Pith. sign in

REVIEW 2 major objections 4 minor 51 references

When a Verified World Model Still Loses: Play-Adequacy vs Prediction-Accuracy in LLM-Synthesized Code World Models

T0 review · 2 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A Code World Model can pass a random-trajectory gate at 100% transition accuracy and still lose systematically at play, because the <1% error sits exactly where competent play goes.

desk verdict The verified-vs-correct gap is a real and important result, but the paper needs to clarify exactly what the gate scores before the quantitative law can be taken at face value. read the letter →

arxiv 2607.14169 v2 pith:UICXWBFK submitted 2026-07-15 cs.AI cs.LG

Javier Aguilar Martín This is my paper · ORCID
classification cs.AIcs.LG
keywords CodeWorldModelsLLMsynthesistransitionaccuracysamplinggaterarerulesMonteCarloTreeSearchimperfect-informationgamesbeliefinference
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

This paper argues that verifying an LLM-synthesized game world model by transition accuracy on random play-throughs is the wrong test for planning. It shows a model can score 100% on that gate, be 98% accurate on states the planner actually visits, and still be systematically outplayed, because the rare transition it gets wrong is the one that decides competent games. The paper quantifies the harm: danger = play cost × (1 − rarity)^N, with the gate-miss factor proven exact and the play cost empirically measured at 0.091 with separated 95% confidence intervals. It also shows that feeding more examples does not repair the gap: an omitted rule is statistically unidentifiable from a sample that misses it, and tested LLMs behave as rule translators rather than rule inferrers. The same mechanism is demonstrated for belief functions in imperfect-information games, where a hand-built witness passes every inference check yet loses every game.

What carries the argument

The central mechanism is a reach-distribution shift between the verification policy (uniform-random play, which concentrates on shallow, common histories) and the deployment policy (Monte Carlo tree search, which concentrates on deep, strategically significant histories). This shift creates a rare-but-pivotal region that a random gate almost never samples and a planner reliably reaches. The quantitative carrier is the danger law danger = play cost × (1 − rarity)^N, whose gate-miss factor is proven exact by Bernoulli sampling and whose play-cost factor is bounded above by the planner's query-hit mass, certified below by measurement, and exact on the Beacon witness. On the inference side, a co

What would settle it

Run the full synthesis pipeline with many more gate-passing seeds and measure play cost: if the incomplete CWM's win rate approaches the fair baseline (≈0.495) instead of staying near 0.404, the verified-vs-correct gap does not transfer to actual synthesis. Alternatively, find one LLM synthesis that, given on-manifold transitions containing a rare rule, later encodes the rule and plays at parity—that would refute 'translation, not inference.' On the inference side, a game with a rare-but-pivotal belief error where a pure random gate still catches the error would falsify the coverage bound's pr

Watch

Extended reading notes

Core claim

The central discovery is that a Code World Model synthesized by an LLM can satisfy the standard acceptance test—transition accuracy 1.0 on a random trajectory sample—and even be ≥98% state-accurate on the distribution the planner actually visits, while being wrong in a rare-but-pivotal region that competent play reliably reaches. The paper calls this the verified-vs-correct gap, and measures its play cost at 0.091 (from a 0.404 win rate against a 0.495 fair baseline). The harm follows a quantitative law, danger = play cost × (1 − rarity)^N, where the gate-miss factor (1 − rarity)^N is proven exact under i.i.d. sampling and play cost is empirically bounded. The paper further proves that an om

Load-bearing premise

The headline 0.091 play cost rests on a hand-written rule-blind version of the game standing in for the LLM-synthesized incomplete model, and the paper states this bridge is empirical rather than proven—if the two are not play-equivalent outside the rare region, the instrument would not isolate the omitted-rule mechanism.

