Pith. sign in

REVIEW 3 major objections 5 minor 166 references

Agents that actively probe environments and distill only control-relevant latents learn world models that transfer across skills and new tasks.

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

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

T0 review · grok-4.5

2026-07-11 19:23 UTC pith:BUYAGZ6J

load-bearing objection Solid closed-loop recipe for task-sufficient latents with real multi-suite gains; formal MIST claim is oversold relative to the surrogate objectives. the 3 major comments →

arxiv 2607.04409 v1 pith:BUYAGZ6J submitted 2026-07-05 cs.LG

Learning Task-Sufficient World Models by Synergizing Agentic Exploration and Structured Modeling

Fan Feng , Yujia Zheng , Minghao Fu , Yongqiang Chen , Guangyi Chen , Kevin Murphy , Biwei Huang , Kun Zhang This is my paper
classification cs.LG
keywords world modelsmodel-based RLtask-sufficient representationsminimal sufficient statesagentic explorationskill discoveryadaptive curriculumrobotic 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.

Standard world models keep too many visual or latent factors that do not matter for control, which hurts sample efficiency and transfer. This paper argues that agents can instead recover task-specific, minimal, and sufficient latent states—called MIST states—by closing a loop between purposeful exploration and structured model learning. The agent actively probes each new environment with skills chosen to expose latent factors that affect observations and rewards, under a curriculum that prioritizes hard tasks. From that data the world model learns a soft mask that keeps only the coordinates needed to predict rewards while discarding the rest. Policies that act only on these compact states reuse control-relevant factors and generalize to new skill combinations and previously unseen tasks on continuous-control and robotic-manipulation benchmarks.

Core claim

A closed-loop synergy of agentic active probing (under an adaptive curriculum) and structured world-model learning recovers task-specific minimal sufficient (MIST) latent representations that capture all control-relevant factors; policies conditioned on those representations achieve improved sample efficiency and generalization across skills, object–skill compositions, and previously unseen tasks.

What carries the argument

MIST states: the coordinates of a Dreamer-style latent vector selected by a soft mask that maximizes reward likelihood (sufficiency) while maximizing mutual information with a task descriptor and penalizing L1 mass (minimality); the data that train this mask come from skills chosen by InfoNCE segment separability under an adaptive curriculum.

Load-bearing premise

That the practical mask-and-mutual-information objectives, plus contrastive skill selection, actually recover the formal minimal set of reward-relevant latent factors rather than a merely useful but incomplete subspace.

What would settle it

In a simulator where ground-truth factored states are known, train MIST-WM and check whether linear probes of the masked latents recover every ground-truth parent of the reward (and its one-step parents) with high R-squared while discarding non-intervenable factors; if key reward parents remain unaligned or non-intervenable factors are retained, the recovery claim fails.

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

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes MIST-WM, a closed-loop framework that couples agentic exploration with structured world-model learning to obtain task-specific, minimal, and sufficient latent states for control. On the agent side, a skill library (DIAYN/METRA) is refined by an InfoNCE segment-separability objective and an adaptive UED-style curriculum to collect informative trajectories. On the model side, Dreamer-v3 latents are gated by a soft mask optimized for reward likelihood (sufficiency) and mutual information with a task descriptor minus an L1 penalty (minimality), with an expandable latent dimension across a task sequence. The authors evaluate representation recovery via linear probes to simulator factors on RoboSuite, single-task sample efficiency on Meta-World and DMControl, and skill/compositional/unseen generalization on RoboSuite, Franka-Kitchen, and Meta-World, with multi-seed results and ablations over exploration, structure terms, backbone, and curriculum.

Significance. If the empirical claims hold, the work is a useful contribution to model-based RL: it targets task-sufficient rather than generic or fully factored latents, and shows that closed-loop data curation plus structure-aware masking can improve sample efficiency and transfer on standard continuous-control and manipulation suites. Strengths include multi-benchmark evaluation (DMControl, Meta-World, RoboSuite, Franka-Kitchen), five-seed reporting, direct R² probing against ground-truth factors, and ablations that isolate exploration, mask/MI/sufficiency terms, backbone, and curriculum. The formal MIST definition (DBN parents of reward and one-step parents) is a clear conceptual target even if the learning objectives are surrogates. The package is of interest to the world-model and unsupervised-skill communities, provided claims about recovering all control-relevant factors are calibrated to the evidence.

