Pith. sign in

REVIEW 2 major objections 4 minor 38 references

This paper proves that predictive accuracy and calibration alone cannot identify whether a forecast's spread comes from hidden state or genuine randomness, and shows that paired interventional tests—ClosurePairs—can recover that split.

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 ↗

ClosurePairs uses paired microstate and disturbance interventions with variance decomposition to identify whether future branching is caused by state aliasing or process noise, which ordinary prediction scores cannot.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Sound non-identifiability proof plus a genuinely useful interventional protocol; the fiber-sampling assumption is real but disclosed. the 2 major comments →

arxiv 2608.00591 v1 pith:ZJQLHR4D submitted 2026-08-01 cs.AI

Why Does the Future Branch? Identifiable Closure Tests for Stochastic Physical World Models

Yibin Dong This is my paper
classification cs.AI
keywords world modelsstochastic simulationidentifiabilitystate aliasingprocess noisevariance decompositioninterventional evaluationproper scoring rules
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.

The reading

A stochastic world model's predicted future distribution does not explain its own width. The paper proves that the same observed conditional future distribution can be produced by systems in which the forecast spread comes mostly from observations aliasing different physical states, or mostly from randomness that remains after the full state is fixed. No estimator that sees only passive transition data can tell these apart, even if it achieves perfect likelihood and calibration. The paper then introduces ClosurePairs, an interventional protocol that samples multiple compatible microstates within an observation fiber and crosses them with repeated exogenous disturbances, yielding unbiased estimates of state aliasing, process noise, and their interaction. If correct, evaluators of stochastic simulators must go beyond proper scoring rules to decide whether to refine sensing or to branch multiple futures.

Core claim

The paper's central claim is an observational non-identifiability result: for any observed conditional future distribution p(Y|Z,A) with variance V, there is a continuum of controlled systems sharing that exact kernel but differing in how V splits into Valias, the variance of microstate-conditional means, and Vproc, the average residual variance at fixed microstate. Since the joint law of (Z,A,Y) is identical across the family, no estimator based only on such samples can consistently recover both components. The paper then constructs ClosurePairs: hold the observation fixed, sample compatible microstates X from q(X|z,a), repeat the same exogenous disturbances E across them, and estimate the

What carries the argument

The law of total variance identity Var(Y|z,a)=Valias+Vproc is the load-bearing object: it splits forecast uncertainty into variation of the microstate-conditional mean (state aliasing) and residual variation at fixed microstate (process stochasticity). ClosurePairs operationalizes it by paired interventions—independent nested repeats for unreplayable noise and balanced crossed functional-ANOVA mean squares for replayable noise. The crossed estimator is the key mechanism: it uses the expected mean squares of a balanced two-way random-effects layout to separate state, noise, and state–noise interaction, so that interaction variance is not misread as additional aliasing.

Load-bearing premise

The protocol requires interventional access to the simulator: it must be possible to sample microstates from the true conditional distribution given a fixed observation and to replay or independently draw the exogenous disturbances; if that access is missing or the fiber sampler does not match the simulator's actual conditional distribution, the estimates answer a different question.

What would settle it

Construct two Gaussian simulators with identical p(Y|Z,A) but different alias fractions, as in Equation (4), train any observational estimator with large samples and many seeds, and check whether it can recover the alias fraction better than chance: any success would refute the paper's Proposition 1. ClosurePairs should recover the alias fraction; an observational estimator that also does so would contradict the proof.

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

If this is right

  • Proper forecast scores such as negative log-likelihood and calibration are provably insensitive to why a future distribution branches; equal likelihood does not imply equal mechanism.
  • ClosurePairs supplies unbiased estimates of state aliasing, process noise, and their nonlinear interaction, so mechanism attribution is achievable when paired interventions are available.
  • The optimal refine-versus-branch decision is not identified from passive data, but plugging consistent ClosurePairs estimates into the decision rule makes it decision-consistent when the regret margin is nonzero.
  • Nested observation resolutions yield a monotone decrease in the aliasing variance component with a resolution-invariant process floor, giving a closure diagnostic for coarse-grained physical models.
  • Under matched total variance, a router using ClosurePairs' components separates REFINE from BRANCH nearly perfectly (99.99%) while a magnitude-only threshold achieves only 66.48%.