Editorial extensions

If this is right

  • Adequacy for planning-oriented world models should be measured on the planner's search distribution or by direct play, not by transition accuracy on randomly sampled transitions.
  • Synthesis pipelines should include held-out play-throughs from a trusted reference planner; replacing one of 2000 random Beacon gates with one reference trajectory changed the verdict from accept to reject.
  • The specification must be complete before synthesis: example transitions do not reliably teach an omitted rule, and off-manifold examples actively corrupt the synthesized model.
  • The danger law gives a threshold test: a rule with random rarity r is safe only when (1 − r)^N is negligible; games with deep rare regions sit in the danger zone while shallow games are provably covered.
  • For imperfect-information games, belief functions must be gated separately from transitions, because the information partition is invisible to any transition accuracy gate.

Reading between the lines

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

  • A natural testable extension would apply the policy-guided mixture gate to other games beyond Beacon, predicting that a few reference-planner trajectories expose rare-region errors that pure random gates systematically miss.
  • The synthesis-residual channel—where the rule is present in the sample but the LLM still fails to encode it—means the danger law is a lower bound under an idealized synthesizer; real pipelines may suffer larger end-to-end harm than the law predicts, as the paper's synthesized replication suggests.
  • The reach-shift diagnosis is not specific to LLM-synthesized code: any learned world model validated on data from a different distribution than the deployment distribution could show the same verified-yet-inadequate pattern, including neural model-based RL and simulation-to-real pipelines.
  • The structural analogy to off-equilibrium-path beliefs points toward a design principle: verification should discipline the unconstrained region the way equilibrium refinements do in game theory, rather than relying purely on random sampling.
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.

Referee Report

2 major / 4 minor

Summary. The paper asks whether passing a random-trajectory transition-accuracy gate certifies that an LLM-synthesized Code World Model is adequate for planning. It argues no, and reports four main results: (1) a constructed army5x5a variant with a rare material-at-cap tiebreak where a rule-blind CWM passes a 100% transition-accuracy gate, is at least 98% state-accurate on the planner's search distribution, and yet loses at play with a measured play cost of 0.091 (seed-clustered CI [0.065,0.117]); (2) a quantitative danger law, danger = play_cost x (1 - rarity)^N, with the gate-miss factor proven exact; (3) a synthesis study showing that the omitted rule is translated when specified but not inferred from examples, across GPT-5.x models and several data regimes; and (4) an imperfect-information extension with a coverage bound for the inference gate, a hand-constructed Beacon witness that passes the inference gate yet loses every game, and two targeted checks that close this witnessed gap. The paper also includes an honest null result on small fully-specified games, explicit limitations, and a reproducible code release.

Significance. If the central claims hold, this is a significant and timely result for the CWM/LLM game-playing paradigm and for sampling-based verification of synthesized world models more broadly. The paper is unusually careful: headline play results are accompanied by pooled Wilson intervals and seed-clustered intervals; the Beacon play cost is proven exactly and checked mechanically; the authors report an honest null and explicitly separate instrumented from synthesized evidence; and they disclose and fix a feedback-channel confound. The danger law is a useful, falsifiable quantitative statement, and the distinction between verification-reach and deployment-reach distributions is conceptually important. The paper also ships reproducible scripts and records exact commands for the headline results.

