Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification

Kaito Takano; Kei Nakagawa; Masanori Hirano

REVIEW 4 major objections 6 minor 36 references

Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Modeling FOMC decisions as a multi-agent LLM debate with hawkish-dovish beliefs improves prediction accuracy over single-shot baselines.

desk verdict A sensible application of multi-agent debate to FOMC prediction that overstates its empirical support; the central claim of 'significant' improvement doesn't hold with one run on 60 hand-picked meetings. read the letter →

arxiv 2511.02469 v1 pith:EIHZT225 submitted 2025-11-04 q-fin.CP cs.AIcs.MA

Kaito Takano , Masanori Hirano , Kei Nakagawa This is my paper

classification q-fin.CPcs.AIcs.MA

keywords FOMCmonetarypolicymulti-agentLLMdebatehawkish-dovishlatentbeliefBeigeBookclassification

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the Federal Reserve's rate decisions are better predicted by simulating the FOMC's internal debate than by asking a single model to classify the same inputs. It builds a committee of seven LLM agents, each assigned a belief profile from strongly hawkish to strongly dovish, who read the Beige Book and macroeconomic indicators, announce an initial Raise/Hold/Lower call, and then revise their calls after seeing each other's outputs over up to ten rounds. A latent 'belief' variable is formalized as mediating how inputs and peer opinions translate into each agent's vote. On a hand-selected set of 60 meetings the full debate model reaches an F1 of 0.48, outperforming versions that drop the debate, the beliefs, the Beige Book, or the macro indicators. The authors argue the Beige Book and the debate itself matter most, with debate shrinking the strong Hold bias observed when agents never interact.

What carries the argument

The latent-belief decomposition (Lemmas 1-2) plus the iterative debate loop. The belief space theta has five stances (strong/moderate hawkish, neutral, moderate/strong dovish) assigned across seven agents; a Bayesian generative story says each agent's vote is drawn from a label distribution conditioned on its stance, while the stance posterior updates from inputs and peers' prior votes. The debate runs up to 10 rounds, halting early on full agreement, with majority vote as final output. This machinery is what converts a collection of independent LLM calls into a deliberative committee whose internal states are interpretable.

What would settle it

Run the same prompts on the full universe of FOMC meetings from 2000-2025 without excluding 2009-2015 or consecutive changes, using a fixed random seed and reporting per-class F1; if single-shot or belief-free baselines match the debate model's F1 on the full universe, the central claim of debate superiority fails.

Watch

Extended reading notes

Core claim

The central claim is that an explicit, belief-mediated multi-agent debate reproduces FOMC outcomes better than static LLM classification. Each agent's output is modeled as P(z|x,v,Z,phi) = sum_theta P(z|theta,phi) P(theta|x,v,Z,phi), where theta is a discrete hawkish-dovish stance. Under the assumption that the label is conditionally independent of inputs once theta is fixed, the posterior over stances factors into the likelihood of the observed text and numbers, the agent's prior, and how well the other agents' previous labels support that stance. The paper reports F1 of 0.476 for the full model versus 0.385 without the Beige Book, 0.399 without belief profiles, and 0.415 without debate rou

Load-bearing premise

The evaluation relies on a hand-filtered, class-balanced set of 60 meetings that excludes the 2009-2015 period and all consecutive rate changes; if that selection is not representative of the full FOMC decision distribution, the reported accuracy ordering may not hold in real-time forecasting.

