Anisotropic Representations Improve Planning in JEPA World Models

arXiv 2026 36 method

TLDR

Introduces AnisoWM with ΛReg, a learnable diagonal covariance target that improves planning in JEPA world models by aligning latent representation geometry with task costs.

Reasoning

The paper provides a focused theoretical analysis of how representation geometry affects planning costs and supports it with empirical gains across four visual control environments. However, it relies only on simulated control benchmarks and does not demonstrate real-world deployment or compare with broader generative world model approaches.

Read-first score

Read-first score 36, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,metric,result

Topical relevance 25%

55.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=github

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 352.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

world simulator

generative world model

video world model

Deep Analysis

Innovations

Identifies that isotropic Gaussian regularization can produce task-misaligned latent planning geometry even with accurate prediction and non-collapsed representations.
Introduces AnisoWM with ΛReg, replacing the fixed isotropic Gaussian target with a learnable diagonal covariance under fixed-trace and anisotropy constraints.
Provides analysis of prediction-driven allocation of target variance and conditions under which the induced metric reduces planning regret.

Methodology

The paper proposes AnisoWM, a latent world model that uses ΛReg with a learnable diagonal covariance target under fixed-trace and anisotropy constraints, while keeping the prediction objective, predictor architecture, and Euclidean planner unchanged. It includes theoretical analysis of target variance allocation and planning regret, and evaluates planning success and latent cost-task agreement across four visual control environments compared to LeWorldModel.

Key Results

AnisoWM improves planning success over LeWorldModel in all four visual control environments, and its latent planning cost shows better agreement with task outcomes.

Resources

Paper