Pith. sign in

REVIEW 4 major objections 6 minor 60 references

A DSGE is a structured world model: its state is the belief state learners seek, and its structure manufactures the off-path coverage pure learning cannot sample.

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 · grok-4.5

2026-07-12 04:34 UTC pith:WIA6Q7U5

load-bearing objection Clean reframing plus a real benchmark: learned world models collapse off-path and DSGE-generated coverage recovers them, with the large regime numbers correctly scoped as synthesis rather than pure OOD. the 4 major comments →

arxiv 2607.03144 v1 pith:WIA6Q7U5 submitted 2026-07-03 econ.GN q-fin.EC

DSGE as a Structured World Model:Benchmarking Counterfactual Generalization in Economic Worlds

Wenli Xu This is my paper
classification econ.GN q-fin.EC
keywords DSGEworld modelsbelief statecounterfactual generalizationpolicy regime shiftoff-path generalizationDSGE-Gymcoverage generation
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.

Modern world models learn transition maps only where trajectories were seen, so they degrade under policy-induced distribution shift and on counterfactual states. This paper argues that a Dynamic Stochastic General Equilibrium model is already a structured world model: its state variable is a belief state in the same sense used by latent world models, but supplied by theory with causal structure and hard cross-equation constraints that hold everywhere. The authors introduce DSGE-Gym, a benchmark of eight solved DSGE environments (from an 11-variable real-business-cycle model to the ECB’s 230-variable New Area-Wide Model) whose train and test splits deliberately sample different regions of the same state space. Learned architectures match the dynamics on the training path yet collapse off-path; the same architectures, trained at fixed sample size on data the DSGE itself generates across rare shocks and counterfactual policy regimes, recover much of that accuracy. Because no single historical sample can ever contain those counterfactual regimes, the result isolates structure’s power to manufacture the missing distribution rather than merely fit observed data.

Core claim

Learned world models match structured dynamics on-path (normalized RMSE roughly 0.004–0.08) but collapse off-path, with 5-sigma tail RMSE rising by up to about 40 times; training the identical architectures at fixed sample size on DSGE-generated rare and counterfactual-policy states roughly halves tail error and cuts policy-regime error by factors of 10–280 wherever the counterfactual rule shifts the ergodic support. The operative mechanism is coverage generation, not imposed constraints.

What carries the argument

The identity that a DSGE structural state is a belief state (Proposition 1): a sufficient statistic of history for future observations, supplied by construction together with the transition map and cross-equation restrictions that hold at every state. DSGE-Gym turns that identity into a measurement by drawing train and off-path test sets from the same solved model.

Load-bearing premise

That the solved DSGE is a trustworthy generator of the missing off-path distribution—i.e., that synthetic coverage remains valid precisely where real data cannot check it.

What would settle it

Train the same architectures on non-structural coverage of equal size and support (for example regime-dummy VARs or parameter-perturbation augmentation) and on a leave-one-regime-out wide set that withholds the exact test-regime parameters; if the large regime recovery disappears, the claim that equilibrium structure itself manufactures the missing distribution fails.

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

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 / 6 minor

Summary. The paper reframes DSGE models as structured world models whose recursive state is a belief state (Proposition 1), and introduces DSGE-Gym: eight DSGE environments with unified interfaces and off-path counterfactual test sets (tail shocks, policy-regime shifts), scaling to the ECB’s 230-variable NAWM. On a controlled one-step prediction task (T1), standard learned architectures (linear, MLP, LSTM, Transformer, NextLat) match the structured dynamics on-path but collapse under 5σ tails; training the same architectures at fixed sample size on DSGE-generated wide coverage roughly halves tail error and, where the counterfactual rule shifts the ergodic support, cuts regime error by large factors. Negative results show that a first-order oracle and static equilibrium penalties do not close the gap. Code and the benchmark are released.