Where Pith is reading between the lines

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

  • If the non-identifiability result carries over to any passive evaluation dataset, then benchmark rankings built purely on predictive likelihood cannot separate world models that differ in mechanism; benchmark suites should include paired-intervention probes or resolution manipulations.
  • The monotone reducibility theorem suggests a practical diagnostic for a learned world model's internal state: if expanding the model's state does not shrink the estimated aliasing component, the added state variables are not capturing the information hidden by the observation fiber.
  • The protocol's requirements—microstate fiber sampling and seed control or independent trials—also point toward physical experiments where initial microstates can be prepared and disturbances repeated, not just software simulators.
  • A testable extension is to use ClosurePairs-style attribution to guide which sensor to add: rather than a binary refine/branch choice, the estimated aliasing variance per hidden coordinate could rank candidate observations by expected variance reduction.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 4 minor

Summary. The paper addresses the problem of attributing predictive variance in stochastic world models to state aliasing (different physical states mapping to the same observation) versus process noise (residual randomness at fixed full state). It proves that this attribution is not identifiable from i.i.d. observational transition data (Proposition 1), then introduces ClosurePairs, an interventional protocol that samples compatible microstates within an observation fiber and crosses them with repeated exogenous disturbances. Under this protocol, nested and crossed estimators are proven unbiased (Theorems 1–2), and monotone reducibility under nested observation resolutions is established (Theorem 3). The paper reports extensive experiments: likelihood-equivalent Gaussian systems, matched-total-variance REFINE/BRANCH routing, 18 nonlinear Langevin conditions with OOD actions, a pixel-conditioned recurrent model, and a stochastic Pendulum closure curve. All experiments show large improvements in attribution MAE and decision regret at fixed NLL when paired labels are used, with negative controls (full-state zero aliasing) and bootstrap/Wilcoxon statistics.

Significance. If the claims hold, the paper makes a significant contribution: it gives a sharp negative result that proper scoring rules cannot identify the source of branching, and a concrete interventional protocol that can identify it under explicit access assumptions. The proofs are elementary but correct, and the experimental design is careful: frozen predictive branches isolate information content, negative controls confirm zero aliasing at full state, OOD actions test generalization, and seed-level bootstrap CIs and paired Wilcoxon tests support the comparisons. The paper also ships reproducible code with unit tests and is candid about limitations, including the need for microstate-fiber and seed-controlled access. The main caveat is that the protocol's estimand is defined relative to a chosen fiber distribution q(X|z,a); the paper states but does not empirically validate this assumption, which limits the strength of the central claim.

major comments (2)
  1. [§6.5, §8, Appendix D] The central identifiability claim (Eqs. 1–2 and Theorem 2) is conditional on the ability to sample compatible microstates from the true conditional distribution q(X|z,a). The paper states this in §8, but the pixel Pendulum study (§6.5) never specifies or validates the distribution over hidden velocities used to generate 'compatible microstates.' Proposition 1 implies that no observational check can certify the sampler, and Appendix D's fiber-validity diagnostic only asks the evaluator to report the sampler. Consequently, the reported alias-fraction MAEs (ID 0.130, OOD 0.170) are relative to an unverified choice of q; a different valid fiber changes the decomposition while preserving p(Y|Z,A), so the headline improvements do not establish an intrinsic property of the environment. Please add a validation experiment with known q, a sensitivity analysis over plausible fiber distributions, or
  2. [§1, §5, §8] The manuscript oscillates between 'measures why futures branch' and 'relative to the declared state.' Because the estimand depends on q(X|z,a) and on the declared full-state boundary, the process component is not a physical invariant. The abstract and introduction should consistently state that ClosurePairs identifies the decomposition for a specified fiber and disturbance interface; otherwise readers may over-interpret the process term as fundamental aleatoric noise. This framing affects the interpretation of every experimental claim, not merely wording.
minor comments (4)
  1. [Eq. (8)] The estimator formulas as rendered appear to have missing division bars; please ensure the final version clearly shows (MS_X - MS_XE)/K and (MS_E - MS_XE)/M.
  2. [§6.5, Table 4] The deep ensemble and MC dropout are different predictive models with different NLL. Please label them as diagnostic observational baselines rather than matched-NLL competitors; the frozen observational row is the appropriate matched control.
  3. [§6.3 and abstract] Clarify that the 99.99% routing accuracy is achieved with paired rollouts at evaluation time (as already stated in §8), and do not present it as a passive online result.
  4. [§4.1] The independent-repeat nested estimator is not evaluated empirically; add a small experiment or state explicitly that it is validated only by Theorem 1.