major comments (3)
  1. [Abstract; Sec. 3.1, Eqs. (2)–(3); Def. 1; App. B.1] Abstract and Sec. 1 claim empirical recovery of task-sufficient latents that 'capture all control-relevant factors,' with MIST formally defined via DBN parents of reward and their one-step parents (Def. 1; App. B.1 Defs. 2–3). Sec. 3.1 instead optimizes practical surrogates—reward likelihood under a soft mask (Eq. 2) and MI(masked states, gi) − λ_M‖m‖1 (Eq. 3)—and explicitly calls this a 'practical surrogate' for a 'MIST-aligned subspace rather than exact latent DBN variables.' These objectives do not enforce the conditional independencies of Defs. 2–3, do not target one-step dynamics parents of reward parents, and depend on gi (RoBERTa embeddings when available, else rewards). The central claim should be restated as recovery of a useful task-aligned subspace, or the paper should add a direct residual-sufficiency check (e.g., reward prediction error with vs. without discarded coordinates
  2. [Sec. 5, RQ1; Fig. 4; Table A9] Fig. 4 and Table A9 show incomplete factor recovery on the controlled RoboSuite stack: gripper openness R²≈0.21 and size≈0, while cube mass/position/orientation and gripper pose are high. The text notes that static or fine-grained factors are hard to identify with contrastive pairs, yet still asserts that the subspace 'retains what matters for the task.' For the 'all control-relevant factors' claim, the paper should either (i) demonstrate that residual unprobed factors are irrelevant for return under the evaluated policies (e.g., policy performance when those coordinates are ablated or when ground-truth residual factors are added), or (ii) qualify the claim to the factors that are intervenable and reward-linked under the probing design. Incomplete recovery is not fatal to the empirical story, but it currently undercuts the strongest wording.
  3. [Sec. 3.2; App. D.2] The adaptive curriculum and active-probing design rely on environment scaffolding that places objects with controlled factor differences (App. D.2: multi-cube mass/friction/orientation variants; multi-object Meta-World/Kitchen setups). Baselines receive the same phase-one budget, which is fair for comparison, but the method’s ability to expose latent factors may depend on this design. The paper should clarify how much of the representation recovery is attributable to the probing objective (Eq. 5) versus the curated multi-object scenes, and discuss applicability when such scaffolding is unavailable (e.g., single-object or real-world settings without factor-controlled variants).
minor comments (5)
  1. [Fig. 2–3; Table A2] Fig. 2(b) and Fig. 3 are dense; a short caption walkthrough of which modules are frozen vs. updated (cf. Table A2) would help readers follow the closed loop.
  2. [Sec. 2, Def. 1] Notation for MIST indices mixes I^{(t)}_{i,1}, I^{(t)}_{i,2}, and U_i; a single consistent symbol for the selected subspace S^{min}_i in the main text would reduce load.
  3. [Sec. 5, RQ2; Fig. 5] Reacher-Hard is the main single-task underperformance vs. TD-MPC2 (Fig. 5, Table A11); a brief discussion of when structure-aware Dreamer objectives lose to model-free control objectives would strengthen the analysis.
  4. [Sec. 6; App. A] App. A limitations correctly flag simulation-only evaluation and scaling to diffusion world models; a sentence in the main conclusion pointing to these would set expectations for practitioners.
  5. [Abstract; Fig. 7] Typos/consistency: 'task-specific, minimal, andsufficient' spacing in abstract; 'w/o su ciency' in Fig. 7; occasional Dreamer-V3 vs Dreamer-v3 capitalization.

Circularity Check

0 steps flagged

No circular derivation: formal MIST is defined independently of the learning objectives; claims are empirical recovery via surrogates and external benchmarks, not results forced by construction or self-citation.