major comments (2)
  1. [§2.2/Table 1; §3.3/Table 8; §5.3] The gate's scoring definition is internally inconsistent, and the exact danger law depends on which definition is meant. Section 2.2 and Table 1 define transition accuracy as agreement on (state, action, next-state) tuples and the gate criterion as transition accuracy = 1.0 on those tuples. The material-at-cap rule, however, changes only the terminal returns(state) vector (draw vs material-based win); it need not alter any (s,a,s') tuple or is_terminal. Under the stated tuple-only metric the gate could never detect a material-at-cap terminal, so the gate-miss probability would be 1, not (1-r)^N, and Proposition 1 would not apply. Yet Table 8 reports that 0/10 material-terminal-present seeds reach gate 1.0, and §5.3 shows refinement feedback printing returns: expected={1: 1.0, ...} got={1: 0.0, ...} on cap states with unequal material. This is only possible if the gate scores terminal ret
  2. [§3.3, Table 5] The abstract and Contribution (1) state that the incomplete CWM is '≥98% state-accurate' on the planner's own search distribution, but the paper never reports a measured value of agree(D_cwm). Table 5 reports only gap_truth = 0.000, and the prose says the divergence region is 'less than 1% of visited states' without counts, seeds, or a precise definition of which states count (played trajectory vs MCTS search tree). Because the 98% number is one of the two accuracy metrics claimed to be blind to the omission, it needs a direct measurement with a confidence interval, or the claim should be replaced by a clearly derived bound if that is what supports it. This is a missing support for a headline empirical number.
minor comments (4)
  1. [Table 9] The danger-curve table uses a rounded play cost of 0.12, which is outside the reported seed-clustered 95% CI [0.065,0.117] for the headline play cost. The threshold shape is insensitive, as stated, but the numerical danger values should use the central estimate 0.091 or be presented as illustrative only.
  2. [§6.7, Proposition 5] The proposition is marked 'self-proved (□)', but the symbol is not defined and the proof is a definitional argument. Consider attaching a short proof block or explicitly labeling it as a definitional observation rather than a theorem.
  3. [§4.2, Figure 4] The x-axes are reversed in the two panels of Figure 4. The reversal is explained in the caption, but a reader may initially misread the direction of the threshold; consider adding an arrow or labeling 'rarer ->' on both panels.
  4. [§3.3, Panel B] Footnote 1 is long and contains important caveats about seed 11. It would be easier to follow if the seed-11 exclusion and the one-directional coincidence were stated in the main text of Table 8's caption.

Circularity Check

1 steps flagged · score 4.0 of 10

Danger-law gate-miss factor is definitional: r is defined as the gate-observable rate, so (1-r)^N restates the input; central play-cost experiments remain independent.

  1. self definitional [Section 4.1 Proposition 1; Section 3.3 rarity definition]
    "A sampling gate draws N i.i.d. uniform-random play-throughs and accepts the CWM if none of them triggers the rule in question (and so reveals a discrepancy between the CWM and the true game). Since each play-through triggers the rule independently with probability r (the 'rarity'), the probability the gate never observes the rule is exactly P(miss) = (1−r)^N."

    The rarity r is defined in Section 3.3 as 'the rate at which the material-at-cap branch produces a decisive, gate-observable result.' Thus r is by definition the per-trial probability that the gate observes the rule. Proposition 1 then 'proves' that the probability the gate never observes the rule in N i.i.d. trials is (1−r)^N, which is simply the complementary Bernoulli probability under that definition — a restatement of the input rather than an independent derivation. The danger law danger = play_cost × (1−r)^N is consequently a definitional decomposition (expected harm = P(miss) × E[harm|miss]) whose threshold shape follows from the exponential function, not from empirical discovery. The paper's own language ('gate-observable result') makes the factor definitional. The independent empi

full rationale

The paper's central empirical findings are measured rather than fitted: the play-cost gap (Table 6, seed-clustered CI) comes from a controlled arena comparison, and the synthesized-pipeline replication (Panel B, Table 8) is a separate end-to-end experiment. The coverage bounds and Beacon play-cost-exactness are standard combinatorial arguments or exhaustive checks. There are no load-bearing self-citations: the only closely related CWM work cited (Lehrach et al.) is by other authors and is used for paradigm reproduction, not to justify the paper's own claims. The one genuinely definitional step is the danger-law's (1−r)^N factor: because r is defined as the probability of a 'gate-observable' rule-triggering event, the miss probability is the complement of that same quantity, so the 'proven exact' factor is a tautology of the definition. The paper stops short of claiming the law is parameter-free, and it explicitly scopes the law to one channel, so this is partial circularity rather than a wholesale reduction of the paper's thesis. Score 4 reflects one definitional 'prediction' in an otherwise empirically independent paper.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard probability and extensive-form game reasoning plus several explicit empirical/domain assumptions. No free parameters are fitted to manufacture the headline numbers; the only hand-chosen constant is the rounded play cost used for one illustrative table. The paper introduces constructed games (army5x5a+material-at-cap, Beacon) as instruments, not as new theoretical entities, and gives independent analytic support for Beacon's behavior via Proposition 4.