Editorial extensions

If this is right

If correct, central-bank watchers can treat FOMC forecasts as an emergent property of belief diversity plus debate, not just text sentiment.
Ablation order suggests qualitative Beige Book text is more load-bearing than the historical rate path for these predictions.
The debate loop reduces the models' default Hold bias compared with round-0 majority voting.
The belief posterior gives a per-meeting, per-agent interpretable stance that could be tracked over time.
The framework transfers to other committees (ECB, BoJ), as the authors state as future work.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The paper's 60-meeting selection excludes entire regimes (2009-2015 ZLB, consecutive hikes) and balances classes artificially; real-time forecasting would face different base rates, and the reported F1 gap may narrow outside the curated sample.
The theoretical decomposition is not actually estimated as a latent-variable model: the beliefs are fixed prompt labels and the posterior is never computed, so the Bayesian formalism is a post-hoc interpretation rather than a fitted mechanism.
A testable extension: vary the belief mixture (e.g., an all-hawkish committee) and measure whether final predictions shift in the expected direction; this would validate whether the belief manipulation is causal or just noise.
Comparing against a single LLM prompted to 'consider both hawkish and dovish views' would isolate whether the benefit comes from the multi-agent structure or just from broader prompting.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this paper isn't a waste of anyone's time, but the headline claim should be read as 'we tried it and saw a modest improvement,' not 'significantly outperforms.' The modeling idea is straightforward: run several LLM agents with hawkish/dovish belief prompts, let them debate for a few rounds, and aggregate a majority vote. That's a legitimate extension of existing multi-agent debate work to a new domain, and the authors correctly identify the FOMC's deliberative structure as a natural fit. The Beige Book ablation is the most interesting result, and the observation that debate shifts the distribution away from a default Hold is a real, if unsystematic, finding.

The empirical core, however, is thinner than the abstract claims. The comparison rests on a single run of each configuration on a hand-filtered set of 60 meetings. Temperature is 1, so outputs are stochastic, yet there are no repeated runs, confidence intervals, or significance tests. Observed F1 gaps (0.476 vs. 0.415 vs. 0.399) are within plausible noise. The paper also cites Experiment (3) – removing macro indicators – in the text and conclusion, but Table 2 doesn't show it. That's a concrete reporting gap, not an interpretive quibble.

The theoretical section is mostly decoration. Lemma 1 is a marginalization identity, and Lemma 2 rests on an independence assumption that directly contradicts the implemented prompt, where the belief string is fixed and the model still consumes the text and other agents' outputs. The 'latent belief' is not inferred; it's a hand-assigned label. So the framing as a Bayesian generative model adds no explanatory power.

There are smaller issues: the sample excludes 2009–2015 and consecutive rate changes, the filtering rules aren't fully specified, and no code or data is released. These are fixable in revision. The paper would be more honest as a feasibility study with clearly labeled preliminary results and no significance language. It's a reasonable candidate for a serious referee at an applied NLP or financial AI venue, but only with the expectation of major revisions.

Referee Report

4 major / 6 minor

Summary. The paper proposes a multi-agent LLM debate framework for classifying FOMC policy rate decisions (Raise/Hold/Lower). Each of seven agents is assigned an explicit 'belief profile' (e.g., Strong Hawkish, Dovish) and receives the Beige Book text, macroeconomic indicators, and historical policy rates. Agents exchange predictions over up to ten rounds, and a final decision is made by majority vote or consensus. The authors also present a Bayesian latent-variable formulation intended to formalize how beliefs mediate inputs and debate. Empirically, they evaluate the method on 60 hand-selected FOMC meetings from 2000–2025 and compare against ablations that remove the Beige Book, macroeconomic indicators, historical policy rates, belief profiles, and debate. The full model reports a macro-F1 of 0.476, which the abstract and conclusion describe as significantly outperforming baselines.

Significance. If the empirical claims were supported, the framework would extend multi-agent LLM debate to a consequential financial time-series classification task, and the explicit belief-profile design offers a transparent way to interpret agent behavior. The paper is clearly written and its ablation structure is sensible. However, the central empirical claim of statistically significant superiority rests on a single stochastic run over a small, hand-filtered sample, and the theoretical latent-variable apparatus is not actually estimated. The contribution is potentially interesting but, as it stands, the evidence does not justify the 'significantly outperforms' language.

major comments (4)