Circularity Check

0 steps flagged

No significant circularity: non-identifiability is an explicit construction and ClosurePairs estimators are direct plug-in estimators with independent experimental validation.

full rationale

The paper's central negative claim is not circular: Proposition 1 explicitly constructs a continuum of systems Y = mu(Z,A) + sqrt(lambda V) U + sqrt((1-lambda)V) E that all share the same p(Y|Z,A) while having different (Valias, Vproc). This is a standard indistinguishability argument, not an imported uniqueness theorem or a self-citation. The positive estimators are derived from the law of total variance and balanced ANOVA: Theorem 1 is proven by expanding E[s_i^2] and Var(row means), and Theorem 2 follows from the classical EMS equations M SX = VXE + K VX, etc. These are direct plug-in estimators of the definitions in Eqs. 1-2, with explicit unbiasedness proofs. The attribution head is trained only on paired intervention labels, while the predictive branch is either frozen or parameter-disjoint; NLL equality is enforced by construction, and none of the reported improvements are obtained by fitting to the outcome being explained. Experimental 'predictions' are evaluated against held-out intervention groups generated by the same protocol, which is a supervised generalization check under the paper's stated generative assumptions, not a circular reduction. Section 8 and Appendix D transparently state the key assumption: 'Fiber sampling must match the target q(X|z,a); otherwise estimates answer a shifted question.' That is an access/validity limitation, not a circularity: no parameter is fitted to the quantity it is later claimed to predict, and no equation reduces to its own input. References to Sobol and Searle et al. are standard external mathematical machinery, and the STORI separation is explicitly disclaimed as not this paper's contribution. There are no load-bearing self-citations. The honest finding is therefore no significant circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim relies on standard statistical assumptions (ANOVA, exchangeability) and strong domain assumptions about simulator access, but introduces no new physical entities or postulated fields. The free parameters listed are experiment-specific settings, not parameters of the identifiability proof.

free parameters (3)
  • impulse amplitude for total-variance matching = tuned on training data, exact value not reported
    In the matched-variance routing experiment (Section 6.3), the noise mechanism's impulse amplitude is tuned on training data so that total predictive variance matches the state mechanism. This is a synthetic-test setting, not a parameter of the core identifiability method.
  • sensor cost = 0.05 (Langevin), 0.065 (pixel), 0.05 (Pendulum)
    Decision regret computations depend on chosen sensor cost values. These are experiment design choices, not part of the statistical identifiability claim.
  • attribution head parameters rho(z,a) = learned from paired labels
    The learned attribution head is a function fitted to interventional targets; it is a supervised model, not a hand-set constant. It is a practical component of the protocol, not a free parameter of the identifiability theorems.
axioms (5)
  • standard math Balanced two-way ANOVA assumptions hold: finite second moments, exchangeability, independent crossings of microstates and disturbances.
    Used in Theorem 2 and Appendix A.3; standard random-effects model assumptions from Searle et al. (1992).
  • domain assumption Exogenous disturbances E are independent of microstate X and action A under the intervention.
    Required for the law of total variance decomposition in Section 2 and for the nested/crossed estimators.
  • domain assumption The simulator allows resetting or sampling microstates within an observation fiber and replaying or independently sampling disturbances.
    Stated in Section 8 (Limitations): the protocol requires this interventional access; without it the method cannot be applied.
  • domain assumption The compatible-microstate distribution q(X|z,a) is known or can be sampled exactly.
    Stated in the Limitations and Protocol Assumptions appendix: fiber sampling must match the target distribution or estimates answer a shifted question.
  • domain assumption Observations are nested sigma-fields for the monotone reducibility theorem.
    Theorem 3 requires G0 ⊂ G1 ⊂ ... ⊂ sigma(X); the paper notes this in Section 4.3 and Appendix D.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Why Does the Future Branch? Identifiable Closure Tests for Stochastic Physical World Models." pith.science (2026). https://pith.science/paper/ZJQLHR4D