Significance. If the scoped claims hold, the paper supplies a reusable, reproducible testbed for counterfactual generalization of world models in economics and a clean empirical demonstration that structure’s practical value is coverage generation of states no single history contains. Strengths include: a controlled protocol (same architectures, fixed N, z-scored RMSE, multiple seeds); replication across real, monetary, fiscal, open-economy, labor, climate, and firm-entry actions; an honest capacity interaction at 230 variables; useful negative results on approximate imposed structure; and full code/data release with datasheet-style documentation. The belief-state bridge (Proposition 1) is elementary but useful as identification. The work is a solid Datasets & Benchmarks / ML-for-economics contribution and a natural foundation for the announced Paper 2 on learned structured latents.

major comments (4)
  1. [Abstract, §5.3, Appendix B] Abstract and §5.3 / Appendix B: the headline regime multipliers (10–280×, and up to ~670× in Table 6) measure synthesis of a distribution that, by design, includes the exact test_regime policy parameters (Appendix B transparency note). The body scopes this correctly as generativity of structure rather than learner OOD extrapolation (§5.3, Remark 2), but the abstract and introduction state the numbers without that design fact. Because these figures will be the most-cited claim, the abstract must state that wide training spans the evaluated regime parameters and that H2 measures manufactured coverage, not withheld-regime generalization. A leave-one-regime-out number (already promised in the release) should appear in the main text if the large multipliers are retained.
  2. [§5.3, Limitations §6.2] §5.3 and Limitations §6.2: the central thesis is that structure helps off-path by manufacturing coverage. The design shows DSGE-generated coverage helps, but does not yet separate equilibrium structure from coverage per se. The paper correctly flags a non-structural control (regime-dummy VAR / parameter-perturbation augmentation) as the key missing piece and defers it to Paper 2. For this manuscript’s claim that structure (not merely a wider training region) is doing the work, either (i) include at least one matched non-structural coverage baseline on the main environments, or (ii) demote language that attributes the gain specifically to equilibrium structure until that control exists. As written, the load-bearing identification is incomplete.
  3. [Abstract, §5.4, Tables 3–6] §5.4 and Tables 3–6 vs. Abstract: regime recovery is large only where the counterfactual shifts the ergodic support (RBC, DMP, E-NK, Firm); in NK/TANK/TCM the counterfactual contracts support and narrow-trained regime error is already small, so gains are small or reverse for expressive models. The abstract’s unqualified “cuts policy-regime error 10–280×” averages over this heterogeneity. The support-shift vs. support-contract distinction should be in the abstract and treated as a primary result, not only in §5.4/§5.6.
  4. [§4 Tasks, §5] §4 Tasks and §5 Setup: the paper studies only T1 (one-step map fidelity with realized ε_{t+1} supplied). World-model value for planning lives in multi-step rollouts and action selection (T2/T3), which are released but not analyzed. The collapse/recovery story may change under compounding error. At minimum, report a short T2 multi-step probe on one or two environments (e.g., RBC and NK) so the off-path claim is not solely a one-step regression result; otherwise, further soften “world model” language in the abstract to “one-step transition model.”
minor comments (6)
  1. [§3.3] Proposition 1 is correctly labeled elementary; consider moving the full proof sketch to an appendix and keeping only the identification statement in the main text to free space for the non-structural control or T2 results.
  2. [Figures 2–5] Figure 1 and Table 1 are clear; ensure log-scale axes in Figures 2–5 are labeled as such in the caption, not only in the text.
  3. [References] NextLat citation appears as arXiv:2511.0XXXX (placeholder). Replace with the final identifier before camera-ready.
  4. [§5.2, Appendix C] §5.2: the static-penalty ablation is important; report the exact penalty weight and collocation construction in the main text or a short appendix table so the negative result is reproducible without reading the code.
  5. [Abstract, §5.5] NAWM is first-order only (§5.5, Limitations §6.1). State this in the abstract’s “scaling to 230 variables” clause so readers do not infer a nonlinear production-scale test.
  6. [§4, Appendix C] Terminology note in Appendix C (policy action vs. shock as input) is helpful; a one-sentence version in §4 would reduce confusion for ML readers.