full rationale

The paper’s load-bearing chain is methodological and empirical, not a first-principles derivation that collapses into its inputs. Definition 1 and Appendix B.1 define MIST indices from DBN parents of reward and their one-step parents via conditional independence; Section 3.1 then introduces separate practical surrogates (reward likelihood under a soft mask, Eq. 2; MI of masked states with task descriptor minus L1, Eq. 3; InfoNCE segment separability, Eq. 5) and explicitly labels them a “practical surrogate … rather than exact latent DBN variables.” That gap is an approximation/correctness issue, not circularity: the formal object is not defined as the surrogate optimum, and the paper does not claim the surrogate equals Def. 1 by construction. Downstream claims (sample efficiency, skill/compositional/unseen generalization) are evaluated on external suites (DMControl, Meta-World, RoboSuite, Franka-Kitchen) against external baselines (Dreamer-V3, DINO-WM, TD-MPC2, METRA, Plan2Explore, etc.), with representation quality checked by linear probes to simulator ground truth. Related-work citations to prior factored/causal RL (including some author overlap) position the contribution; they are not uniqueness theorems or load-bearing premises that force the empirical outcomes. Task-descriptor MI and curriculum proxies are modeling choices, not fitted parameters renamed as predictions of the same quantity. No self-definitional loop, no fitted-input-as-prediction, no uniqueness imported from the authors, and no renaming of a known closed-form result. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The central empirical claim rests on standard MBRL/POMDP machinery plus several paper-specific modeling choices and many tuned coefficients. Formal MIST is defined cleanly from DBN parents; almost all experimental success depends on surrogate losses, skill libraries, curriculum proxies, and expandable latent heuristics rather than identification theorems.

free parameters (7)
  • minimality L1 coefficient λ_M / minimal-score coef = 1e-3 / 1e-2 by suite
    Balances MI vs sparsity in the mask objective; set to 1e-3 (DMControl) or 1e-2 (others).
  • sufficiency score coefficient = 1e-3 / 1e-2 by suite
    Weights reward-likelihood term used to train the mask.
  • latent expansion increment Δd = 32
    Heuristic growth of Dreamer latent dim when a new task arrives; not identified from data.
  • initial latent dimensionality d1 = 512 or 256
    Suite-specific overparameterized start (512 visual suites, 256 DMControl).
  • MI annealing schedule (α_start, α_end) and sparsity target schedule = cosine anneal; sparsity target 0.35 at adaptation
    Cosine schedules control when MI and gate sparsity become active; chosen for training stability.
  • curriculum temperature η, domain-randomization prob p_DR, ensemble size M = η=0.1, p_DR=0.3, M=5
    UED-style environment selection hyperparameters.
  • skill/contrastive and RL hyperparameters (lr, entropy coef, network sizes, steps) = see Tables A4–A7
    Large set of training knobs for MISL, InfoNCE, SAC/PPO, and world-model training budgets.
axioms (6)
  • domain assumption Environment is a factored POMDP with expandable latent state factors and a task-specific DBN whose reward parents define control-relevant structure.
    Sec. 2 formal setup; without factorization/expandability the MIST index set is not well-defined as stated.
  • ad hoc to paper Soft-mask sufficiency (reward likelihood) and minimality (MI with task descriptor − L1) are adequate surrogates for formal MIST indices.
    Sec. 3.1 explicitly labels this a practical surrogate for DBN-defined MIST states.
  • ad hoc to paper Skills that maximize InfoNCE separability of observation segments under different latent-factor values yield data informative enough to identify task-relevant factors.
    Sec. 3.2 information-gain construction; no identification guarantee.
  • domain assumption Task descriptors (RoBERTa-encoded instructions when available, else rewards) are available at train and test and carry the right task identity for MI.
    Sec. 3.1; performance of minimality depends on this side information.
  • domain assumption Dreamer-v3 RSSM latents form a usable unstructured basis that can be frozen while masks and skills are learned, then expanded by fine-tuning.
    Sec. 3.1 and Appendix C.4–C.5 backbone assumptions.
  • standard math Standard variational world-model, SAC/PPO, MINE, and MISL training theory apply as used.
    Background optimization tools imported from cited literature.
invented entities (3)
  • MIST states / MIST index set U_i no independent evidence
    purpose: Name the minimal sufficient task-specific latent coordinates (reward parents and their one-step parents) used for policy and reward decoding.
    Definition 1; operationalized via learned masks rather than recovered DBN nodes.
  • MIST-WM closed-loop framework no independent evidence
    purpose: Couple active probing skill library + adaptive curriculum with structured masking world-model learning.
    Core proposed system (Fig. 1, Sec. 3); evaluated only inside this paper's experiments.
  • Active probing skills selected by segment separability no independent evidence
    purpose: Collect trajectories that expose latent factors for structure learning.
    Sec. 3.2; built on MISL but specialized via InfoNCE on latent-factor contrasts.

pith-pipeline@v1.1.0-grok45 · 32054 in / 4097 out tokens · 46635 ms · 2026-07-11T19:23:01.188028+00:00 · methodology

