Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

ActSWM: Action-Sensitive World Models for Long-Horizon Planning in Open-World Games

T0 review · 3 major / 5 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Latent world models fail at planning when futures stop depending on actions; ActSWM keeps those futures separable.

desk verdict Solid engineering paper: names Context Collapse, freezes a readout + hinge to keep JEPA rollouts action-sensitive, and shows real MineStudio and multi-game recovery gains—with one diagnostic that overlaps the training contrast. read the letter →

arxiv 2607.26712 v1 pith:74S5DKEW submitted 2026-07-29 cs.RO

Zhenfeng Gan , ZiTong Zeng , Jiajun Cheng , Yeke Song , Yongyi Tang , Xueqian Wang This is my paper · ORCID
classification cs.RO
keywords latentworldmodelsactionsensitivitycontextcollapsemodel-predictivecontrolJEPAMinecraftplanningrecoveryopen-worldgames
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

Latent world models let agents plan by rolling out futures in a compact code space instead of pixels. This paper argues that matching future states is not enough: if different action sequences produce nearly the same predicted futures, a planner cannot tell good controls from bad ones. The authors name that failure Context Collapse—rollouts that still look like plausible futures while becoming almost insensitive to the actions they are conditioned on. ActSWM counters it with a transition-separation rule: keep alternative-action futures distinguishable over long horizons, and keep each local latent transition informative enough that the action can be recovered. They enforce this with a hinge that pushes recorded-action rollouts away from all-zero-action rollouts, plus a frozen action readout that forces transitions to stay action-discriminative. On Minecraft step-drift tests the action gap stays large; closed-loop MineStudio planning gains large success lifts on sustained tasks; and CEM can recover controls from offline videos across several games much better than baselines.

What carries the argument

Transition-separation principle, implemented as multi-step prediction plus an action-contrastive hinge (recorded vs all-zero futures) and a parameter-frozen action readout on encoded and predicted transitions, so gradients reshape the latent dynamics rather than move a trainable decoder boundary.

What would settle it

Rerun the step-drift and CEM planning tests with non-zero alternative action sequences drawn from the same action-block library the planner searches; if ActSWM’s action gaps and task-success gains collapse while prediction similarity stays high, the central claim fails.

Watch

Extended reading notes

Core claim

A planning-useful latent dynamics model must keep futures under different actions distinguishable and make the action of each local transition recoverable. Enforcing that as a constraint on latent rollouts—not only as an auxiliary head—mitigates Context Collapse, preserves large action-dependent rollout gaps, and improves long-horizon closed-loop planning and offline action recovery.

Load-bearing premise

Treating recorded actions versus all-zero actions as the main contrast is assumed to stand in for the planner-relevant property that different real control sequences produce separable futures.

Editorial extensions

If this is right

  • Future-state similarity alone is an incomplete training and evaluation target for planning-oriented latent world models.
  • Closed-loop MPC in open-world games should improve when latent rollouts preserve action-dependent separation over long horizons.
  • The same constrained dynamics can support recovering player controls from offline gameplay video via search in latent space.
  • Diagnostics that report an action gap (recorded vs counterfactual actions) become a practical check for controllability, not only prediction quality.

Reading between the lines

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

  • Any world-model planner that scores candidate sequences in latent space may need an explicit anti-collapse term whenever context is long enough to predict motion from history alone.
  • Frozen readouts as hard constraints may transfer to other inverse-dynamics regularizers where joint training lets the head absorb the loss without separating transitions.
  • If zero-action counterfactuals overstate separation, training against sampled planner-like alternatives could close the gap between the diagnostic and real CEM search.
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

3 major / 5 minor

Summary. The paper identifies Context Collapse in latent world models: autoregressive predictors can match encoded futures under recorded actions while producing nearly identical rollouts under alternative (here, all-zero) action sequences, undermining MPC. ActSWM addresses this via a transition-separation principle implemented as (i) multi-step JEPA prediction, (ii) a cosine hinge that separates recorded-action vs all-zero rollouts (Eqs. 5a–5b), and (iii) a frozen action readout on encoded and predicted transitions (Eqs. 8a–8c, Appendix E). On Minecraft VPT step-drift, ActSWM reports a step-31 action gap of 0.760 versus ~0–0.002 for LeWM-style baselines while retaining high GT similarity; in matched CEM planning on MineStudio it raises success (e.g., stone mining 19/20 vs 10/20); and in cross-game CEM action recovery it improves CEM-over-random gaps by up to ~16.5× with higher active-key accuracy in some domains.