[§5.4, Table 2] Experiment (3) ('remove macroeconomic indicators'), defined in §5.3, is missing from Table 2. The text in §5.4 states that 'results from Experiment (3) confirm that both quantitative and textual macroeconomic indicators are crucial,' and the conclusion repeats this claim. Without the corresponding F1/precision/recall row, this conclusion is unsupported and the ablation comparison is incomplete. This is load-bearing because the relative ranking of ablations is the paper's core empirical contribution.
[§5.2, Table 2, Abstract] The claim that the proposed method 'significantly outperforms' baselines is not statistically substantiated. Temperature is set to 1, making outputs stochastic, yet each configuration is run only once. On n=60 observations, macro-F1 differences of 0.476 vs 0.415 (No Debate) and 0.399 (Remove Belief) may be within sampling noise; no confidence intervals, significance tests, or repeat runs are reported. The word 'significantly' in the abstract and conclusion is therefore unjustified. The authors should run multiple seeded trials, report variance, and apply appropriate tests (e.g., bootstrap or paired tests) to support the ranking.
[§5.2] The evaluation sample is a hand-filtered, class-balanced subset of 60 meetings (15 Raise, 30 Hold, 15 Lower), with periods such as 2009–2015 and consecutive rate changes excluded as 'too trivial.' The paper does not report the full meeting universe, the number of excluded cases, or the selection algorithm/seed. This non-transparent filtering undermines the generality of the 'significantly outperforms' claim and risks severe selection bias. The authors should justify the exclusions, report results on the full available sample or at least quantify how many meetings are dropped and why.
[§4.1, Assumption 1, Lemma 2] The theoretical framework is disconnected from the implementation. The 'latent belief' θ is not inferred from data; it is a fixed string inserted into the prompt by the authors (Table 1, §5.2). Assumption 1 (Eq. 11) claims the output is conditionally independent of x, v, and Z given θ, but the actual prompts in §5.2 feed x, v, and Z to the agent along with the belief, so the assumption is directly contradicted by the method. Moreover, the posterior in Eq. (13) is never computed in the experiments. The theory section therefore does not provide evidence for a 'mediating role' of beliefs; it is a tautological decomposition (Lemma 1) followed by an untested assumption. The authors should either reframe the belief profiles as explicit conditions, not latent variables, or show how the posterior is estimated and used.

minor comments (6)

[§1] Typo: 'different beliefs' should be 'different beliefs' (grammar) or 'different belief profiles.'
[§5.4] The text references 'Figure 5' for policy decision changes, but the paper contains only Table 5 (Transition Matrix). Re-number or add the figure.
[Appendix A.2] The citation 'Assumption 11' should be 'Assumption 1' (or Eq. 11) to avoid confusion.
[§5.3] The 'Remove Text' prompt contains a formatting artifact: 'Youwillbegivenbeigeassociatedmacroeconomicnumericaldata' is missing spaces. Please proofread all prompt variants.
[References] Reference [31] is duplicated in the text as '[31,31]'. Also, some entries (e.g., [12]) are incomplete.
[§5.2] The experiments use a single LLM family (GPT-4o-mini). The conclusion acknowledges this as a limitation, but the paper would be stronger if at least one additional model were tested for robustness.

Circularity Check

1 steps flagged · score 4.0 of 10

Belief-mediation 'theorem' is Assumption 1 restated and the belief is a prompt input; main F1 ablation is independent.

self definitional [Section 4.1, Assumption 1 and Lemma 2 (Eqs. 11-13); Section 5.2 prompt template]
"Assumption 1 (Conditional Independence Given Latent Belief): ... P(z_i^{(t+1)} | θ, x, v, Z^{(t)}, φ_i) = P(z_i^{(t+1)} | θ, φ_i) (∀θ∈Θ, t≥0) ... Under this assumption, the output probability ... reduces to: P_LLM(z_i^{(t+1)} | x, v, Z^{(t)}, φ_i) = Σ_{θ∈Θ} P(z_i^{(t+1)} | θ, φ_i) P(θ | x, v, Z^{(t)}, φ_i)."