free parameters (1)
  • round play-cost constant used for danger-curve table = 0.12 (headline measurements span 0.091–0.131)
    Table 9 computes danger using a rounded constant play cost. It is not fitted to make the law match—the threshold shape is robust to the choice—but it is a hand-chosen summary value for display.
assumptions (5)
  • domain assumption Random-play rule firings are i.i.d. Bernoulli with probability r per play-through (Proposition 1).
    This makes (1−r)^N exact. It assumes the uniform-random policy is memoryless and that rule-firing events across independent play-throughs are independent, which is standard but not proven for all games.
  • domain assumption Only distinct fresh play-throughs count toward N; re-used refinement trajectories contribute nothing (Section 4, 'What N counts').
    The numerical instances of the danger law and Corollary 2 (N≈40, N=2000) depend on this accounting. The paper verifies it against its own implementation, but it is not a theorem about other pipelines.
  • domain assumption Detectability: when the inference gate visits an erring info-set, the elementwise comparison of consistent-state sets surfaces the error (Theorem 1).
    Stated explicitly as a hypothesis. It holds for the paper's own gate by construction, but would not hold for a gate that checks only coarse summaries.
  • domain assumption Competent MCTS reach of the cap region stays roughly constant as the rarity knob varies, making play cost approximately rarity-invariant (Remark 1).
    This is the load-bearing empirical regularity behind the cheap/expensive split and the threshold danger curve. The paper labels it a mechanistic correlate, not a proof, and it may fail for other games or planners.
  • domain assumption GPT-5.x mini/large results and a few cross-family probes are representative enough to support the scoped 'translation, not inference' claim.
    Section 8 explicitly limits the universal form to a conjecture; the empirical finding is scoped to the tested models, regimes, and seeds. This is an honest boundary rather than a hidden assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When a Verified World Model Still Loses: Play-Adequacy vs Prediction-Accuracy in LLM-Synthesized Code World Models." pith.science (2026). https://pith.science/paper/UICXWBFK

@misc{pith2026260714169,
  author       = {Pith},
  title        = {Pith review of: When a Verified World Model Still Loses: Play-Adequacy vs Prediction-Accuracy in LLM-Synthesized Code World Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UICXWBFK}},
  note         = {Machine review of arXiv:2607.14169}
}
abstract

Large language models can synthesize a game's rules as executable code - a Code World Model (CWM) - which a classical planner then searches over. Such models are typically accepted when they reach high transition accuracy on sampled trajectories. We argue this is the wrong notion of adequacy for planning. We show four things. (1) An LLM-synthesized CWM can pass a sampling gate at 100% transition accuracy and be $\geq 98\%$ state-accurate on the planner's own search distribution, yet lose systematically at play, because the $<1\%$ it gets wrong is exactly the pivotal dynamics; the play cost of the omitted rule is $0.091$ (seed-clustered 95% CI $[0.065,0.117]$, $n=4800$). We call this the verified-vs-correct gap, and confirm it end-to-end through the synthesis pipeline. (2) The harm follows a quantitative law, $\mathrm{danger}=\mathrm{play\_cost}\times(1-\mathrm{rarity})^N$, whose $(1-\mathrm{rarity})^N$ gate-miss factor is proven exact and whose play cost is empirically bounded. (3) The failure is not repaired by more data: LLM synthesis behaves as rule translation, not rule inference, and did not infer the omitted rule across models (GPT-5.x) and data regimes (including DAgger and targeted examples). (4) The same mechanism recurs on the belief-inference function of imperfect-information CWMs: we prove a coverage bound (a size-$N$ gate is identifying when $N\gtrsim b^{d_{\max}}$), explaining why shallow games such as Kuhn poker show no gap, and hand-construct Beacon, a verified-but-wrong inference function that passes the gate yet loses every game. These results suggest adequacy for planning-oriented world models should be measured on the search distribution or by play directly, not by prediction accuracy on sampled transitions.