Circularity Check

1 steps flagged

No load-bearing circular derivation; mild design overlap on regime H2 is disclosed and scoped as generativity, not OOD prediction-by-construction.

specific steps
  1. fitted input called prediction [§5.3; Appendix B (train_wide / test_regime construction)]
    "Transparency on overlap: the four wide regimes include the parameter setting used to build test_regime, so the H2 regime result quantifies the value of a structural generator synthesizing the counterfactual distribution (which no single history contains; §5, Remark 2), not a learner extrapolating to a withheld regime."

    For the policy-regime split, the wide training distribution is constructed to contain the same counterfactual policy parameters later used as test_regime. Large regime RMSE reductions (10–280×, up to ~670×) therefore largely reflect training on support that already covers the evaluation regime—i.e., learning under manufactured in-support coverage—rather than predicting a regime never present in training. The paper discloses and scopes this as measuring generativity of structure, so the step is mild design circularity of claim framing, not a forced identity equating input parameters to output RMSE.

full rationale

This is primarily an empirical benchmark paper. Proposition 1 is an elementary restatement that a recursive equilibrium state is a sufficient statistic, explicitly labeled as identification rather than a novel derivation; it does not force the experimental numbers. H1 (off-path collapse under narrow training) is a genuine empirical comparison against the same DGP used as ground truth—standard for synthetic benchmarks, not circular. H2’s tail recovery (rare-shock coverage at fixed sample size) is likewise an empirical capacity/coverage result. The only soft point is the regime half of H2: Appendix B states that train_wide includes the exact policy-parameter setting used for test_regime, so the large regime multipliers measure what happens once a structural generator synthesizes a distribution that already contains the test support, not withheld-regime extrapolation. The paper is explicit about this (§5.3, Remark 2, Limitations) and scopes the claim as structure’s generativity. That design choice weakens the rhetorical force of “counterfactual generalization” but does not make RMSE recovery a mathematical identity: networks can still fail to absorb wide coverage (as seen for some expressive models on contracting-support regimes and at NAWM scale). There is no self-citation chain, no uniqueness theorem imported from the author, and no fitted scalar renamed as a prediction. Score 2 reflects one minor, disclosed design overlap rather than circular derivation of the central claims.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

The central empirical claim rests on standard rational-expectations DSGE technology, the identification of the recursive state with a belief state, and the design choice that “wide” coverage may include the tested counterfactual parameters. No new physical entities are postulated; free parameters are ordinary calibration constants taken from the source models or modest architecture hyperparameters shared across environments.

free parameters (3)
  • Wide-training regime band and shock-scale segments
    The composition of train_wide (five 600-period 1–5σ segments + four 500-period counterfactual regimes) is a design choice that directly determines the measured H2 gains; the paper notes that test_regime parameters lie inside this span.
  • Baseline architecture hyperparameters (layers, latent dim, lr, epochs)
    Shared modest tuning (MLP 2×128, Transformer 2 layers/4 heads/64-d, etc.) is fixed for protocol uniformity; results could shift under stronger off-path-robustness interventions the paper does not exhaust.
  • Environment-specific calibrations (β, α, ϕπ, b, μ̄, fE, …)
    Taken from published sources (Nispi Landi, Coenen et al., standard DMP/BGM values) and verified at steady state; they define the ground-truth maps against which RMSE is scored.
axioms (4)
  • standard math The recursive (Bellman) state of a rational-expectations equilibrium is a belief state / sufficient statistic of history (Proposition 1).
    Elementary Markov property restated; used to equate DSGE state variables with the object NextLat/Belief-State Transformers learn.
  • domain assumption Deep parameters θ are invariant to policy-rule parameters ψ (Lucas critique / Proposition 2 factorization).
    Standard structural assumption that underwrites re-composition under counterfactual policy; Remark 2 correctly limits its scope for jump variables.
  • domain assumption A pruned third-order (or first-order for NAWM) perturbation solution is an adequate ground-truth generator for the environments studied.
    All train/test data and oracles are produced from these solutions; misspecification of the DSGE class would invalidate synthetic coverage.
  • ad hoc to paper Normalized one-step RMSE on z-scored variables is a sufficient probe of world-model quality for the claims made (T1).
    T2 multi-step and T3 planning are released but not analyzed; the paper treats one-step map fidelity as the necessary first probe.
invented entities (1)
  • DSGE-Gym benchmark suite no independent evidence
    purpose: Unified world-model interface and off-path counterfactual test sets for eight (plus unbenchmarked) DSGE environments.
    New artifact constructed for the paper; independent evidence will come from community use after release.