Lemma 2, offered as the theoretical demonstration that the latent belief mediates input perception and interaction, is not derived from properties of LLMs; it is Assumption 1 restated. Assumption 1 literally asserts that x, v, and Z^(t) are conditionally independent of the output once θ is known, which is the definition of θ as a full mediator. The posterior formula (13) is then obtained by adding further conditional-independence assumptions and is never estimated or used in the experiments. In the experiments θ is not latent: it is supplied by the authors as the prompt field 'Belief: {Beliefk}' with fixed belief profiles, so the reported belief-behavior associations are induced by the prompt. The mediation 'result' therefore reduces to its own input assumption by construction.

full rationale

The empirical prediction chain (prompts -> majority vote -> F1) involves no fitted parameters and no derived predictions: Eqs. (2)-(5) are sampling statements, and Table 2 is an ablation comparison over actual inputs. That comparison is self-contained and not circular, although it is statistically underpowered. The circular step is confined to the interpretability claim: the paper says it 'theoretically demonstrates' that the latent belief mediates input and interaction, but Lemma 1 is just the law of total probability, and Lemma 2 follows directly from Assumption 1, which is the mediation claim itself. The posterior in Eq. (13) is never estimated, and the 'latent' belief is in fact an author-assigned prompt string in Section 5.2. This partially undermines the beliefs-shape-forecasts interpretation, but it does not make the F1 ablation circular, because the ablations vary real inputs rather than derived quantities. The only author-overlapping citation, [33], is used as a dataset source and is not load-bearing for the central claim; [9] is external. Missing significance tests and the omitted (3) ablation row are validity concerns, not circularity. Score 4 rather than 6+ because the central empirical claim retains independent content.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claim depends on hand-set hyperparameters (agent count, belief profile, temperature, sample filtering) and on strong conditional-independence assumptions that are not validated. The 'latent' belief is actually an observed prompt condition, so the theoretical mediation story is not empirically grounded.

free parameters (4)

Sample class balance and exclusion rules = 15 Raise / 30 Hold / 15 Lower; exclude 2009-2015 and consecutive rate changes
Chosen to make the task non-trivial and balanced; determines the measured accuracy and is not justified against the full FOMC distribution.
Agent count and belief profile = n=7; 1 Strong Hawkish, 1 Moderately Hawkish, 3 Neutral, 1 Moderately Dovish, 1 Strong Dovish
Hand-set to facilitate majority voting; no sensitivity analysis is reported.
Temperature and maximum rounds = temperature=1, T=10
Chosen for diverse debate patterns; no sensitivity analysis is reported.
Belief prompt descriptions = Table 1 wording for each belief category
Hand-authored; these descriptions could encode answer biases and are not varied.

assumptions (5)

ad hoc to paper Assumption 1: given the latent belief theta, an agent's output is independent of the input text x, numerical data v, and previous responses Z^(t)
Makes Lemma 2 tractable but is contradicted by the Section 5.2 prompt, which always includes text, indicators, rates, and other predictions alongside the stated belief.
ad hoc to paper Conditional independence of x, v, and Z^(t) given theta, and of the n previous outputs given theta (Eqs. 17 and 19)
Additional independence assumptions introduced inside the Appendix proof; never checked or mentioned in the main text.
standard math Law of total probability and Bayes' theorem
Used in the proofs of Lemma 1 and Lemma 2.
domain assumption GPT-4o-mini zero-shot responses with assigned beliefs are a faithful proxy for FOMC members' deliberation
The entire simulation rests on this; the paper provides no validation against actual FOMC member behavior.
domain assumption Beige Book 'overall economic activity'/'summary' sentences plus unemployment and inflation capture the information relevant to rate decisions
Input selection is limited to these sources; other potentially relevant indicators are ignored without discussion.

invented entities (1)