Significance. If the result holds, the work cleanly separates prediction fidelity from planner-relevant controllability and gives a practical training recipe (hinge + frozen readout) that improves closed-loop open-world planning without abandoning latent JEPA efficiency. Strengths include a matched-backbone ablation suite (context length, multi-step rollout, joint vs frozen readout), an explicit Lipschitz separation argument for the frozen readout (Appendix E), external closed-loop MineStudio evaluation under a shared CEM planner, and a multi-game offline recovery protocol with a clear CEM-over-random gap. These make action sensitivity a falsifiable, operational property rather than only an auxiliary inverse-dynamics head. The contribution is incremental relative to WAM/inverse-dynamics and LeWM, but the diagnostic framing and planning gains are of clear interest to model-based control and game-agent communities.

major comments (3)
  1. [Method (Eqs. 5, 10); Q1; Fig. 3; Table 5; Appendix C] The headline Context Collapse diagnostic and the training hinge use the same contrast. Eqs. (5a–5b) maximize separation between recorded-action and all-zero rollouts; Q1 then reports Δ_k = s_gt_k − s_0_k (Eq. 10, Fig. 3, Table 5) on that identical pair, including the claimed ~380× gap (0.760 vs ~0.002). This does not make the method circular with Q2/Q3, but it does make the strongest diagnostic number a near-direct optimization target rather than an independent test that “distinct control sequences induce separable latent consequences.” Appendix C’s CEM library (look-down+attack, forward/back, place+pitch, stack chunks, etc.) never ranks the constant zero sequence. Please add at least one Q1-style evaluation under non-zero alternatives drawn from the planner’s action-block distribution (or random non-zero sequences from the empirical pool), and reframe the recorded-vs-zero gap as a contr
  2. [Abstract; Q2; Figure 4; Table 8] Closed-loop gains are important external evidence, but the reporting mixes relative and absolute figures in a way that overstates robustness. The abstract claims “up to 45% on stone mining and 30% on pillar building,” while §Q2 text says “90.0% on stone mining and 54.5% on pillar building” (relative to LeWM’s base rates), and Table 8 gives 10/20→19/20 and 11/20→17/20 over only 20 trials with no confidence intervals or multiple random seeds for environment/CEM stochasticity. For the central claim that action-sensitive rollouts improve long-horizon interactive planning, please standardize absolute success rates, add binomial CIs or repeated seeds, and clarify whether the abstract’s percentages are absolute points or relative lifts.
  3. [Q3; Table 1; Appendix D] Q3 supports steerability but is uneven and partially under-explained as evidence for the same mechanism. ActSWM’s Gap improves in all three games (Table 1), yet active accuracy falls on Apex (0.576→0.269) while rising sharply on GTA (0.050→0.761), and LeWM uses H=3 while ActSWM uses H=32 (Appendix D), confounding architecture/objective with context length. G-IDM’s zero active accuracy is a weak foil if it collapses to the inactive class. To tie Q3 to the transition-separation claim, please ablate H-matched LeWM vs ActSWM, report variance across the 15 windows, and discuss the Apex active-accuracy drop rather than only the best-case 16.5× Gap.
minor comments (5)
  1. [Figure 3] Figure 3 caption and panel labels are dense; the right-panel gap annotation “=0.760” is easy to miss. Consider a small table callout of GT / Zero / Gap next to the bar chart.
  2. [Figure 3; Method] Notation switches between p_θ, q_ϕ0, and “IDM” in figure legends (Fig. 3 “+IDM”) without defining IDM in the main method text; align terminology with “action readout.”
  3. [Related Work; References] Related Work cites several 2026 arXiv preprints heavily; ensure camera-ready versions and consistent spelling of method names (LeWorldModel / LeWM).
  4. [Introduction; Method] Typos / spacing artifacts from line breaks: “Wetherefore,” “refertothis,” “isenforcedasaconstraint,” “multimodalagentsystems,” etc., appear throughout the Introduction and Method; a full copy-edit pass is needed.
  5. [Appendix A; ActSWM Architecture] Appendix A notes idm.stop_grad=true is misleadingly named; a one-sentence clarification in the main text that gradients still flow into latents would help reimplementers.