pith-pipeline@v1.1.0-grok45 · 26999 in / 3510 out tokens · 35562 ms · 2026-07-12T04:34:18.141283+00:00 · methodology

0 comments
read the original abstract

Modern world models -- Dreamer, transformer world models (IRIS, Genie), and JEPA / next-latent architectures -- learn dynamics from observed trajectories but share a weakness: their transition map is disciplined only where data were seen, so it degrades under policy-induced distribution shift and on counterfactual states off the training path. We argue that a Dynamic Stochastic General Equilibrium (DSGE) model is a structured world model: its state is a belief state -- the very object a latent world model learns, but supplied with causal structure and hard cross-equation constraints. We introduce DSGE-Gym, a benchmark of eight DSGE environments with off-path counterfactual test sets, scaling to the ECB's 230-variable New Area-Wide Model. We find that (i)learned world models match the dynamics on-path but collapse off-path (5{\sigma} tail RMSE up to \sim 40 the on-path level), and (ii)training the same architectures on data the DSGE generates across rare and counterfactual-policy states -- coverage only a structural model can synthesize -- roughly halves tail error and cuts policy-regime error 10--280 where the counterfactual rule shifts the ergodic support. Because such coverage cannot be sampled from any single history, this measures structure's ability to manufacture the missing distribution. DSGE-Gym and all code are released as a reproducible testbed for counterfactual generalization.

Figures

Figures reproduced from arXiv: 2607.03144 by Wenli Xu.

Figure 1
Figure 1. Figure 1: DSGE as a structured world model. (a) A learned world model and a DSGE describe the same object—an encoder mapping observations to a belief state zt (the latent a world model is trained to learn) and a transition Γ that rolls it forward under a policy action at. They differ in origin: the learned map is fit from sampled trajectories, whereas the DSGE map Γ and its cross￾equation constraints Φ are deduced f… view at source ↗
Figure 2
Figure 2. Figure 2: DSGE-generated coverage recovers counterfactual generalization. The same architectures [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coverage replication on NK (top) and TANK (bottom). Same architectures trained on [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Coverage replication on the two-country open economy (TCM; 45 variables, 6 shocks). [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Scaling to the ECB New Area-Wide Model (NAWM; 230 variables, 21 shocks). The [PITH_FULL_IMAGE:figures/full_fig_p013_5.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

60 extracted references · 5 linked inside Pith

  1. [2]

    International Conference on Machine Learning (ICML) , year=

    Learning Latent Dynamics for Planning from Pixels , author=. International Conference on Machine Learning (ICML) , year=

  2. [3]

    International Conference on Learning Representations (ICLR) , year=

    Dream to Control: Learning Behaviors by Latent Imagination , author=. International Conference on Learning Representations (ICLR) , year=

  3. [4]

    International Conference on Learning Representations (ICLR) , year=

    Mastering Atari with Discrete World Models , author=. International Conference on Learning Representations (ICLR) , year=

  4. [6]