Hawkish-dovish latent belief theta
purpose: Claimed mediating variable that interprets inputs and others' outputs into a policy label
The belief is never inferred from data; it is a fixed string inserted in the prompt, so the stance-dependent behavior is partly constructed by the authors rather than discovered.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification." pith.science (2026). https://pith.science/paper/EIHZT225

@misc{pith2026251102469,
  author       = {Pith},
  title        = {Pith review of: Modeling Hawkish-Dovish Latent Beliefs in Multi-Agent Debate-Based LLMs for Monetary Policy Decision Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EIHZT225}},
  note         = {Machine review of arXiv:2511.02469}
}

read the original abstract

Accurately forecasting central bank policy decisions, particularly those of the Federal Open Market Committee(FOMC) has become increasingly important amid heightened economic uncertainty. While prior studies have used monetary policy texts to predict rate changes, most rely on static classification models that overlook the deliberative nature of policymaking. This study proposes a novel framework that structurally imitates the FOMC's collective decision-making process by modeling multiple large language models(LLMs) as interacting agents. Each agent begins with a distinct initial belief and produces a prediction based on both qualitative policy texts and quantitative macroeconomic indicators. Through iterative rounds, agents revise their predictions by observing the outputs of others, simulating deliberation and consensus formation. To enhance interpretability, we introduce a latent variable representing each agent's underlying belief(e.g., hawkish or dovish), and we theoretically demonstrate how this belief mediates the perception of input information and interaction dynamics. Empirical results show that this debate-based approach significantly outperforms standard LLMs-based baselines in prediction accuracy. Furthermore, the explicit modeling of beliefs provides insights into how individual perspectives and social influence shape collective policy forecasts.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 2 linked inside Pith

[1]

Journal of conflict resolution41(2), 203–226 (1997)

Axelrod, R.: The dissemination of culture: A model with local convergence and global polarization. Journal of conflict resolution41(2), 203–226 (1997)

1997
[2]

Financial Stability Review (Autumn) (2024)

Balsategui, I., Gorjón, S., Marqués, J.M.: Artificial intelligence in the financial sys- tem: implications and progress from a central bank perspective. Financial Stability Review (Autumn) (2024)

2024
[3]

International Monetary Fund (2024)

Brandão-Marques, M.L., Meeks, M.R., Nguyen, V.: Monetary Policy with Uncer- tain Inflation Persistence. International Monetary Fund (2024)

2024
[4]

Journal of monetary economics71, 119–132 (2015)

Bruno, V., Shin, H.S.: Capital flows and the risk-taking channel of monetary policy. Journal of monetary economics71, 119–132 (2015)

2015
[5]

In: The Twelfth International Conference on Learning Representations (2024)

Chan, C.M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., Fu, J., Liu, Z.: Chateval: Towards better LLM-based evaluators through multi-agent debate. In: The Twelfth International Conference on Learning Representations (2024)

2024
[6]

Journal of Macroeco- nomics67, 103279 (2021)

Couture, C.: Financial market effects of fomc projections. Journal of Macroeco- nomics67, 103279 (2021)

2021
[7]

In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics

Devlin,J.,Chang,M.W.,Lee,K.,Toutanova,K.:BERT:Pre-trainingofDeepBidi- rectional Transformers for Language Understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics. pp. 4171–4186. Association for Computational Linguistics (2019)

2019
[8]

In: Forty-first International Conference on Machine Learning (2023)

Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factual- ity and reasoning in language models through multiagent debate. In: Forty-first International Conference on Machine Learning (2023)

2023

Show all 36 references

[9]

Advances in Neural Information Processing Systems37, 28938–28964 (2024)

Estornell, A., Liu, Y.: Multi-llm debate: Framework, principals, and interventions. Advances in Neural Information Processing Systems37, 28938–28964 (2024)

2024
[10]

Applied Economics Letters pp

Fanta, N., Horvath, R.: Artificial intelligence and central bank communication: the case of the ecb. Applied Economics Letters pp. 1–8 (2024)

2024
[11]

In: 2023 14th IIAI International Congress on Advanced Applied Informatics (IIAI-AAI)