@misc{pith2026260800591,
  author       = {Pith},
  title        = {Pith review of: Why Does the Future Branch? Identifiable Closure Tests for Stochastic Physical World Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZJQLHR4D}},
  note         = {Machine review of arXiv:2608.00591}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Stochastic world models are usually evaluated by the accuracy and calibration of their predicted futures. These criteria leave a decision-relevant ambiguity: the same conditional future distribution can arise because an observation aliases different physical states, or because the dynamics remain random after the declared full state is fixed. We prove that this attribution is not identifiable from ordinary transition data, even with an optimal probabilistic predictor. We introduce ClosurePairs, an interventional evaluation protocol that crosses compatible microstates with repeated exogenous disturbances. A two-way variance decomposition identifies state aliasing, process noise, and their nonlinear interaction; an independent-repeat variant applies when disturbances cannot be reused. On likelihood-equivalent Gaussian systems, paired supervision reduces alias-fraction error 15.96-fold at identical test NLL. Across 18 nonlinear Langevin conditions, it reduces attribution MAE from 0.372 to 0.051 and sensing regret from 0.0138 to 0.0003 without changing NLL. On a pixel-conditioned recurrent model, a frozen shared-state probe reduces alias-fraction MAE against a deep ensemble from 0.584 to 0.130 in distribution and from 0.630 to 0.170 out of distribution over ten seeds. Finally, in a matched-total-variance REFINE/BRANCH test, a total-variance router reaches 66.48 percent plus or minus 1.06 percent accuracy, whereas ClosurePairs reaches 99.99 percent plus or minus 0.02 percent and improves selected NLL from -2.087 to -2.717 over five seeds. ClosurePairs therefore measures why futures branch, information that proper forecast scores cannot identify.

Figures

Figures reproduced from arXiv: 2608.00591 by Yibin Dong.

Figure 1
Figure 1. Figure 1: Forecast equivalence does not imply mechanism equivalence. Left: five neural attribu￾tions produce identical likelihood. Right: across observationally equivalent systems, the observational split stays at its arbitrary initialization, whereas ClosurePairs recovers the true split. Error bars are standard deviations over three seeds. estimator is therefore identical at parameter values with different targets.… view at source ↗
Figure 2
Figure 2. Figure 2: Nonlinear multi-step results. Orange and blue models have identical predictive parameters and NLL. Paired supervision recovers the mechanism, especially the zero-aliasing full-state negative controls, and reduces sensing regret under both ID and OOD actions. Each bar averages three seeds; q, v denotes full-state observation [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pixel recurrent model. Fraction and alias-variance MAE are measured against crossed intervention targets. The right panel reports real active-sensing regret. Bars are means and error bars are standard deviations over ten independent training seeds. CP denotes ClosurePairs. a three-member deep ensemble (Lakshminarayanan et al., 2017), and Monte Carlo dropout (Gal & Ghahramani, 2016). Training actions lie in… view at source ↗
Figure 4
Figure 4. Figure 4: Pendulum components at angle-only resolution. Hidden-velocity aliasing falls with horizon for the bounded outcome, while noise and nonlinear state–noise interaction grow. At full state (not plotted), aliasing is zero and process totals are 0.00085, 0.01395, 0.04451, 0.10202. counterfactuals, and a crossed estimator that remains valid under state–noise interaction. Recent work also cautions that a generic a… 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

