Pith. sign in

REVIEW 3 major objections 5 minor 26 references

This paper claims that recovering the qualitative form of a dynamics by active probing, then fitting its parameters, turns interaction traces into executable code world models that a planner can use, reaching 69% mean planning success and b

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 03:00 UTC pith:PZJC6EVW

load-bearing objection VPW is a credible, well-scoped demonstration of structure-then-fit program induction for planning, but its headline margin depends on hand-authored hypothesis classes and single-run baselines; worth refereeing, not without revisions. the 3 major comments →

arxiv 2607.25236 v1 pith:PZJC6EVW submitted 2026-07-28 cs.CL cs.RO

VisualPatchWorld: Code World Models as Latent Structured Representations for Planning

Jiaxin Bai , Jiaxuan Xiong This is my paper
classification cs.CL cs.RO
keywords world modelsprogram inductionactive probingmodel-predictive controlplanningscene graphscode generationcontact-rich manipulation
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.

VisualPatchWorld (VPW) claims that world dynamics can be recovered from interaction traces as executable Python programs that behave like small physics engines, and that the key is to first select the qualitative form of the dynamics through short active probes, then fit the free parameters from multi-step rollout loss. On four control domains spanning navigation, reaching, cube manipulation, and contact-rich pushing, the induced programs used as scorers inside model-predictive control achieve 69.0% mean planning success, 23.5 points above the strongest prior code-based world model. Induced scoring alone approaches the ground-truth physics engine on navigation and grasping, while contact-rich pushing benefits from a hybrid scheme that re-scores the top 30% of candidate plans in the engine. The paper argues that separating structure selection from parameter fitting is what makes code world models planner-usable rather than merely locally predictive.

Core claim

VPW's central claim is that the qualitative form of a dynamics must be chosen before its parameters are fit, and that this form can be discovered by probing the environment rather than generated open-endedly. A shared two-level induction runs on every domain: Level 1 uses short discriminating experiments to select among a small, hand-supplied family of dynamical sketches (contact-driven control, linear navigation, grip-gated motion, joint-space kinematics), and Level 2 fits the chosen template's constants by minimizing accumulated prediction error over the planning horizon. The resulting program is an executable simulator—inspectable, editable, and usable to score candidate actions in recedi

What carries the argument

The load-bearing object is the 'dynamical sketch': a compact, hand-authored family of qualitative transition rules (for example, contact-driven versus free motion, joint-space versus Cartesian kinematics) from which Level 1 active probing selects one, and which Level 2 then fits. The sketch family is the human-supplied prior; everything else—probing, rollout fitting, planning—is shared across domains. The two-level separation prevents free-form code generation from fitting local transitions while encoding the wrong overall dynamics, which the paper shows is the failure mode of prior executable world models.

Load-bearing premise

Each domain's true dynamics is assumed to lie within the small set of hand-written qualitative sketches the authors supply; active probing can only choose among these candidates, so if the correct form is absent the fitted program has the wrong structure and planning success collapses, as the pushing ablations show.

What would settle it

Run VPW on the pushing domain with the contact-driven sketch removed from the candidate family; if induced-only planning success remains at or above 22%, then the particular qualitative form is not the cause of the gain; if it falls to near zero, the structure-selection claim is confirmed.

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

If this is right

  • Induced code scoring recovers 91–96% of the ground-truth engine's success on navigation and cube manipulation, so code world models can serve as drop-in scoring models for closed-loop planning in these domains.
  • Structure selection is the dominant factor: on reaching, recovering joint-space kinematics plus forward kinematics raises planning success from the single digits to 72%, and on pushing only the contact-driven sketch yields nonzero success.
  • Hybrid scoring—using the induced model to search and the physics engine to re-score the top 30% of candidates—raises pushing success from 22% to 96% while using about 70% fewer engine queries than full-engine planning.
  • The multi-step rollout objective is necessary for planner-usable dynamics: a linear template can achieve low one-step error yet fail to support goal-directed pushing.
  • Image-derived scene graphs are a workable replanning interface, with tool-based extraction reaching 61% mean success under induced scoring, close to the 69% obtained with simulator-state observation.