Circularity Check

1 steps flagged · score 3.0 of 10

Q1’s headline action-gap partly tracks the hinge objective’s recorded-vs-zero contrast; planning and recovery results remain independent.

  1. fitted input called prediction [Method Eqs. 5a–5b; Q1 Eqs. 10a–10c; Fig. 3; Table 5]
    "ℓ_k = max(0, cos(ẑ^gt_{t+k}, ẑ^0_{t+k}) − (1−m)), L_hinge = (1/K) Σ ℓ_k ... s^gt_k = cos(ẑ^gt_{t+k}, z_{t+k}), s^0_k = cos(ẑ^0_{t+k}, z_{t+k}), Δ_k = s^gt_k − s^0_k. ... ActSWM ... achieves the largest action gap. ... ActSWM, ours 0.923 0.163 0.760"

    Training explicitly penalizes similarity between recorded-action and all-zero rollouts while L_pred fits the recorded-action rollout to true future latents. Q1’s primary sensitivity number is the gap between those same two conditionings’ similarities to the true future. The pair (high s^gt, low s^0) is therefore partly an optimized training target reported as diagnostic evidence of action sensitivity, not an independent probe under planner-relevant non-zero alternatives (CEM’s chunk library). Related to—but not identical with—the hinge (hinge separates ẑ^gt from ẑ^0; Δ compares each to z); still a partial by-construction readout of the trained contrast.

full rationale

ActSWM is an empirical methods paper, not a first-principles derivation. The transition-separation principle is implemented by L_hinge (recorded vs all-zero rollouts) plus a frozen readout, then Q1 reports Δ_k on that same recorded-vs-zero pair. L_pred aligns GT-conditioned rollouts to encoded futures while L_hinge pushes GT and zero rollouts apart, so a large step-31 gap is partly encouraged by training rather than a fully held-out test of “distinct control sequences induce separable futures.” That is mild fitted-objective/diagnostic overlap, not definitional collapse of the paper’s claim: closed-loop MineStudio success and cross-game CEM-over-random recovery use different contrasts, costs, and environments and are not forced by Eq. 5. The Appendix E Lipschitz readout argument is a standard constraint justification, not a self-citation uniqueness import. No load-bearing self-citation chain or renamed known theorem was found. Score 3 reflects one partial by-construction diagnostic, with the central planning-utility claim still externally tested.

Assumptions & free parameters 5 free parameters · 4 assumptions · 3 invented entities

Claims rest on standard latent-MPC practice plus paper-specific modeling choices: JEPA-style latent prediction sufficiency, zero-action counterfactuals as the action-sensitivity probe, a frozen random readout as a stable separation constraint, and CEM with task-specific action chunks as the measure of planning utility. Hyperparameters and the hinge margin are free knobs; Context Collapse and the transition-separation principle are invented organizing entities without independent evidence outside these experiments.