Figures

Figures reproduced from arXiv: 2607.14169 by the authors.

Figure 1
Figure 1. The structural diagnosis underlying every gap in this paper, schematically. The verifica [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The headline result of Table [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. A measured mechanistic correlate of the cheap/expensive split [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The danger law danger = play cost (1−r) N is a threshold in rarity, and the same law recurs on both halves of the CWM contract. Solid curves are the analytic law; markers are the measured operating points ( [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith

  1. [1]

    Proceedings of the Fifth

    Claessen, Koen and Hughes, John , title =. Proceedings of the Fifth. 2000 , pages =

    2000

  2. [2]

    and Powley, Edward J

    Cowling, Peter I. and Powley, Edward J. and Whitehouse, Daniel , title =. 2012 , volume =

    2012

  3. [3]

    Artificial Intelligence , year =

    Frank, Ian and Basin, David , title =. Artificial Intelligence , year =

  4. [4]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , series =

    Gao, Luyu and Madaan, Aman and Zhou, Shuyan and Alon, Uri and Liu, Pengfei and Yang, Yiming and Callan, Jamie and Neubig, Graham , title =. Proceedings of the 40th International Conference on Machine Learning (ICML) , series =

  5. [5]

    2025 , eprint =

    Verma, Vivek and Huang, David and Chen, William and Klein, Dan and Tomlin, Nicholas , title =. 2025 , eprint =

    2025

  6. [6]

    International Conference on Learning Representations (ICLR) , year =

    Hafner, Danijar and Lillicrap, Timothy and Ba, Jimmy and Norouzi, Mohammad , title =. International Conference on Learning Representations (ICLR) , year =

  7. [7]

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

    Janner, Michael and Fu, Justin and Zhang, Marvin and Levine, Sergey , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  8. [8]

    and Wilson, Robert , title =

    Kreps, David M. and Wilson, Robert , title =. Econometrica , year =

Show all 51 references
  1. [9]

    and Amos, Brandon and Yadan, Omry and Calandra, Roberto , title =

    Lambert, Nathan O. and Amos, Brandon and Yadan, Omry and Calandra, Roberto , title =. Proceedings of the 2nd Conference on Learning for Dynamics and Control (L4DC) , series =. 2020 , note =

    2020

  2. [10]

    , title =

    Lehrach, Wolfgang and Hennes, Daniel and Lazaro-Gredilla, Miguel and Lou, Xinghua and Wendelken, Carter and Li, Zun and Dedieu, Antoine and Grau-Moya, Jordi and Lanctot, Marc and Iscen, Atil and Schultz, John and Chiam, Marcus and Gemp, Ian and Zielinski, Piotr and Singh, Sati...

    2025

  3. [11]

    Liang, Jacky and Huang, Wenlong and Xia, Fei and Xu, Peng and Hausman, Karol and Ichter, Brian and Florence, Pete and Zeng, Andy , title =

  4. [12]

    and Buro, Michael and Furtak, Timothy , title =

    Long, Jeffrey Richard and Sturtevant, Nathan R. and Buro, Michael and Furtak, Timothy , title =. Proceedings of the

  5. [13]

    Bandit Based

    Kocsis, Levente and Szepesv. Bandit Based. Proceedings of the 17th European Conference on Machine Learning (ECML) , pages =

  6. [14]

    and Kroese, Dirk P

    Rubinstein, Reuven Y. and Kroese, Dirk P. , title =

  7. [15]

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , booktitle =

    Ross, St. A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , booktitle =

  8. [16]

    Games and Economic Behavior , year =

    von Stengel, Bernhard , title =. Games and Economic Behavior , year =

  9. [17]

    and Cowling, Peter I

    Whitehouse, Daniel and Powley, Edward J. and Cowling, Peter I. , title =

  10. [18]