0 comments
read the original abstract

Learning and planning in imagination using world models provides an effective paradigm for training agents for decision-making. However, existing approaches often rely on high-dimensional latent spaces or generic visual embeddings that retain many factors irrelevant to control, limiting efficiency and generalization across tasks. To this end, we study how agents can learn world models with representations that are task-specific, minimal, and sufficient for decision-making. We achieve this via a closed-loop synergy between the agent and the world model, in which structured world-model learning distills task-sufficient representations from informative interaction data. On the agent side, agents actively probe the environment to collect informative trajectories that expose task-relevant latent factors, guided by an adaptive curriculum. On the world-model side, we learn structured representations over observations to distill compact, task-sufficient latent states from the collected interaction data. This synergy enables the empirical recovery of task-sufficient latent representations that capture all control-relevant factors. Leveraging these representations, the resulting policies achieve improved sample efficiency and generalization, including generalization across skills, object-skill compositions, and previously unseen tasks on standard continuous-control and robotic-manipulation benchmarks.

Figures

Figures reproduced from arXiv: 2607.04409 by Biwei Huang, Fan Feng, Guangyi Chen, Kevin Murphy, Kun Zhang, Minghao Fu, Yongqiang Chen, Yujia Zheng.

Figure 1
Figure 1. Figure 1: Overview on MIST-WM: co-evolve on the agentic ex￾ploration and structured world models. The agent conducts active interventions to curate meaningful behavior data to identify the key latent factors in each new environment under an adaptive cur￾riculum. With the curated data, a structured world model with minimal, sufficient, and task-specific representation is learned. a closed-loop synergy centered on lea… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Example DBNs that describe the generative process of two tasks. The grey and white nodes indicate observable and latent variables. The blue nodes are the newly expanded variables for Task 2. Red edges are the links from state variables to rewards. For simplicity, we omit the factorization of the state–observation mapping s → o. (b) Computational flow of our proposed framework compared with the Dreamer-… view at source ↗
Figure 3
Figure 3. Figure 3: Method overview. On the agent side, for each new environment, the agent selects skills from its library to perform active probing, collects informative data, under an adaptive curriculum. Conditioned on these informative data, the structured world model is learned to distill the MIST states. For the minimality score, we encourage the mask to retain only the necessary state dimensions by maximizing the mu￾t… view at source ↗
Figure 5
Figure 5. Figure 5: Results (across 5 seeds) on single-task learning, including robotic manipulation tasks (Meta World) and locomotion (DMControl). Shaded areas indicate the standard errors. Lift Pick & Place Stacking Door-open Handle-press Light Slide MT-1 MT-2 Coffee-push Drawer-open Button-press Drawer-close Door-unlock MIST-WM Factored-Dreamer DINO-WM Dreamer-v3 Lift Pick & Place Stacking Door-open Handle-press Light Slid… view at source ↗
Figure 6
Figure 6. Figure 6: Success rate (across 5 seeds) on skill and compositional generalization, including tasks in Stacking (Robosuite), MT-1 & 2 (kitchen), and Door-open, Door-close, Coffee-push (Meta-World). Error bars indicate the standard errors. Curiosity Replay METRA Random Kitchen Meta-World Skill (Kitchen) Com (MW) Unseen (MW) w/o MI w/o su ciency w/o masking ` w/o curriculum Normalized Scores w/ DINO w/ R3M [PITH_FULL_… view at source ↗
Figure 7
Figure 7. Figure 7: Ablation studies. We consider studies on exploration policy (Row 1); World model terms (Row 2); and Backbones and curriculum (Row 3). Pink bars are original MIST-WM and error bars indicate the standard errors. whereas for Meta-World and RoboSuite, they use Soft Actor￾Critic (SAC, (Haarnoja et al., 2018)). Results are in [PITH_FULL_IMAGE:figures/full_fig_p008_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

166 extracted references · 37 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2402.05052 , year=

    Causal representation learning from multiple distributions: A general setting , author=. arXiv preprint arXiv:2402.05052 , year=

    Pith/arXiv arXiv

  2. [2]

    International conference on machine learning , pages=

    Interventional causal representation learning , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  3. [4]

    Nature Communications , volume=

    Recurrent neural networks with explicit representation of dynamic latent variables can mimic behavioral patterns in a physical inference task , author=. Nature Communications , volume=. 2022 , publisher=

    2022

  4. [5]