free parameters (5)
  • hinge margin m = 0.3
    Cosine hinge threshold in Eq. 5; set to 0.3 in Table 3, directly shapes how separated GT vs zero rollouts must be.
  • loss weights λ_hinge, λ_readout, λ_sig, α_pred = 0.5, 1.0, 0.09, 1.0
    Trade off prediction, SigReg, separation, and readout; chosen as 0.5 / 1.0 / 0.09 / 1.0 rather than derived.
  • training rollout horizon K and context H = H=32, K=12 (train); Keval=32; Kplan=12
    H=32, K=12 for training; Keval=32 and Kplan=12 at test—design choices that define long-horizon claims.
  • CEM planner hyperparameters and task action-chunk libraries = e.g. 512 candidates, 6 iters, task-specific chunk weights
    Candidate counts, elite fraction, and hand-designed Minecraft chunks/weights (Table 7) strongly affect closed-loop success and are not learned end-to-end.
  • frozen readout initialization ϕ0 = random init, hidden dim 512, trainable=no
    Randomly initialized fixed decoder; separation guarantee depends on this random linear/MLP geometry remaining a useful probe.
assumptions (4)
  • domain assumption Latent next-embedding prediction (without pixel reconstruction) is a sufficient dynamics interface for MPC cost J(a) based on latent distance to reference goals.
    Inherited from JEPA/LeWM/Dreamer-style planning; stated in Preliminaries and Planning objective.
  • ad hoc to paper Distinguishability under recorded vs all-zero future actions is an adequate operational test of action sensitivity for planning.
    Explicitly chosen in Modeling Action Sensitivity and reused for L_hinge and Q1; not shown equivalent to separation under the planner’s real alternative chunks.
  • standard math A parameter-frozen, locally Lipschitz action readout induces a positive margin between latent transitions of different actions when readout error is small (Eq. 3 / Appendix E).
    Triangle inequality + Lipschitz; correct as an inequality, but does not prove the random frozen net is an optimal or tight constraint.
  • domain assumption Offline VPT/MineStudio and cleaned multi-game H5 windows are representative enough that step-drift and CEM recovery predict interactive planning utility.
    Underpins transfer from Q1/Q3 diagnostics to Q2 claims; data pipelines in Appendices F–G.
invented entities (3)
  • Context Collapse
    purpose: Name the failure where autoregressive latent predictors match future states yet yield nearly identical rollouts under different actions.
    Central diagnostic construct; evidenced only via the paper’s GT-vs-zero cosine protocol, not an externally standardized benchmark.
  • Action-Readout Separation Principle / transition-separation principle
    purpose: Justify freezing q_ϕ0 and treating action recoverability as a constraint on latent transitions and rollouts.
    Organizing principle for ActSWM; related to prior inverse-dynamics work but packaged here as the paper’s design law.
  • ActSWM
    purpose: The proposed JEPA-based model combining multi-step prediction, hinge rollout separation, SigReg, and frozen readout.
    Method entity evaluated against LeWM variants; no external replication yet.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ActSWM: Action-Sensitive World Models for Long-Horizon Planning in Open-World Games." pith.science (2026). https://pith.science/paper/74S5DKEW

@misc{pith2026260726712,
  author       = {Pith},
  title        = {Pith review of: ActSWM: Action-Sensitive World Models for Long-Horizon Planning in Open-World Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/74S5DKEW}},
  note         = {Machine review of arXiv:2607.26712}
}
read the original abstract

Latent world models support efficient model-predictive control by optimizing future control sequences in latent space and replanning in a receding-horizon manner. However, existing latent predictors often lack stable long-horizon rollout ability, and prediction accuracy alone does not ensure that rollouts remain responsive to the actions being planned. We identify Context Collapse, a failure mode in which autoregressive latent predictors maintain high similarity to future states while producing nearly indistinguishable futures under different action sequences. To address this issue, we propose ActSWM, an action-sensitive latent world model grounded in a transition-separation principle: a planning-useful latent dynamics model should keep alternative-action futures distinguishable and make the action associated with each local transition recoverable. Under this principle, action sensitivity is enforced as a constraint on latent rollouts rather than treated only as an auxiliary prediction target, encouraging predicted futures to preserve action-dependent differences over long horizons. Across step-drift analysis, closed-loop Minecraft planning, and cross-game local action recovery, ActSWM preserves larger action-dependent rollout gaps than existing baselines, improves task success in long-horizon interactive settings, and enables world-model-based action recovery from offline gameplay videos.

Figures

Figures reproduced from arXiv: 2607.26712 by the authors.

Figure 1
Figure 1. Overview of ActSWM. Latent world models can suffer from context collapse, where long-horizon rollouts match [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. ActSWM framework. ActSWM combines JEPA-based latent prediction, multi-step rollout training, an action-sensitivity [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Q1 offline step-drift diagnostic on Minecraft VPT trajectories. The left panel tracks recorded-action prediction quality [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Success-rate comparison on Minecraft planning [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Reference trajectories for the three Minecraft planning tasks. Each row shows the target video segment for one task [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Overview of the gameplay data cleaning pipeline. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Multi-track game data construction framework. GTA V and Apex share web-video discovery, segmentation, and [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Metric Non-Collapse in Learned World Models for Control: Approximation Theory, Finite-Sample Geometric Guarantees, and Deterministic Planning Transfer

    math.OC 2026-08 conditional novelty 7.0 of 10

    A mathematically justified local-global metric regularizer converts approximately optimized empirical world models into provably non-collapsed encoders with controlled planning transfer for deterministic nonlinear control.