Where Pith is reading between the lines

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

  • The method's reach is bounded by the human-supplied sketch library; extending the family to new qualitative behaviors (friction, deformation, nonlinear springs, multi-contact) is the natural next test, and the paper's four domains are all ones where a small discrete set of laws plausibly holds.
  • Because Level 1 requires a simulator that can be reset and stepped, the pipeline is a simulator-to-code distillation rather than a fully model-free learner; whether the same probing sketches transfer to real-robot settings, where resets are expensive, is an open question.
  • The hybrid re-scoring pattern suggests a general compute-accuracy trade-off: a cheap induced model proposes, an expensive accurate model verifies a shortlist; this could transfer to other surrogate-based planning pipelines beyond world models.
  • The observed failure of the vision-language path to produce metric coordinates hints that perception, not dynamics induction, is the current bottleneck for fully visual code world models.

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

3 major / 5 minor

Summary. The paper proposes VisualPatchWorld (VPW), a two-level method for inducing executable code world models from state–action traces. Level 1 selects a qualitative dynamical sketch by active probing over a small, per-environment hypothesis family (contact PD vs memoryless, quasi-static vs momentum, contact-driven vs action-driven, etc.). Level 2 fits the sketch's free parameters by multi-step rollout loss with multi-restart selection. The resulting program is then used as the scoring model inside CEM-MPC, with oracle or image-derived scene graphs supplying replan state; an optional hybrid mode re-scores the top 30% of candidate plans in the MuJoCo engine. On four LeWM domains the paper reports 69.0% mean success for Oracle+Induced scoring, 23.5 points above the best ported programmatic baseline, and near-engine performance on navigation and grasp-rich control, with a residual gap on contact-rich pushing that hybrid scoring largely closes.

Significance. The paper is valuable in establishing a concrete pipeline that turns interaction data into inspectable, planner-usable code world models, and in separating qualitative structure selection from parameter fitting. It explicitly evaluates on held-out starts/goals, uses multi-restart fitting, provides component ablations, and reports five-seed confidence intervals for VPW conditions. If the central comparison is trustworthy, the two-level induction idea is a useful step beyond free-form code synthesis for planning. However, the strength of the claim rests on two assumptions that the paper does not fully secure: that the per-domain sketch families are complete enough to contain the true dynamics, and that the ported baselines faithfully represent the prior methods. These are load-bearing for the headline 23.5-point margin and for the generality claim.

major comments (3)
  1. [§5.1, Table 5, Table 6] The method's correctness depends on the human-supplied Level 1 sketch family containing the true dynamics. In all four environments the recovered sketch is exactly the ground-truth class (Table 5), and the PushT ablation (Table 6) shows that a linear template collapses to 0% planning success. No experiment is reported in which the correct qualitative form is absent from the family or in which distractor families compete. The claim that VPW 'chooses the correct qualitative form' is therefore only validated within hand-authored families that the authors already know to include the true law. This is a load-bearing limitation for the paper's generality claim; the manuscript should either add family-robustness experiments (e.g., omitting the true sketch on one or more domains, or widening the family with plausible but incorrect forms) or substantially soften the general-routing statements.
  2. [Appendix D, Table 1] The main-table comparison against prior code world models may understate the baselines. Appendix D states that WorldCoder omits online optimism and live environment interaction, PoE-World omits online expert debugging and hierarchical Atari planners, and CWM-Game omits value-function synthesis and hidden-state inference. These are not cosmetic omissions: the cited methods' reported strengths often come from exactly those components. Since the headline claim is a 23.5-point gain over the strongest code baseline, the paper needs to either justify that the omitted components are inapplicable under the shared LeWM interface, or run the full original methods where feasible. At minimum, multi-seed intervals for the baselines are needed; the current reporting convention (§5.1) gives only seed-42 point estimates for all code baselines, while five-seed intervals are provided only for VPW conditio
  3. [§2, Algorithm 3] There is a mismatch between the stated problem and the induction procedure. Section 2 defines the mining task from a dataset of transitions D={(o_t,a_t,o_{t+1})}, which suggests offline recorded traces, but Algorithm 3 requires a black-box simulator with reset and step calls to perform active probing. Level 2 can fit from offline traces, but Level 1 cannot. The paper should clarify whether the reported results assume online reset/step access, and, if so, state that explicitly as a requirement; otherwise the title claim of recovering code world models from interaction trajectories is incomplete. If an offline-only variant is intended, the authors should explain how sketch selection would work without environment access.
minor comments (5)
  1. [Table 1] Several cells appear to have lost column separators in the rendered text, e.g., '988 0 66' and '9818 0 66' in the PatchWorld and POMDP-Coder rows. Please format the table so each number is clearly aligned with its environment column.
  2. [Algorithm 4] The rollout loss in line 4 is missing the inner transpose/norm formatting; as typeset it reads as ||\hat{g}_t(θ)-g_t||^2 without indicating which norm. Please clarify the norm used.
  3. [§4.1] The tool path is described as an 'engineered perception module' with environment-specific extractors. This is acceptable, but the paper should state more prominently that the reported main results depend on these hand-designed extractors, not on a general visual abstraction method.
  4. [Appendix F] The five-seed intervals are reported only for certain VPW conditions. It would be helpful to also report the five-seed means for the stronger ported baselines (at least PatchWorld and POMDP-Coder) so the reader can assess whether the 23.5-point margin is meaningful relative to baseline variability.
  5. [§5.6] For Cube, Oracle+Hybrid (84%) is slightly below Oracle+Induced (86%) on n=50. The text says hybrid 'need not help' when induced is near ceiling, but the raw numbers suggest a small decrease. A brief note about this being a one-trial difference would prevent confusion.