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

    Zinkevich, Martin and Johanson, Michael and Bowling, Michael and Piccione, Carmelo , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

  11. [19]

    Advances in Neural Information Processing Systems 22 (NeurIPS) , year =

    Lanctot, Marc and Waugh, Kevin and Zinkevich, Martin and Bowling, Michael , title =. Advances in Neural Information Processing Systems 22 (NeurIPS) , year =

  12. [21]

    Kuhn, H. W. , title =. Contributions to the Theory of Games, Volume I , editor =

  13. [22]

    Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence (UAI) , pages =

    Southey, Finnegan and Bowling, Michael and Larson, Bryce and Piccione, Carmelo and Burch, Neil and Billings, Darse and Rayner, Chris , title =. Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence (UAI) , pages =

  14. [23]

    2020 , howpublished =

    Erickson, Alek , title =. 2020 , howpublished =

    2020

  15. [24]

    , title =

    Wilson, Edwin B. , title =. Journal of the American Statistical Association , volume =

  16. [25]

    International Journal of Game Theory , volume =

    Selten, Reinhard , title =. International Journal of Game Theory , volume =

  17. [26]

    Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search , booktitle =

    Coulom, R. Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search , booktitle =

  18. [27]

    QuickCheck : A lightweight tool for random testing of Haskell programs

    Koen Claessen and John Hughes. QuickCheck : A lightweight tool for random testing of Haskell programs. In Proceedings of the Fifth ACM SIGPLAN International Conference on Functional Programming (ICFP) , pages 268--279, 2000

    2000

  19. [28]

    Efficient selectivity and backup operators in monte-carlo tree search

    R \'e mi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In Computers and Games (CG 2006), volume 4630 of Lecture Notes in Computer Science, pages 72--83. Springer, 2006

    2006

  20. [29]

    Cowling, Edward J

    Peter I. Cowling, Edward J. Powley, and Daniel Whitehouse. Information set Monte Carlo tree search. IEEE Transactions on Computational Intelligence and AI in Games , 4 0 (2): 0 120--143, 2012

    2012

  21. [30]

    Alek Erickson. Trike. Abstract board game. https://boardgamegeek.com/boardgame/307379/trike, 2020. Accessed 2026-07-06

    2020

  22. [31]

    Search in games with incomplete information: A case study using Bridge card play

    Ian Frank and David Basin. Search in games with incomplete information: A case study using Bridge card play. Artificial Intelligence, 100 0 (1--2): 0 87--123, 1998

    1998

  23. [32]

    PAL : Program-aided language models

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. PAL : Program-aided language models. In Proceedings of the 40th International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, pages 10764--...

    2023

  24. [33]

    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

  25. [34]

    When to trust your model: Model-based policy optimization

    Michael Janner, Justin Fu, Marvin Zhang, and Sergey Levine. When to trust your model: Model-based policy optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2019

    2019

  26. [35]

    Bandit based Monte - Carlo planning

    Levente Kocsis and Csaba Szepesv \'a ri. Bandit based Monte - Carlo planning. In Proceedings of the 17th European Conference on Machine Learning (ECML), pages 282--293, 2006

    2006

  27. [36]

    Kreps and Robert Wilson

    David M. Kreps and Robert Wilson. Sequential equilibria. Econometrica, 50 0 (4): 0 863--894, 1982

    1982

  28. [37]

    H. W. Kuhn. A simplified two-person poker. In H. W. Kuhn and A. W. Tucker, editors, Contributions to the Theory of Games, Volume I, number 24 in Annals of Mathematics Studies, pages 97--103. Princeton University Press, 1950

    1950

  29. [38]

    Lambert, Brandon Amos, Omry Yadan, and Roberto Calandra

    Nathan O. Lambert, Brandon Amos, Omry Yadan, and Roberto Calandra. Objective mismatch in model-based reinforcement learning. In Proceedings of the 2nd Conference on Learning for Dynamics and Control (L4DC), volume 120 of Proceedings of Machine Learning Research, pages 761--770...