    International Conference on Learning Representations (ICLR) , year=

    Transformers are Sample-Efficient World Models , author=. International Conference on Learning Representations (ICLR) , year=

  5. [7]

    International Conference on Machine Learning (ICML) , year=

    Genie: Generative Interactive Environments , author=. International Conference on Machine Learning (ICML) , year=

  6. [8]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Diffusion for World Modeling: Visual Details Matter in Atari , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  7. [9]

    Open Review , year=

    A Path Towards Autonomous Machine Intelligence , author=. Open Review , year=

  8. [10]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , author=. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

  9. [12]

    arXiv preprint arXiv:2511.0XXXX , year=

    Next-Latent Prediction Transformers Learn Compact World Models , author=. arXiv preprint arXiv:2511.0XXXX , year=

  10. [13]

    International Conference on Machine Learning (ICML) , year=

    Understanding Self-Predictive Learning for Reinforcement Learning , author=. International Conference on Machine Learning (ICML) , year=

  11. [14]

    International Conference on Learning Representations (ICLR) , year=

    Bridging State and History Representations: Understanding Self-Predictive RL , author=. International Conference on Learning Representations (ICLR) , year=

  12. [15]

    International Conference on Learning Representations (ICLR) , year=

    Belief State Transformers , author=. International Conference on Learning Representations (ICLR) , year=

  13. [16]

    Artificial Intelligence , volume=

    Planning and Acting in Partially Observable Stochastic Domains , author=. Artificial Intelligence , volume=

  14. [17]

    Journal of Mathematical Analysis and Applications , volume=

    Sufficient Statistics in the Optimum Control of Stochastic Systems , author=. Journal of Mathematical Analysis and Applications , volume=

  15. [19]

    International Economic Review , volume=

    Deep Equilibrium Nets , author=. International Economic Review , volume=

  16. [20]

    Journal of Monetary Economics , volume=

    Deep Learning for Solving Dynamic Economic Models , author=. Journal of Monetary Economics , volume=

  17. [21]

    Econometrica , volume=

    Financial Frictions and the Wealth Distribution , author=. Econometrica , volume=

  18. [24]

    Econometrica , volume=

    Time to Build and Aggregate Fluctuations , author=. Econometrica , volume=

  19. [25]

    Carnegie-Rochester Conference Series on Public Policy , volume=

    Econometric Policy Evaluation: A Critique , author=. Carnegie-Rochester Conference Series on Public Policy , volume=

  20. [26]

    American Economic Review , volume=

    Shocks and Frictions in US Business Cycles: A Bayesian DSGE Approach , author=. American Economic Review , volume=

  21. [27]

    American Economic Review , volume=

    Monetary Policy According to HANK , author=. American Economic Review , volume=

  22. [28]

    2023 , note=

    MacroModelling.jl: A Julia Package for Developing and Solving Dynamic Stochastic General Equilibrium Models , author=. 2023 , note=

    2023

  23. [29]

    2018 , note=

    The Real Business Cycle Model , author=. 2018 , note=

    2018

  24. [30]

    2021 , note=

    The New Keynesian Model and the TANK Model , author=. 2021 , note=

    2021

  25. [31]

    Journal of the European Economic Association , volume=

    Understanding the Effects of Government Spending on Consumption , author=. Journal of the European Economic Association , volume=

  26. [32]

    Journal of Economic Theory , volume=

    Limited Asset Markets Participation, Monetary Policy and (Inverted) Aggregate Demand Logic , author=. Journal of Economic Theory , volume=

  27. [33]

    ECB Working Paper / International Finance , year=

    The New Area-Wide Model of the Euro Area: A Micro-Founded Open-Economy Model for Forecasting and Policy Analysis , author=. ECB Working Paper / International Finance , year=

  28. [34]

    Diffusion for world modeling: Visual details matter in atari

    Eloi Alonso, Adam Jelley, Vincent Micheli, Anssi Kanervisto, Amos Storkey, Tim Pearce, and Fran c ois Fleuret. Diffusion for world modeling: Visual details matter in atari. In Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  29. [35]