    International Conference on Machine Learning , pages=

    Ready policy one: World building through active learning , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  5. [6]

    Advances in Neural Information Processing Systems , volume=

    Discovering and achieving goals via world models , author=. Advances in Neural Information Processing Systems , volume=

  6. [7]

    Nature , volume=

    People construct simplified mental representations to plan , author=. Nature , volume=. 2022 , publisher=

    2022

  7. [8]

    Neuron , volume=

    Linking connectivity, dynamics, and computations in low-rank recurrent neural networks , author=. Neuron , volume=. 2018 , publisher=

    2018

  8. [9]

    Advances in Neural Information Processing Systems , volume=

    Neural foundations of mental simulation: Future prediction of latent representations on dynamic scenes , author=. Advances in Neural Information Processing Systems , volume=

  9. [10]

    Proceedings of the Eighth international joint conference on Artificial intelligence-Volume 2 , pages=

    Planning using a temporal world model , author=. Proceedings of the Eighth international joint conference on Artificial intelligence-Volume 2 , pages=

  10. [11]

    ACM Sigart Bulletin , volume=

    Dyna, an integrated architecture for learning, planning, and reacting , author=. ACM Sigart Bulletin , volume=. 1991 , publisher=

    1991

  11. [12]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  12. [13]

    Artificial intelligence , volume=

    Between MDPs and semi-MDPs: A framework for temporal abstraction in reinforcement learning , author=. Artificial intelligence , volume=. 1999 , publisher=

    1999

  13. [14]

    International conference on machine learning , pages=

    Mutual information neural estimation , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  14. [15]

    Advances in Neural Information Processing Systems , volume=

    Identifying selections for unsupervised subtask discovery , author=. Advances in Neural Information Processing Systems , volume=

  15. [17]

    arXiv preprint arXiv:2301.04104 , year=

    Mastering diverse domains through world models , author=. arXiv preprint arXiv:2301.04104 , year=

    Pith/arXiv arXiv

  16. [18]

    arXiv preprint arXiv:2402.16354 , year=

    Language-guided skill learning with temporal variational inference , author=. arXiv preprint arXiv:2402.16354 , year=

    Pith/arXiv arXiv

  17. [19]

    Advances in Neural Information Processing Systems , volume=

    Disentangled unsupervised skill discovery for efficient hierarchical reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  18. [20]

    International conference on machine learning , pages=

    Causal curiosity: Rl agents discovering self-supervised experiments for causal representation learning , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  19. [21]

    International Conference on Machine Learning , pages=

    Citris: Causal identifiability from temporal intervened sequences , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  20. [22]

    Uncertainty in Artificial Intelligence , pages=

    Biscuit: Causal representation learning from binary interactions , author=. Uncertainty in Artificial Intelligence , pages=. 2023 , organization=

    2023

  21. [23]

    Advances in Neural Information Processing Systems , volume=

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence , author=. Advances in Neural Information Processing Systems , volume=

  22. [24]

    arXiv preprint arXiv:2410.22325 , year=

    Robots Pre-train Robots: Manipulation-Centric Robotic Representation from Large-Scale Robot Datasets , author=. arXiv preprint arXiv:2410.22325 , year=

    Pith/arXiv arXiv

  23. [25]

    arXiv preprint arXiv:2310.06824 , year=

    The geometry of truth: Emergent linear structure in large language model representations of true/false datasets , author=. arXiv preprint arXiv:2310.06824 , year=

    Pith/arXiv arXiv

  24. [26]

    arXiv preprint arXiv:2311.03658 , year=

    The linear representation hypothesis and the geometry of large language models , author=. arXiv preprint arXiv:2311.03658 , year=

    Pith/arXiv arXiv

  25. [27]

    Advances in Neural Information Processing Systems , volume=

    Reft: Representation finetuning for language models , author=. Advances in Neural Information Processing Systems , volume=

  26. [28]

    arXiv preprint arXiv:2502.12179 , year=

    Identifiable Steering via Sparse Autoencoding of Multi-Concept Shifts , author=. arXiv preprint arXiv:2502.12179 , year=

    arXiv

  27. [30]

    International conference on artificial intelligence and statistics , pages=

    Variational autoencoders and nonlinear ica: A unifying framework , author=. International conference on artificial intelligence and statistics , pages=. 2020 , organization=