Circularity Check

0 steps flagged

No circular derivation; VPW's planning gains are evaluated on held-out starts/goals, with the hand-authored sketch families disclosed as an explicit input/limitation rather than a concealed fit.

full rationale

VPW's derivation chain is not circular in the sense of the target being used as an input. Level 2 parameters are fit by minimizing multi-step rollout error on training sim-state traces (Algorithm 4) and then evaluated on 50 held-out starts/goals under frozen planners (Section 5.1), so planning success is not a fitted target. The recovered qualitative sketches come from Level-1 active probing over a small candidate family, which the paper explicitly discloses as an environment-specific input ('Each domain supplies a compact Level 1 hypothesis family', Section 5.1; 'Scope and setup... Each domain supplies a compact Level 1 sketch family under a shared induction procedure', Section 6). This is a prior/limitation rather than a circular reduction: Table 6 shows that a wrong template (linear) yields 0% planning success despite low one-step error, so the reported 69.0% mean is not forced by the fitting loss. Active probing does use black-box reset/step access, but only to select among disclosed hypotheses, not to define the held-out evaluation labels. Self-citations ([1] PatchWorld, [22] NewtonBench) appear as baselines/related work, not as a load-bearing justification or uniqueness theorem; the core two-level induction is implemented directly in Algorithms 3-5. Hybrid re-scoring is explicitly separated from induced-only results (Tables 1 vs 7) and is reported as an optional engine check, not a hidden input. The main unresolved weakness is the hand-authored completeness of each sketch family and the lack of out-of-distribution tests, but these are correctness/scope concerns, not circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The central claim rests on two empirical choices: (1) hand-authored per-domain sketch families that contain the correct dynamics class, and (2) faithful baseline ports. These are the main sources of unknown risk in the comparison.

free parameters (7)
  • Per-domain sketch family (hypothesis class) = n/a (hand-authored; e.g., contact PD, linear navigation, grip-gated, joint-space FK)
    Level 1 selects among these; including the correct qualitative class is what drives planning gains. If the family omits the true dynamics, Level 2 cannot recover it.
  • PushT PD gains KP, KV, SCALE = ≈92, ≈26, ≈98
    Fitted in Level 2 from simulator-state traces; reported in Figure 7.
  • PushT contact spring constants = not reported numerically
    Normal/tangential springs in the quasi-static block push; fitted from traces and central to contact behavior.
  • Two-room linear transition matrix W = not reported
    Ridge fit on agent/goal coordinates and distance features; recovered linear navigation.
  • Cube ridge-regression action map = not reported
    Fitted 5-D action to 6-D state deltas; underlies grip-gated contact.
  • Reacher joint dynamics B and FK parameters = not reported
    Ridge fit qpos_{t+1} = qpos_t + B·action plus fitted forward kinematics.
  • Planner budgets per environment (CEM samples, iterations, frame stride) = Table 9
    Chosen so MuJoCo ground-truth success is at least 90%; these protocol choices affect all rows equally but inflate the ceiling.
axioms (5)
  • ad hoc to paper Each environment's true transition law is a member of the small per-domain hypothesis family.
    Level 1 cannot select a sketch outside the family; the four families were chosen by the authors to match the LeWM domains.
  • domain assumption The target simulator can be reset and stepped for active probing without cost.
    Algorithm 3 calls reset()/step() to choose the sketch; real-robot or expensive environments may not permit this.
  • domain assumption Oracle or tool-derived scene graphs contain sufficient state for Markovian dynamics.
    Stage 1 abstraction must preserve geometric/relational cues; if state is partially observed, f_theta is misspecified.
  • ad hoc to paper Baseline ports faithfully represent the cited methods despite omitted components.
    Appendix D removes online interaction, hidden-state inference, and hierarchical planners; the 'strongest code baseline' margin depends on these ports.
  • ad hoc to paper Frozen planners with MuJoCo success at least 90% are a fair common testbed.
    Planner budgets chosen per environment; favorable ceilings may compress the gap between induced and engine scoring.