    Self-supervised learning from images with a joint-embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  30. [36]

    Deep equilibrium nets

    Marlon Azinovic, Luca Gaegauf, and Simon Scheidegger. Deep equilibrium nets. International Economic Review, 63 0 (4): 0 1471--1525, 2022

    2022

  31. [37]

    Economics-inspired neural networks with stabilizing homotopies

    Marlon Azinovic, Harold Cole, and Felix Kubler. Economics-inspired neural networks with stabilizing homotopies. arXiv preprint arXiv:2303.14802, 2023

    Pith/arXiv arXiv 2023

  32. [38]

    Revisiting feature prediction for learning visual representations from video

    Adrien Bardes, Quentin Garrido, Jean Ponce, Xinlei Chen, Michael Rabbat, Yann LeCun, Mahmoud Assran, and Nicolas Ballas. Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471, 2024

    Pith/arXiv arXiv 2024

  33. [39]

    Florin O. Bilbiie. Limited asset markets participation, monetary policy and (inverted) aggregate demand logic. Journal of Economic Theory, 140 0 (1): 0 162--196, 2008

    2008

  34. [40]

    Genie: Generative interactive environments

    Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, et al. Genie: Generative interactive environments. In International Conference on Machine Learning (ICML), 2024

    2024

  35. [41]

    The new area-wide model of the euro area: A micro-founded open-economy model for forecasting and policy analysis

    G \"u nter Coenen, Roland Straub, and Mathias Trabandt. The new area-wide model of the euro area: A micro-founded open-economy model for forecasting and policy analysis. ECB Working Paper / International Finance, 2008. New Area-Wide Model (NAWM), Euro Area--US

    2008

  36. [42]

    Financial frictions and the wealth distribution

    Jes \'u s Fern \'a ndez-Villaverde, Samuel Hurtado, and Galo Nu \ n o. Financial frictions and the wealth distribution. Econometrica, 91 0 (3): 0 869--901, 2023

    2023

  37. [43]

    David L \'o pez-Salido, and Javier Vall \'e s

    Jordi Gal \' , J. David L \'o pez-Salido, and Javier Vall \'e s. Understanding the effects of government spending on consumption. Journal of the European Economic Association, 5 0 (1): 0 227--270, 2007

    2007

  38. [44]

    World models

    David Ha and J \"u rgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018

    Pith/arXiv arXiv 2018

  39. [45]

    Learning latent dynamics for planning from pixels

    Danijar Hafner, Timothy Lillicrap, Ian Fischer, Ruben Villegas, David Ha, Honglak Lee, and James Davidson. Learning latent dynamics for planning from pixels. In International Conference on Machine Learning (ICML), 2019

    2019

  40. [46]

    Dream to control: Learning behaviors by latent imagination

    Danijar Hafner, Timothy Lillicrap, Jimmy Ba, and Mohammad Norouzi. Dream to control: Learning behaviors by latent imagination. In International Conference on Learning Representations (ICLR), 2020

    2020

  41. [47]

    Mastering atari with discrete world models

    Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. In International Conference on Learning Representations (ICLR), 2021

    2021

  42. [48]

    Mastering diverse domains through world models

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse domains through world models. arXiv preprint arXiv:2301.04104, 2023

    Pith/arXiv arXiv 2023

  43. [49]

    Hu et al

    Edward S. Hu et al. Belief state transformers. In International Conference on Learning Representations (ICLR), 2025

    2025

  44. [50]

    Littman, and Anthony R

    Leslie Pack Kaelbling, Michael L. Littman, and Anthony R. Cassandra. Planning and acting in partially observable stochastic domains. Artificial Intelligence, 101 0 (1-2): 0 99--134, 1998

    1998

  45. [51]