Fujiwara, M., Suimon, Y., Nakagawa, K.: Treasury yield spread prediction with sentiments of beige book and macroeconomic data. In: 2023 14th IIAI International Congress on Advanced Applied Informatics (IIAI-AAI). pp. 337–342. IEEE (2023)

2023
[12]

Econometrics9(4), 36 (2021)

Fulton, C., Hubrich, K.: Forecasting us inflation in real time. Econometrics9(4), 36 (2021)

2021
[13]

arXiv (2024), https://arxiv.org/abs/2403.03336

Gatto, J., Basak, M., Srivastava, Y., Bohlman, P., Preum, S.M.: Scope of large language models for mining emerging opinions in online health discourse. arXiv (2024), https://arxiv.org/abs/2403.03336

2024 arXiv
[14]

The Quarterly Journal of Economics 133(2), 801–870 (2018)

Hansen, S., McMahon, M., Prat, A.: Transparency and deliberation within the fomc: A computational linguistics approach. The Quarterly Journal of Economics 133(2), 801–870 (2018)

2018
[15]

Research Bulletin50(2018)

Jarociński, M., Karadi, P.: The macroeconomic impact of news about policy and news about the economy in ecb announcements. Research Bulletin50(2018)

2018
[16]

Journal of monetary Economics48(2), 339–372 (2001)

Kim, S.: International transmission of us monetary policy shocks: Evidence from var’s. Journal of monetary Economics48(2), 339–372 (2001)

2001
[17]

(eds.): Language Models are Few-Shot Learners, vol

Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H. (eds.): Language Models are Few-Shot Learners, vol. 33. Curran Associates, Inc. (2020)

2020
[18]

In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing

Liang, T., He, Z., Jiao, W., Wang, X., Wang, Y., Wang, R., Yang, Y., Shi, S., Tu, Z.: Encouraging divergent thinking in large language models through multi-agent debate. In: Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. pp. 17889–17904 (2024)

2024
[19]

López-Moctezuma, G.: Sequential deliberation in collective decision-making: The case of the fomc (2016) Modeling Hawkish-Dovish Latent Beliefs 17

2016
[20]

Nature397(6719), 498–500 (1999)

Lux, T., Marchesi, M.: Scaling and Criticality in a Stochastic Multi-agent Model of a Financial Market. Nature397(6719), 498–500 (1999)

1999
[21]

Finance Research Letters49, 103130 (2022)

Nakagawa, K., Suimon, Y.: Inflation rate tracking portfolio optimization method: Evidence from japan. Finance Research Letters49, 103130 (2022)

2022
[22]

The Quarterly Journal of Economics133(3), 1283–1330 (2018)

Nakamura, E., Steinsson, J.: High-frequency identification of monetary non- neutrality: the information effect. The Quarterly Journal of Economics133(3), 1283–1330 (2018)

2018
[23]

OpenAI: ChatGPT (2023), https://openai.com/blog/chatgpt/

2023
[24]

OpenAI: GPT-4 Technical Report (2023), https://arxiv.org/abs/2303.08774

2023 arXiv
[25]

In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology

Park, J.S., O’Brien, J., Cai, C.J., Morris, M.R., Liang, P., Bernstein, M.S.: Gener- ative agents: Interactive simulacra of human behavior. In: Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. UIST ’23, Association for Computing Machinery...

2023
[26]

Qu, Y., Wang, J.: Performance and biases of large language models in public opin- ionsimulation.HumanitiesandSocialSciencesCommunications11(1),1–13(2024)

2024
[27]

Radford, A., Narasimhan, K., Salimans, T., Sutskever, I.: Improving Language Un- derstanding by Generative Pre-Training (2018), https://cdn.openai.com/research- covers/language-unsupervised/language_understanding_paper.pdf

2018
[28]

Radford, A., Wu, J.,Child,R., Luan, D.,Amodei,D.,Sutskever, I.:Language Mod- els are Unsupervised Multitask Learners (2019), https://cdn.openai.com/better- language-models/language_models_are_unsupervised_multitask_learners.pdf