    2020

  28. [31]

    arXiv preprint arXiv:2402.09236 , year=

    Learning interpretable concepts: Unifying causal representation learning and foundation models , author=. arXiv preprint arXiv:2402.09236 , year=

    Pith/arXiv arXiv

  29. [32]

    UAI 2022 Workshop on Causal Representation Learning , year=

    Intervention design for causal representation learning , author=. UAI 2022 Workshop on Causal Representation Learning , year=

    2022

  30. [34]

    ICLR , year=

    Cross-Entropy Is All You Need To Invert the Data Generating Process , author=. ICLR , year=

  31. [35]

    Advances in Neural Information Processing Systems , volume=

    Optimistic active exploration of dynamical systems , author=. Advances in Neural Information Processing Systems , volume=

  32. [36]

    Artificial intelligence , volume=

    Planning and acting in partially observable stochastic domains , author=. Artificial intelligence , volume=. 1998 , publisher=

    1998

  33. [37]

    arXiv preprint arXiv:2506.06658 , year=

    Self-Adapting Improvement Loops for Robotic Learning , author=. arXiv preprint arXiv:2506.06658 , year=

    arXiv

  34. [38]

    Advances in Neural Information Processing Systems , volume=

    Text-aware diffusion for policy learning , author=. Advances in Neural Information Processing Systems , volume=

  35. [39]

    Neural computation , volume=

    Improving generalization for temporal difference learning: The successor representation , author=. Neural computation , volume=. 1993 , publisher=

    1993

  36. [40]

    arXiv preprint arXiv:2504.15369 , year=

    Solving new tasks by adapting internet video knowledge , author=. arXiv preprint arXiv:2504.15369 , year=

    Pith/arXiv arXiv

  37. [41]

    Nature , pages=

    Mastering diverse control tasks through world models , author=. Nature , pages=. 2025 , publisher=

    2025

  38. [42]

    International Conference on Machine Learning, PMLR , year=

    Temporal Difference Learning for Model Predictive Control , author=. International Conference on Machine Learning, PMLR , year=

  39. [43]

    Advances in neural information processing systems , volume=

    Mastering atari games with limited data , author=. Advances in neural information processing systems , volume=

  40. [44]

    International Conference on Learning Representations , year=

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

  41. [45]

    International Conference on Learning Representations , year=

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

  42. [46]

    Nature , volume=

    Mastering atari, go, chess and shogi by planning with a learned model , author=. Nature , volume=. 2020 , publisher=

    2020

  43. [47]

    2024 , url=

    Nicklas Hansen and Hao Su and Xiaolong Wang , booktitle=. 2024 , url=

    2024

  44. [49]

    Forty-second International Conference on Machine Learning , year=

    Founder: Grounding foundation models in world models for open-ended embodied decision making , author=. Forty-second International Conference on Machine Learning , year=

  45. [50]

    2025 , url=

    Gaoyue Zhou and Hengkai Pan and Yann LeCun and Lerrel Pinto , booktitle=. 2025 , url=

    2025

  46. [52]

    International Conference on Machine Learning , pages=

    Denoised MDPs: Learning World Models Better Than the World Itself , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  47. [53]

    Advances in Neural Information Processing Systems , volume=

    Learning world models with identifiable factorization , author=. Advances in Neural Information Processing Systems , volume=

  48. [54]

    arXiv preprint arXiv:2306.12672 , year=

    From word models to world models: Translating from natural language to the probabilistic language of thought , author=. arXiv preprint arXiv:2306.12672 , year=

    Pith/arXiv arXiv

  49. [55]

    Artificial Life Conference Proceedings , pages=

    Identifying necessary conditions for open-ended evolution through the artificial life world of chromaria , author=. Artificial Life Conference Proceedings , pages=. 2014 , organization=

    2014

  50. [56]

    arXiv preprint arXiv:2107.12808 , year=

    Open-ended learning leads to generally capable agents , author=. arXiv preprint arXiv:2107.12808 , year=

    Pith/arXiv arXiv

  51. [57]

    arXiv preprint arXiv:2502.08623 , year=

    Robot data curation with mutual information estimators , author=. arXiv preprint arXiv:2502.08623 , year=

    Pith/arXiv arXiv

  52. [58]

    arXiv preprint arXiv:2506.19121 , year=

    CUPID: Curating Data your Robot Loves with Influence Functions , author=. arXiv preprint arXiv:2506.19121 , year=

    arXiv

  53. [59]

    arXiv preprint arXiv:2403.05110 , year=

    Efficient data collection for robotic manipulation via compositional generalization , author=. arXiv preprint arXiv:2403.05110 , year=