Reference graph

Works this paper leans on

44 extracted references · cited by 1 Pith paper

  1. [1]

    Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos , url =

    Baker, Bowen and Akkaya, Ilge and Zhokov, Peter and Huizinga, Joost and Tang, Jie and Ecoffet, Adrien and Houghton, Brandon and Sampedro, Raul and Clune, Jeff , booktitle =. Video PreTraining (VPT): Learning to Act by Watching Unlabeled Online Videos , url =

  2. [2]

    STEVE-1: A Generative Model for Text-to-Behavior in Minecraft , url =

    Lifshitz, Shalev and Paster, Keiran and Chan, Harris and Ba, Jimmy and McIlraith, Sheila , booktitle =. STEVE-1: A Generative Model for Text-to-Behavior in Minecraft , url =

  3. [3]

    JARVIS-1: Open-World Multi-Task Agents With Memory-Augmented Multimodal Language Models , year=

    Wang, Zihao and Cai, Shaofei and Liu, Anji and Jin, Yonggang and Hou, Jinbing and Zhang, Bowei and Lin, Haowei and He, Zhaofeng and Zheng, Zilong and Yang, Yaodong and Ma, Xiaojian and Liang, Yitao , journal=. JARVIS-1: Open-World Multi-Task Agents With Memory-Augmented Multimodal Language Models , year=

  4. [4]

    Zihao Wang and Shaofei Cai and Zhancun Mu and Haowei Lin and Ceyao Zhang and Xuejie Liu and Qing Li and Anji Liu and Xiaojian Ma and Yitao Liang , booktitle=. Omni. 2024 , url=

    2024

  5. [5]

    2025 , eprint=

    Game-TARS: Pretrained Foundation Models for Scalable Generalist Multimodal Game Agents , author=. 2025 , eprint=

    2025

  6. [6]

    2025 , eprint=

    Lumine: An Open Recipe for Building Generalist Agents in 3D Open Worlds , author=. 2025 , eprint=

    2025

  7. [7]

    2024 , eprint=

    Mastering Diverse Domains through World Models , author=. 2024 , eprint=

    2024

  8. [8]

    2019 , eprint=

    Learning Latent Dynamics for Planning from Pixels , author=. 2019 , eprint=

    2019

Show all 44 references
  1. [9]

    2024 , eprint=

    TD-MPC2: Scalable, Robust World Models for Continuous Control , author=. 2024 , eprint=

    2024

  2. [10]

    Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , year=

    Assran, Mahmoud and Duval, Quentin and Misra, Ishan and Bojanowski, Piotr and Vincent, Pascal and Rabbat, Michael and LeCun, Yann and Ballas, Nicolas , booktitle=. Self-Supervised Learning from Images with a Joint-Embedding Predictive Architecture , year=

  3. [11]

    2024 , eprint=

    Revisiting Feature Prediction for Learning Visual Representations from Video , author=. 2024 , eprint=

    2024

  4. [12]

    2026 , eprint=

    LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels , author=. 2026 , eprint=

    2026

  5. [13]

    2026 , eprint=

    Enhancing Policy Learning with World-Action Model , author=. 2026 , eprint=

    2026

  6. [14]

    2022 , eprint=

    Temporal Difference Learning for Model Predictive Control , author=. 2022 , eprint=

    2022

  7. [15]

    Diffusion for World Modeling: Visual Details Matter in Atari , url =

    Alonso, Eloi and Jelley, Adam and Micheli, Vincent and Kanervisto, Anssi and Storkey, Amos and Pearce, Tim and Fleuret, Fran. Diffusion for World Modeling: Visual Details Matter in Atari , url =. Advances in Neural Information Processing Systems , doi =

  8. [16]

    2025 , eprint=

    DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning , author=. 2025 , eprint=

    2025

  9. [17]

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , url =

    Bengio, Samy and Vinyals, Oriol and Jaitly, Navdeep and Shazeer, Noam , booktitle =. Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , url =

  10. [18]

    2011 , eprint=

    A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning , author=. 2011 , eprint=

    2011

  11. [19]

    Professor Forcing: A New Algorithm for Training Recurrent Networks , url =

    Lamb, Alex M and ALIAS PARTH GOYAL, Anirudh Goyal and Zhang, Ying and Zhang, Saizheng and Courville, Aaron and Bengio, Yoshua , booktitle =. Professor Forcing: A New Algorithm for Training Recurrent Networks , url =

  12. [20]

    Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence , pages =

    Talvitie, Erik , title =. Proceedings of the Thirtieth Conference on Uncertainty in Artificial Intelligence , pages =. 2014 , isbn =

    2014

  13. [21]

    2026 , eprint=

    Hierarchical Planning with Latent World Models , author=. 2026 , eprint=

    2026

  14. [22]