    Violante

    Greg Kaplan, Benjamin Moll, and Giovanni L. Violante. Monetary policy according to hank. American Economic Review, 108 0 (3): 0 697--743, 2018

    2018

  46. [52]

    Kydland and Edward C

    Finn E. Kydland and Edward C. Prescott. Time to build and aggregate fluctuations. Econometrica, 50 0 (6): 0 1345--1370, 1982

    1982

  47. [53]

    A path towards autonomous machine intelligence

    Yann LeCun. A path towards autonomous machine intelligence. Open Review, 2022. Version 0.9.2

    2022

  48. [54]

    Robert E. Lucas. Econometric policy evaluation: A critique. Carnegie-Rochester Conference Series on Public Policy, 1: 0 19--46, 1976

    1976

  49. [55]

    Deep learning for solving dynamic economic models

    Lilia Maliar, Serguei Maliar, and Pablo Winant. Deep learning for solving dynamic economic models. Journal of Monetary Economics, 122: 0 76--101, 2021

    2021

  50. [56]

    Transformers are sample-efficient world models

    Vincent Micheli, Eloi Alonso, and Fran c ois Fleuret. Transformers are sample-efficient world models. In International Conference on Learning Representations (ICLR), 2023

    2023

  51. [57]

    Macromodelling.jl: A julia package for developing and solving dynamic stochastic general equilibrium models, 2023

    Thore M \"u ller. Macromodelling.jl: A julia package for developing and solving dynamic stochastic general equilibrium models, 2023. Julia package

    2023

  52. [58]

    Bridging state and history representations: Understanding self-predictive rl

    Tianwei Ni, Benjamin Eysenbach, Erfan Seyedsalehi, et al. Bridging state and history representations: Understanding self-predictive rl. International Conference on Learning Representations (ICLR), 2024

    2024

  53. [59]

    The real business cycle model, 2018

    Valerio Nispi Landi. The real business cycle model, 2018. Lecture notes, Bank of Italy

    2018

  54. [60]

    The new keynesian model and the tank model, 2021

    Valerio Nispi Landi. The new keynesian model and the tank model, 2021. Lecture notes, Bank of Italy

    2021

  55. [61]

    A survey of reinforcement learning for economics

    Pranjal Rawat. A survey of reinforcement learning for economics. arXiv preprint arXiv:2603.08956, 2026

    arXiv 2026

  56. [62]

    Equilibrium world models

    Andreas Schaab and Simon Scheidegger. Equilibrium world models. arXiv preprint arXiv:2606.23463, 2026

    Pith/arXiv arXiv 2026

  57. [63]

    Shocks and frictions in us business cycles: A bayesian dsge approach

    Frank Smets and Rafael Wouters. Shocks and frictions in us business cycles: A bayesian dsge approach. American Economic Review, 97 0 (3): 0 586--606, 2007

    2007

  58. [64]

    Sufficient statistics in the optimum control of stochastic systems

    Charlotte Striebel. Sufficient statistics in the optimum control of stochastic systems. Journal of Mathematical Analysis and Applications, 12 0 (3): 0 576--592, 1965

    1965

  59. [65]

    Understanding self-predictive learning for reinforcement learning

    Yunhao Tang, Zhaohan Daniel Guo, Pierre Harvey Richemond, et al. Understanding self-predictive learning for reinforcement learning. International Conference on Machine Learning (ICML), 2023

    2023

  60. [66]

    Hu, Tim Pearce, Pratyusha Sharma, Akshay Krishnamurthy, Riashat Islam, Alex Lamb, and John Langford

    Jayden Teoh, Manan Tomar, Kwangjun Ahn, Edward S. Hu, Tim Pearce, Pratyusha Sharma, Akshay Krishnamurthy, Riashat Islam, Alex Lamb, and John Langford. Next-latent prediction transformers learn compact world models. arXiv preprint arXiv:2511.0XXXX, 2025. Microsoft Research

    2025