    Pith/arXiv arXiv

  54. [60]

    The Twelfth International Conference on Learning Representations , year=

    Reward-Free Curricula for Training Robust World Models , author=. The Twelfth International Conference on Learning Representations , year=

  55. [61]

    Advances in neural information processing systems , volume=

    Emergent complexity and zero-shot transfer via unsupervised environment design , author=. Advances in neural information processing systems , volume=

  56. [62]

    Advances in Neural Information Processing Systems , volume=

    Improving environment novelty quantification for effective unsupervised environment design , author=. Advances in Neural Information Processing Systems , volume=

  57. [63]

    Advances in Neural Information Processing Systems , volume=

    Cqm: Curriculum reinforcement learning with a quantized world model , author=. Advances in Neural Information Processing Systems , volume=

  58. [64]

    Probabilistic Graphical Models, M

    Dynamic bayesian networks , author=. Probabilistic Graphical Models, M. Jordan , volume=

  59. [65]

    Artificial intelligence , volume=

    Stochastic dynamic programming with factored representations , author=. Artificial intelligence , volume=. 2000 , publisher=

    2000

  60. [66]

    2009 , publisher=

    Causality , author=. 2009 , publisher=

    2009

  61. [67]

    International Conference on Learning Representations , year=

    Diversity is All You Need: Learning Skills without a Reward Function , author=. International Conference on Learning Representations , year=

  62. [68]

    Chongyi Zheng and Jens Tuyls and Joanne Peng and Benjamin Eysenbach , booktitle=. Can a. 2025 , url=

    2025

  63. [70]

    International Conference on Learning Representations , year=

    Categorical Reparameterization with Gumbel-Softmax , author=. International Conference on Learning Representations , year=

  64. [72]

    Cognition , volume=

    The double-edged sword of pedagogy: Instruction limits spontaneous exploration and discovery , author=. Cognition , volume=. 2011 , publisher=

    2011

  65. [73]

    , author=

    Word, thought, and deed: the role of object categories in children's inductive inferences and exploratory play. , author=. Developmental Psychology , volume=. 2008 , publisher=

    2008

  66. [74]

    International conference on machine learning , pages=

    Planning to explore via self-supervised world models , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  67. [75]

    arXiv preprint arXiv:2310.06114 , volume=

    Learning interactive real-world simulators , author=. arXiv preprint arXiv:2310.06114 , volume=

    Pith/arXiv arXiv

  68. [78]

    2025 , eprint=

    Ctrl-World: A Controllable Generative World Model for Robot Manipulation , author=. 2025 , eprint=

    2025

  69. [79]

    Proceedings of the Sixteenth conference on Uncertainty in artificial intelligence , pages=

    Dynamic Bayesian multinets , author=. Proceedings of the Sixteenth conference on Uncertainty in artificial intelligence , pages=

  70. [80]

    2024 , url=

    Seohong Park and Oleh Rybkin and Sergey Levine , booktitle=. 2024 , url=

    2024

  71. [81]

    2009 , publisher=

    Directional statistics , author=. 2009 , publisher=

    2009

  72. [83]

    International Conference on Machine Learning , pages=

    Curious Replay for Model-based Adaptation , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  73. [84]

    Conference on robot learning , pages=

    Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning , author=. Conference on robot learning , pages=. 2020 , organization=

    2020

  74. [85]

    , author=

    Metrics for Finite Markov Decision Processes. , author=. UAI , volume=

  75. [86]

    Software Impacts , volume=

    dm\_control: Software and tasks for continuous control , author=. Software Impacts , volume=. 2020 , publisher=

    2020

  76. [88]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Scalable methods for computing state similarity in deterministic markov decision processes , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  77. [89]

    International Conference on Learning Representations , year=

    Learning Invariant Representations for Reinforcement Learning without Reconstruction , author=. International Conference on Learning Representations , year=

  78. [93]

    Advances in Neural Information Processing Systems , volume=

    Factored adaptation for non-stationary reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  79. [96]

    International conference on machine learning , pages=

    Curl: Contrastive unsupervised representations for reinforcement learning , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  80. [97]

    International Conference on Machine Learning , pages=

    Learning task informed abstractions , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

Showing first 80 references.