pith-pipeline@v1.3.0-alltime-deepseek · 20558 in / 11188 out tokens · 113047 ms · 2026-08-01T03:00:05.120521+00:00 · methodology

0 comments
read the original abstract

Different research lines use the term world model in different ways, yet they share a common aim: to capture how the world evolves under action in a form that supports perception, simulation, and planning. Two prominent realizations are neural predictors that learn dynamics in continuous vector spaces, and hand-built physics engines that expose explicit state and physical laws. Neural predictors scale from data but leave the form of the dynamics implicit; physics engines are inspectable and editable but difficult to construct at scale. We introduce VisualPatchWorld (VPW), which represents world dynamics as code. VPW first selects a qualitative dynamical form with short active probes, then fits that form's free parameters from recorded state-action traces by minimizing multi-step prediction error. The resulting programs can be rolled forward like a simulator, inspected in source form, and used inside model-predictive control; image-derived scene graphs can supply the live state at replan time. Across comparisons with prior code-based world models, VPW attains 69.0% mean planning success and exceeds the strongest code baseline by 23.5 points. The largest gains arise when choosing the correct qualitative dynamics is essential. Under the same planner, the induced models approach ground-truth engine success on navigation and grasp-rich control; a residual gap remains for contact-rich pushing, and checking a shortlist of promising plans in the engine closes most of that gap. These results establish a practical route toward automatically constructed code world models that are useful for planning. Code is available at https://github.com/HKBU-KnowComp/VisualPatchWorld/.

Figures

Figures reproduced from arXiv: 2607.25236 by Jiaxin Bai, Jiaxuan Xiong.

Figure 1
Figure 1. Figure 1: World models share one agent–environment loop, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: VisualPatchWorld pipeline. Level 1 selects a dynamical sketch by active probing; Level 2 fits its parameters from [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Two-level program induction (shared across domains), illustrated on a contact-push example. Level 1 selects a [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Rollout analysis for code baselines versus VPW, paired with Table [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Tool-extracted scene graphs versus oracle graphs on four LeWM environments. Extraction preserves object recall [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: VLM PushT scene graphs versus oracle graphs. Semantic structure is recovered, but metric coordinates are misaligned, [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Induced executable world models. Each panel pairs a frame and oracle scene graph with induced rules, one symbolic [PITH_FULL_IMAGE:figures/full_fig_p015_7.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

26 extracted references · 14 linked inside Pith

  1. [1]

    Pan, and Yangqiu Song

    Jiaxin Bai, Yue Guo, Yifei Dong, Jiaxuan Xiong, Tianshi Zheng, Yixia Li, Tianqing Fang, Yufei Li, Yisen Gao, Haoyu Huang, Zhongwei Xie, Hong Ting Tsang, Zihao Wang, Lihui Liu, Jeff Z. Pan, and Yangqiu Song. 2026. PatchWorld: Gradient- Free Optimization of Executable World Models. arXiv:2605.30880 [cs.AI] https: //arxiv.org/abs/2605.30880

    Pith/arXiv arXiv 2026

  2. [2]

    Tenenbaum, Tomás Lozano-Pérez, and Leslie Pack Kaelbling

    Aidan Curtis, Hao Tang, Thiago Veloso, Kevin Ellis, Joshua B. Tenenbaum, Tomás Lozano-Pérez, and Leslie Pack Kaelbling. 2025. LLM-Guided Probabilistic Pro- gram Induction for POMDP Model Estimation. InProceedings of The 9th Con- ference on Robot Learning (Proceedings of Machine Learning Research, Vol. 305), Joseph Lim, Shuran Song, and Hae-Won Park (Eds.)...