38 extracted references · 29 canonical work pages · 2 internal anchors

  1. [2]

    Nature , volume=

    Mastering Diverse Control Tasks through World Models , author=. Nature , volume=. 2025 , doi=

    work page 2025

  2. [3]

    Advances in Neural Information Processing Systems , volume=

    What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? , author=. Advances in Neural Information Processing Systems , volume=

    work page

  3. [5]

    Journal of the American Statistical Association , volume=

    Strictly Proper Scoring Rules, Prediction, and Estimation , author=. Journal of the American Statistical Association , volume=

  4. [7]

    On the Benefits of Memory for Modeling Time-Dependent

    Ruiz, Ricardo Buitrago and Marwah, Tanya and Gu, Albert and Risteski, Andrej , booktitle=. On the Benefits of Memory for Modeling Time-Dependent

    work page

  5. [8]

    Advances in Neural Information Processing Systems , volume=

    Learning Stochastic Multiscale Models , author=. Advances in Neural Information Processing Systems , volume=

    work page

  6. [9]

    Barsainyan, Aryan Amit and Lim, Jing Yu and Liu, Dianbo , journal=

    work page

  7. [10]

    Tung, Hsiao-Yu and Ding, Mingyu and Chen, Zhenfang and Bear, Daniel and Gan, Chuang and Tenenbaum, Josh and Yamins, Dan and Fan, Judith and Smith, Kevin , booktitle=

    work page

  8. [13]

    Anonymous , howpublished=

    work page

  9. [14]

    Mixture Density Networks , author=

    work page

  10. [15]

    Monte Carlo Methods in Financial Engineering , author=

    work page

  11. [16]

    Mathematical Modelling and Computational Experiments , volume=

    Sensitivity Estimates for Nonlinear Mathematical Models , author=. Mathematical Modelling and Computational Experiments , volume=

    work page

  12. [17]

    Variance Components , author=

    work page

  13. [18]

    Progress of Theoretical Physics , volume=

    Transport, Collective Motion, and Brownian Motion , author=. Progress of Theoretical Physics , volume=

    work page

  14. [19]

    Physical Review , volume=

    Memory Effects in Irreversible Thermodynamics , author=. Physical Review , volume=

    work page

  15. [21]

    Advances in Neural Information Processing Systems , volume=

    Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles , author=. Advances in Neural Information Processing Systems , volume=

    work page

  16. [22]

    Proceedings of the 33rd International Conference on Machine Learning , pages=

    Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning , author=. Proceedings of the 33rd International Conference on Machine Learning , pages=

    work page

  17. [23]

    CaliBench : Are the stochastic dynamics of video world models physically calibrated? OpenReview submission rWE29rkvDz, 2026

    Anonymous. CaliBench : Are the stochastic dynamics of video world models physically calibrated? OpenReview submission rWE29rkvDz, 2026

    work page 2026

  18. [24]

    STORI : A benchmark and taxonomy for stochastic environments

    Aryan Amit Barsainyan, Jing Yu Lim, and Dianbo Liu. STORI : A benchmark and taxonomy for stochastic environments. arXiv preprint arXiv:2509.01793, 2025

    work page arXiv 2025

  19. [25]

    Rethinking aleatoric and epistemic uncertainty

    Freddie Bickford Smith, Jannik Kossen, Eleanor Trollope, Mark van der Wilk, Adam Foster, and Tom Rainforth. Rethinking aleatoric and epistemic uncertainty. arXiv preprint arXiv:2412.20892, 2024

    Pith/arXiv arXiv 2024

  20. [26]

    Christopher M. Bishop. Mixture density networks. Technical Report NCRG/94/004, Aston University, 1994

    1994

  21. [27]

    What-If World: A Causal Benchmark for General World Models in Embodied Scenarios

    Kunlin Cai, Rui Song, Jinghuai Zhang, Kaiyuan Zhang, Pranav Bodapati, Alicia Yu, Fnu Suya, Mohammad Rostami, Jiaqi Ma, and Yuan Tian. What-if world: A causal benchmark for general world models in embodied scenarios. arXiv preprint arXiv:2605.27589, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  22. [28]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning, pp.\ 1050--1059, 2016

    work page 2016

  23. [29]