2019
[29]

American economic review90(3), 429–457 (2000)

Romer,C.D.,Romer,D.H.:Federalreserveinformationandthebehaviorofinterest rates. American economic review90(3), 429–457 (2000)

2000
[30]

Financial Management 48(4), 1069–1094 (2019)

Routledge, B.R.: Machine learning and asset allocation. Financial Management 48(4), 1069–1094 (2019)

2019
[31]

The American Economic Review59(2), 488–493 (1969)

Schelling, T.C.: Models of Segregation. The American Economic Review59(2), 488–493 (1969)

1969
[32]

MIT press (2013)

Schonhardt-Bailey, C.: Deliberating American monetary policy: A textual analysis. MIT press (2013)

2013
[33]

Proceedings of The 37th Annual Conference of the Japanese Society for Artificial Intelligence pp

Takano, K., Hasegawa, N., Naito, A., Nakagawa, K.: Construction of the frb beige book corpus and analysis (japanese). Proceedings of The 37th Annual Conference of the Japanese Society for Artificial Intelligence pp. 3Xin4–33 (2023)

2023
[34]

Takata,R.,Masumori,A.,Ikegami,T.:Spontaneousemergenceofagentindividual- itythroughsocialinteractionsinlargelanguagemodel-basedcommunities.Entropy 26(12), 1092 (2024)

2024
[35]

In: Advances in Neural Information Processing Systems

Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, L.u., Polosukhin, I.: Attention is all you need. In: Advances in Neural Information Processing Systems. vol. 30. Curran Associates, Inc
[36]

arXiv (2024)

Yu, H., Hong, Z., Cheng, Z., Zhu, K., Xuan, K., Yao, J., Feng, T., You, J.: Re- searchtown: Simulator of human research community. arXiv (2024)

2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Journal of conflict resolution41(2), 203–226 (1997)

Axelrod, R.: The dissemination of culture: A model with local convergence and global polarization. Journal of conflict resolution41(2), 203–226 (1997)

1997

[2] [2]

Financial Stability Review (Autumn) (2024)

Balsategui, I., Gorjón, S., Marqués, J.M.: Artificial intelligence in the financial sys- tem: implications and progress from a central bank perspective. Financial Stability Review (Autumn) (2024)

2024

[3] [3]

International Monetary Fund (2024)

Brandão-Marques, M.L., Meeks, M.R., Nguyen, V.: Monetary Policy with Uncer- tain Inflation Persistence. International Monetary Fund (2024)

2024

[4] [4]

Journal of monetary economics71, 119–132 (2015)

Bruno, V., Shin, H.S.: Capital flows and the risk-taking channel of monetary policy. Journal of monetary economics71, 119–132 (2015)

2015

[5] [5]

In: The Twelfth International Conference on Learning Representations (2024)

Chan, C.M., Chen, W., Su, Y., Yu, J., Xue, W., Zhang, S., Fu, J., Liu, Z.: Chateval: Towards better LLM-based evaluators through multi-agent debate. In: The Twelfth International Conference on Learning Representations (2024)

2024

[6] [6]

Journal of Macroeco- nomics67, 103279 (2021)

Couture, C.: Financial market effects of fomc projections. Journal of Macroeco- nomics67, 103279 (2021)

2021

[7] [7]

In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics

Devlin,J.,Chang,M.W.,Lee,K.,Toutanova,K.:BERT:Pre-trainingofDeepBidi- rectional Transformers for Language Understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics. pp. 4171–4186. Association for Computational Linguistics (2019)

2019

[8] [8]

In: Forty-first International Conference on Machine Learning (2023)

Du, Y., Li, S., Torralba, A., Tenenbaum, J.B., Mordatch, I.: Improving factual- ity and reasoning in language models through multiagent debate. In: Forty-first International Conference on Machine Learning (2023)

2023