    2020 arXiv

  30. [39]

    Monte C arlo sampling for regret minimization in extensive games

    Marc Lanctot, Kevin Waugh, Martin Zinkevich, and Michael Bowling. Monte C arlo sampling for regret minimization in extensive games. In Advances in Neural Information Processing Systems 22 (NeurIPS), pages 1078--1086, 2009

    2009

  31. [40]

    Wolfgang Lehrach, Daniel Hennes, Miguel Lazaro-Gredilla, Xinghua Lou, Carter Wendelken, Zun Li, Antoine Dedieu, Jordi Grau-Moya, Marc Lanctot, Atil Iscen, John Schultz, Marcus Chiam, Ian Gemp, Piotr Zielinski, Satinder Singh, and Kevin P. Murphy. Code World Models for general ...

    2025

  32. [41]

    Code as policies: Language model programs for embodied control

    Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In IEEE International Conference on Robotics and Automation (ICRA) , pages 9493--9500, 2023

    2023

  33. [42]

    Sturtevant, Michael Buro, and Timothy Furtak

    Jeffrey Richard Long, Nathan R. Sturtevant, Michael Buro, and Timothy Furtak. Understanding the success of perfect information Monte Carlo sampling in game tree search. In Proceedings of the AAAI Conference on Artificial Intelligence , pages 134--140, 2010

    2010

  34. [43]

    Gordon, and J

    St \'e phane Ross, Geoffrey J. Gordon, and J. Andrew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the 14th International Conference on Artificial Intelligence and Statistics (AISTATS), volume 15 of Proceed...

    2011

  35. [44]

    Rubinstein and Dirk P

    Reuven Y. Rubinstein and Dirk P. Kroese. Simulation and the Monte Carlo Method . Wiley, 3rd edition, 2017

    2017

  36. [45]

    Reexamination of the perfectness concept for equilibrium points in extensive games

    Reinhard Selten. Reexamination of the perfectness concept for equilibrium points in extensive games. International Journal of Game Theory, 4 0 (1): 0 25--55, 1975

    1975

  37. [46]

    Bayes' bluff: Opponent modelling in poker

    Finnegan Southey, Michael Bowling, Bryce Larson, Carmelo Piccione, Neil Burch, Darse Billings, and Chris Rayner. Bayes' bluff: Opponent modelling in poker. In Proceedings of the 21st Conference on Uncertainty in Artificial Intelligence (UAI), pages 550--558, 2005

    2005

  38. [47]

    Solving large imperfect information games using CFR +

    Oskari Tammelin. Solving large imperfect information games using CFR +. arXiv preprint arXiv:1407.5042, 2014

    2014 arXiv

  39. [48]

    Measuring general intelligence with generated games, 2025

    Vivek Verma, David Huang, William Chen, Dan Klein, and Nicholas Tomlin. Measuring general intelligence with generated games, 2025. The gg-bench benchmark; repository vivek3141/gg-bench

    2025

  40. [49]

    Efficient computation of behavior strategies

    Bernhard von Stengel. Efficient computation of behavior strategies. Games and Economic Behavior, 14 0 (2): 0 220--246, 1996

    1996

  41. [50]

    Powley, and Peter I

    Daniel Whitehouse, Edward J. Powley, and Peter I. Cowling. Determinization and information set Monte Carlo tree search for the card game Dou Di Zhu . In IEEE Conference on Computational Intelligence and Games (CIG) , 2011

    2011

  42. [51]

    Edwin B. Wilson. Probable inference, the law of succession, and statistical inference. Journal of the American Statistical Association, 22 0 (158): 0 209--212, 1927

    1927

  43. [52]

    Regret minimization in games with incomplete information

    Martin Zinkevich, Michael Johanson, Michael Bowling, and Carmelo Piccione. Regret minimization in games with incomplete information. In Advances in Neural Information Processing Systems (NeurIPS), 2007

    2007

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.