    2025

  3. [3]

    Nicola Dainese, Matteo Merler, Minttu Alakuijala, and Pekka Marttinen. 2024. Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search. InAdvances in Neural Information Processing Sys- tems. arXiv:2405.15383 [cs.AI]

    Pith/arXiv arXiv 2024

  4. [4]

    Fan Feng, Phillip Lippe, and Sara Magliacane. 2025. Learning Interactive World Model for Object-Centric Reinforcement Learning. InAdvances in Neural Infor- mation Processing Systems. arXiv:2511.02225 [cs.LG]

    arXiv 2025

  5. [5]

    Dibya Ghosh, Abhishek Gupta, Ashwin Reddy, Justin Fu, Coline Devin, Benjamin Eysenbach, and Sergey Levine. 2021. Learning to Reach Goals via Iterated Supervised Learning. InInternational Conference on Learning Representations. arXiv:1912.06088 [cs.LG]

    Pith/arXiv arXiv 2021

  6. [6]

    John Gkountouras, Matthias Lindemann, Phillip Lippe, Efstratios Gavves, and Ivan Titov. 2025. Language Agents Meet Causality: Bridging LLMs and Causal World Models. InInternational Conference on Learning Representations. arXiv:2410.19923 [cs.LG]

    Pith/arXiv arXiv 2025

  7. [7]

    Arya Grayeli, Atharva Sehgal, Omar Costilla-Reyes, Miles Cranmer, and Swarat Chaudhuri. 2024. Symbolic Regression with a Learned Concept Library. In Advances in Neural Information Processing Systems. arXiv:2409.09359 [cs.LG]

    Pith/arXiv arXiv 2024

  8. [8]

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. 2025. Mas- tering diverse control tasks through world models.Nature640 (2025), 647–653. doi:10.1038/s41586-025-08744-2

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

  9. [9]

    Zaid Khan, Archiki Prasad, Elias Stengel-Eskin, Jaemin Cho, and Mohit Bansal

  10. [10]

    Ilya Kostrikov, Ashvin Nair, and Sergey Levine. 2022. Offline Reinforcement Learning with Implicit Q-Learning. InInternational Conference on Learning Rep- resentations. arXiv:2110.06169 [cs.LG]

    Pith/arXiv arXiv 2022

  11. [11]

    Schultz, Marcus Chiam, Ian Gemp, Piotr Zielinski, Satinder Singh, and Kevin P

    Wolfgang Lehrach, Daniel Hennes, Miguel Lázaro-Gredilla, Xinghua Lou, Carter Wendelken, Zun Li, Antoine Dedieu, Jordi Grau-Moya, Marc Lanctot, Atıl Işcen, John R. Schultz, Marcus Chiam, Ian Gemp, Piotr Zielinski, Satinder Singh, and Kevin P. Murphy. 2025. Code World Models for General Game Playing.arXiv preprint arXiv:2510.04542(2025)

    arXiv 2025

  12. [12]

    2026.A Functional Taxonomy of World Models: Renderers, Simulators, Planners, and the Loop That Connects Them

    Fei-Fei Li. 2026.A Functional Taxonomy of World Models: Renderers, Simulators, Planners, and the Loop That Connects Them. World Labs blog, Substack. https: //drfeifei.substack.com/p/a-functional-taxonomy-of-world-models

    2026

  13. [13]

    Lucas Maes, Quentin Le Lidec, Damien Scieur, Yann LeCun, and Randall Balestriero. 2026. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312 [cs.LG] https://arxiv.org/abs/2603.1 9312

    Pith/arXiv arXiv 2026

  14. [14]

    Seohong Park, Kevin Frans, Benjamin Eysenbach, and Sergey Levine. 2025. OG- Bench: Benchmarking Offline Goal-Conditioned RL. InInternational Conference on Learning Representations. arXiv:2410.20092 [cs.LG]