    Monte Carlo Methods in Financial Engineering

    Paul Glasserman. Monte Carlo Methods in Financial Engineering. Springer, 2004

    work page 2004

  24. [30]

    Tilmann Gneiting and Adrian E. Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American Statistical Association, 102 0 (477): 0 359--378, 2007

    2007

  25. [31]

    World models

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

    Pith/arXiv arXiv 2018

  26. [32]

    Mastering diverse control tasks through world models

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse control tasks through world models. Nature, 640: 0 647--653, 2025. doi:10.1038/s41586-025-08744-2

    doi:10.1038/s41586-025-08744-2 2025

  27. [33]

    Ilersich and Prasanth Nair

    Andrew F. Ilersich and Prasanth Nair. Learning stochastic multiscale models. In Advances in Neural Information Processing Systems, volume 38, 2025

    work page 2025

  28. [34]

    What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems, volume 30, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? In Advances in Neural Information Processing Systems, volume 30, 2017

    work page 2017

  29. [35]

    Simple and scalable predictive uncertainty estimation using deep ensembles

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. In Advances in Neural Information Processing Systems, volume 30, 2017

    2017

  30. [36]

    Transport, collective motion, and brownian motion

    Hazime Mori. Transport, collective motion, and brownian motion. Progress of Theoretical Physics, 33 0 (3): 0 423--455, 1965

    work page 1965

  31. [37]

    On the benefits of memory for modeling time-dependent PDE s

    Ricardo Buitrago Ruiz, Tanya Marwah, Albert Gu, and Andrej Risteski. On the benefits of memory for modeling time-dependent PDE s. In International Conference on Learning Representations, 2025

    work page 2025

  32. [38]

    Benjamin Sanderse, Panos Stinis, Romit Maulik, and Shady E. Ahmed. Scientific machine learning for closure models in multiscale problems: A review. arXiv preprint arXiv:2403.02913, 2024

    Pith/arXiv arXiv 2024

  33. [39]

    Searle, George Casella, and Charles E

    Shayle R. Searle, George Casella, and Charles E. McCulloch. Variance Components. Wiley, 1992

    work page 1992

  34. [40]

    Ilya M. Sobol. Sensitivity estimates for nonlinear mathematical models. Mathematical Modelling and Computational Experiments, 1 0 (4): 0 407--414, 1993

    work page 1993

  35. [41]

    What Can Latent World Models Know? Physical Parameter Identifiability in Multimodal Predictive Representations

    Kaizhen Tan, Xin Xu, Siru Tao, Hanzhe Hong, Yang Feng, and Heqing Du. What can latent world models know? physical parameter identifiability in multimodal predictive representations. arXiv preprint arXiv:2607.27017, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  36. [42]

    Balis, Gianluca De Cola, Tristan Deleu, Manuel Goul ã o, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al

    Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U. Balis, Gianluca De Cola, Tristan Deleu, Manuel Goul ã o, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al. Gymnasium: A standard interface for reinforcement learning environments. arXiv preprint arXiv:2407.17032, 2024

    Pith/arXiv arXiv 2024

  37. [43]

    Physion++ : Evaluating physical scene understanding that requires online inference of different physical properties

    Hsiao-Yu Tung, Mingyu Ding, Zhenfang Chen, Daniel Bear, Chuang Gan, Josh Tenenbaum, Dan Yamins, Judith Fan, and Kevin Smith. Physion++ : Evaluating physical scene understanding that requires online inference of different physical properties. In Advances in Neural Information Processing Systems Datasets and Benchmarks Track, volume 36, 2023

    work page 2023

  38. [44]

    Memory effects in irreversible thermodynamics

    Robert Zwanzig. Memory effects in irreversible thermodynamics. Physical Review, 124 0 (4): 0 983--992, 1961

    work page 1961

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.