmoud Abdelmoneum, Pierfrancesco Beneventano, and Tomaso Poggio. pAI/MSc: ML theory research with humans on the loop, 2026. arXiv:2604.20622

    Pith/arXiv arXiv 2026

  20. [20]

    Human-in-the-loop machine learning: a state of the art.Artificial Intelligence Review, 56(4):3005– 3054, 2023

    Eduardo Mosqueira-Rey, Elena Hernández-Pereira, David Alonso-Ríos, José Bobes-Bascarán, and Ángel Fernández-Leal. Human-in-the-loop machine learning: a state of the art.Artificial Intelligence Review, 56(4):3005– 3054, 2023

    2023

  21. [21]

    Human-in-the-loop software development agents

    Wannita Takerngsaksiri, Jirat Pasuksmit, Patanamon Thongtanunam, Chakkrit Tantithamthavorn, Ruixiong Zhang, Fan Jiang, Jing Li, Evan Cook, Kun Chen, and Ming Wu. Human-in-the-loop software development agents. In 2025 IEEE/ACM 47th International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP), pages 342–352. IEEE, 2025

    2025

  22. [22]

    Artifacts in the A&A meta-model for multi-agent systems

    Andrea Omicini, Alessandro Ricci, and Mirko Viroli. Artifacts in the A&A meta-model for multi-agent systems. Autonomous Agents and Multi-Agent Systems, 17(3):432–456, 2008

    2008

  23. [23]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate. InInternational Conference on Machine Learning (ICML), 2024

    2024

  24. [24]

    US Monopoly candy market

    Chen Zhu, Xiaolu Wang, and Weilong Zhang. pAI-Econ-claude: Human-in-the-loop agentic theoretical mod- eling for empirical economists. https://github.com/maxwell2732/pAI-Econ-claude , 2026. Software repository and Claude Code Skill. Appendix The appendices below provide the full task prompts, the evaluator-level score tables, and the complete task-by-task ...

    2026