    Pith/arXiv arXiv 2025

  15. [15]

    Wasu Top Piriyakulkij, Yichao Liang, Hao Tang, Adrian Weller, Marta Kryven, and Kevin Ellis. 2025. PoE-World: Compositional World Modeling with Products of Programmatic Experts. InAdvances in Neural Information Processing Systems. arXiv:2505.10819 [cs.AI]

    arXiv 2025

  16. [16]

    Parshin Shojaee, Kazem Meidani, Shashank Gupta, Amir Barati Farimani, and Chandan K. Reddy. 2025. LLM-SR: Scientific Equation Discovery via Program- ming with Large Language Models. InInternational Conference on Learning Representations. arXiv:2404.18400 [cs.LG]

    Pith/arXiv arXiv 2025

  17. [17]

    Vlad Sobal, Wancong Zhang, Kyunghyun Cho, Randall Balestriero, Tim G. J. Rudner, and Yann LeCun. 2025. Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models. InAdvances in Neural Information Processing Systems. arXiv:2502.14819 [cs.LG]

    arXiv 2025

  18. [18]

    Hao Tang, Keya Hu, Jin Peng Zhou, Sicheng Zhong, Wei-Long Zheng, Xu- jie Si, and Kevin Ellis. 2024. Code Repair with LLMs gives an Exploration- Exploitation Tradeoff. InAdvances in Neural Information Processing Systems. arXiv:2405.17503 [cs.SE]

    Pith/arXiv arXiv 2024

  19. [19]

    Hao Tang, Darren Yan Key, and Kevin Ellis. 2024. WorldCoder, a Model- Based LLM Agent: Building World Models by Writing Code and Interacting with the Environment. InAdvances in Neural Information Processing Systems. arXiv:2402.12275 [cs.LG]

    Pith/arXiv arXiv 2024

  20. [20]

    Zizhao Wang, Kaixin Wang, Li Zhao, Peter Stone, and Jiang Bian. 2025. Dyn- O: Building Structured World Models with Object-Centric Representations. In Advances in Neural Information Processing Systems

    2025

  21. [21]

    Jialong Wu, Shaofeng Yin, Ningya Feng, and Mingsheng Long. 2025. RLVR- World: Training World Models with Reinforcement Learning. InAdvances in Neural Information Processing Systems. arXiv:2505.13934 [cs.LG]

    arXiv 2025

  22. [22]

    Nguyen, Baixuan Xu, Zhaowei Wang, Jiayang Cheng, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Tian- qing Fang, Yangqiu Song, Ginny Y

    Tianshi Zheng, Kelvin Kiu-Wai Tam, Newt Hue-Nam K. Nguyen, Baixuan Xu, Zhaowei Wang, Jiayang Cheng, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Tian- qing Fang, Yangqiu Song, Ginny Y. Wong, and Simon See. 2026. NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents. InInter- national Conference on Learning Representations. arXiv:2510....

    arXiv 2026

  23. [23]

    Yaofeng Desmond Zhong, Jiequn Han, and Georgia Olympia Brikis. 2022. Dif- ferentiable Physics Simulations with Contacts: Do They Have Correct Gradi- ents w.r.t. Position, Velocity and Control? arXiv:2207.05060 [cs.LG] https: //arxiv.org/abs/2207.05060

    Pith/arXiv arXiv 2022

  24. [24]

    Gaoyue Zhou, Hengkai Pan, Yann LeCun, and Lerrel Pinto. 2025. DINO- WM: World Models on Pre-trained Visual Features Enable Zero-shot Plan- ning. InProceedings of the 42nd International Conference on Machine Learn- ing (Proceedings of Machine Learning Research, Vol. 267). PMLR, 79115–79135. arXiv:2411.04983 [cs.RO] VisualPatchWorld: Code World Models as La...

    Pith/arXiv arXiv 2025

  25. [2026]

    InInternational Conference on Learning Representations

    One Life to Learn: Inferring Symbolic World Models for Stochastic Envi- ronments from Unguided Exploration. InInternational Conference on Learning Representations. arXiv:2510.12088 [cs.AI]

    Pith/arXiv arXiv

  26. [3184]

    https://proceedings.mlr.press/v305/curtis25a.html