Delta-JEPA: Learning Action-Sensitive World Models via Latent Difference Decoding

arXiv 2026 43 method

TLDR

Delta-JEPA learns action-sensitive world models by decoding actions from latent differences, avoiding pixel reconstruction and improving planning in continuous control.

Reasoning

The paper introduces a novel method to prevent action-insensitive representations in world models by supervising latent displacements, which is simple and effective. However, it is only evaluated on simulated continuous-control tasks, lacking real-world validation.

Read-first score

Read-first score 43, weighted from topical fit, citation, graph, method, reproducibility, and recency signals.

Recency 6%

100

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

Methodology quality 18%

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

Topical relevance 29%

61.4

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 18%

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

Citation impact 18%

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

Citation velocity 12%

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: 330.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

interactive world model

video world model

world simulator

generative world model

Deep Analysis

Innovations

Latent Difference Action Decoder (LDAD) that reconstructs actions from latent displacement between consecutive observations, rather than from concatenated endpoint embeddings.
Displacement-level supervision that regularizes transition geometry to prevent representation collapse and encourage action-distinguishable latent changes.
Reconstruction-free world model (Delta-JEPA) that uses only latent prediction and action reconstruction, avoiding pixel reconstruction and distribution-matching regularizers.

Methodology

Delta-JEPA is an end-to-end reconstruction-free world model that combines latent forward prediction with a Latent Difference Action Decoder (LDAD) trained to reconstruct actions from the latent displacement between consecutive observations. It is trained solely with latent prediction and action reconstruction losses, without pixel reconstruction or distribution matching. The model is evaluated on four visual continuous-control tasks against JEPA-based and representation-learning world model baselines, with ablations comparing displacement-based versus endpoint concatenation decoding.

Key Results

Delta-JEPA improves planning performance over JEPA-based and representation-learning world model baselines across four visual continuous-control tasks. Displacement-based action decoding is consistently more effective than endpoint concatenation, and action-sensitivity analyses show clearer action-conditioned latent responses.

Resources

Paper