← All AI/ML open problems

JEPA & world models · Embeddings & representations · RL & post-training / Open question

Learning a World Model Worth Planning With

A predictive representation is useful to an agent only if it retains the information needed to choose actions, not just the easiest features to forecast. Efficiency payoff: Learn compact state from experience that reduces real interactions while preserving decision-relevant information.

OpenEditorially scopedFramework needs definitionCompiled 2026-11-19

Seed catalog · Showing the bundled record.

Problem statement

Consider horizon-HH POMDPs with SS hidden states, AA actions and a finite observation alphabet, whose emission matrix has smallest singular value at least α>0\alpha>0. Tasks use rewards rθ(s,a)=θ⊤ψ(s,a)∈[0,1]r_\theta(s,a)=\theta^\top\psi(s,a)\in[0,1] for ∥θ∥2≤1\|\theta\|_2\leq1 and a shared kk-dimensional reward basis. Determine when action-conditioned joint-embedding prediction learns a history representation permitting sup⁡θ(Vθ⋆−Vθπ^θ)≤ε\sup_\theta(V^\star_\theta-V^{\hat\pi_\theta}_\theta)\leq\varepsilon, charging exploration, reward labels and auxiliary predictions rather than assuming the representation is Markov.

Assumptions and scope

  • Fix hidden dynamics, observation maps, task/reward family and exploration coverage; specify what history is encoded and whether the latent predictor is action-conditioned.
  • Distinguish low prediction error, noncollapsed embeddings and control sufficiency. Information irrelevant to prediction may still determine which action is good.
  • Charge reward labels, interventions and trajectory collection; define multi-step planning regret and model misspecification rather than assuming the learned state is Markov.

What would settle it?

A learnability and planning-regret theorem with a necessary/sufficient supervision or exploration condition in a restricted latent process, including examples where predictive training loses decision-relevant information.

Selected for a theoretical or computational target. A fitted scaling curve or improved benchmark score alone does not establish the stated learning-theoretic guarantee.

Status and evidence

Cui et al. already relate JEPA prediction error to planning regret through action-conditioned spectral structure. Controlled studies show predictive objectives can discard exogenous control-relevant features. The target extends to partially observed task families with explicit information-acquisition costs, not the first JEPA generalization bound. This is an editorial, source-motivated research target. The preliminary review through 2026-11-19 does not certify novelty or exclude an unlocated resolution.

Preliminary source check; not an expert-verified open status or an independent audit of proposed proofs. The compilation date is not a literature-review cutoff.

Sources and further reading

  1. Cui et al. — A Generalization Theory for JEPA-Based World Models ↗partial theory · source checked 2026-11-19

    Primary abstract checked: action-conditioned co-occurrence factorization, finite-sample prediction/planning-regret bounds and a latent-dimension trade-off. A prediction-to-planning theory is already available under the paper's framework.

  2. Pendharkar — Predictive Objectives Discard Exogenous Control-Relevant Features ↗empirical motivation · source checked 2026-11-19

    Primary abstract checked: controlled experiments separate predictability, controllability and reward relevance; reward grounding recovers omitted features in the tested environments. This is motivation, not a general sample-complexity theorem.

  3. Bhamidipaty et al. — Imperfect World Models are Exploitable ↗partial theory · source checked 2026-11-19

    Abstract checked: model exploitation, limits of reward-hacking analogies and a relaxed safe planning horizon are already studied.

  4. Hafner et al. — Mastering Diverse Domains through World Models ↗related algorithm · source checked 2026-11-20

    Primary abstract checked: DreamerV3 world-model learning from experience, including Minecraft without human data/curricula. This is an empirical baseline, not unrestricted efficient learnability.

  5. Ineffable Intelligence — mission, beliefs and A Note from Dave ↗research agenda · source checked 2026-11-20

    Primary site checked: experience-driven, continual learning without reliance on human data is the stated agenda. The site supplies no corresponding theorem; rewards, environment access and architectures are still inductive assumptions.