    , year =

    Johnson, William and Lindenstrauss, J. , year =. Extensions of Lipschitz mappings into a Hilbert space , volume =

  15. [23]

    Random Features for Large-Scale Kernel Machines , url =

    Rahimi, Ali and Recht, Benjamin , booktitle =. Random Features for Large-Scale Kernel Machines , url =

  16. [24]

    and Cownden, Daniel and Tweed, Douglas B

    Lillicrap, Timothy P. and Cownden, Daniel and Tweed, Douglas B. and Akerman, Colin J. , date =. Random synaptic feedback weights support error backpropagation for deep learning , volume =. doi:10.1038/ncomms13276 , abstract =

    doi

  17. [25]

    2016 , eprint=

    Direct Feedback Alignment Provides Learning in Deep Neural Networks , author=. 2016 , eprint=

    2016

  18. [26]

    2025 , eprint=

    MineStudio: A Streamlined Package for Minecraft AI Agent Development , author=. 2025 , eprint=

    2025

  19. [27]

    and Mannor, Shie and Rubinstein, Reuven Y

    de Boer, Pieter-Tjerk and Kroese, Dirk P. and Mannor, Shie and Rubinstein, Reuven Y. , date =. A tutorial on the cross-entropy method , volume =. doi:10.1007/s10479-005-5724-z , abstract =

    doi

  20. [28]

    2022 , eprint=

    MineDojo: Building Open-Ended Embodied Agents with Internet-Scale Knowledge , author=. 2022 , eprint=

    2022

  21. [29]

    Deep RL Workshop NeurIPS 2021 , year=

    Benchmarking the Spectrum of Agent Capabilities , author=. Deep RL Workshop NeurIPS 2021 , year=

    2021

  22. [30]

    NeurIPS 2023 Foundation Models for Decision Making Workshop , year=

    Skill Reinforcement Learning and Planning for Open-World Long-Horizon Tasks , author=. NeurIPS 2023 Foundation Models for Decision Making Workshop , year=

    2023

  23. [31]

    Intrinsically-Motivated and Open-Ended Learning Workshop @NeurIPS2023 , year=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Intrinsically-Motivated and Open-Ended Learning Workshop @NeurIPS2023 , year=

  24. [32]

    2026 , eprint=

    D2E: Scaling Vision-Action Pretraining on Desktop Data for Transfer to Embodied AI , author=. 2026 , eprint=

    2026

  25. [33]

    2024 , eprint=

    Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion , author=. 2024 , eprint=

    2024

  26. [34]

    2026 , eprint=

    Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding , author=. 2026 , eprint=

    2026

  27. [35]

    2026 , eprint=

    DWM: Separating World Effects from Actions in Latent World Models , author=. 2026 , eprint=

    2026

  28. [36]

    2026 , note=

    Temporal Discriminative World Models , author=. 2026 , note=

    2026

  29. [37]

    2026 , eprint=

    Beyond the Next Step: Variable-Length Latent World Models for Long-Horizon Planning , author=. 2026 , eprint=

    2026

  30. [38]

    2026 , eprint=

    Co-Evolving Latent Action World Models , author=. 2026 , eprint=

    2026

  31. [39]

    2026 , eprint=

    Sensorimotor World Models: Perception for Action via Inverse Dynamics , author=. 2026 , eprint=

    2026

  32. [40]

    2025 , eprint=

    Learning Actionable World Models for Industrial Process Control , author=. 2025 , eprint=

    2025

  33. [41]

    2025 , eprint=

    Latent Action World Models for Control with Unlabeled Trajectories , author=. 2025 , eprint=

    2025

  34. [42]

    2026 , eprint=

    Open-World Reinforcement Learning over Long Short-Term Imagination , author=. 2026 , eprint=

    2026

  35. [43]

    Improving Transformer World Models for Data-Efficient

    Antoine Dedieu and Joseph Ortiz and Xinghua Lou and Carter Wendelken and J Swaroop Guntupalli and Wolfgang Lehrach and Miguel Lazaro-Gredilla and Kevin Patrick Murphy , booktitle=. Improving Transformer World Models for Data-Efficient. 2025 , url=

    2025

  36. [44]

    2026 , eprint=

    World2Act: Latent Action Post-Training from World Model Dynamics , author=. 2026 , eprint=

    2026

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.