AVL-JEPA: Preventing Causal Dynamics Information Collapse In Joint Embedding Predictive Architecture World Models

arXiv 2026 31.5 method

TLDR

AVL-JEPA prevents causal dynamics information collapse in JEPA world models by adding action and vision invariance pathways, improving control under perturbations.

Reasoning

The paper identifies a clear failure mode and proposes a method with supporting ablations on four control tasks. However, the abstract lacks baseline details and statistical rigor, and the tasks may be simulated rather than physically real-world.

Read-first score

Read-first score 31.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.

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=evaluation,result

Topical relevance 25%

48.6

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=none

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

Frontier

Rank sensitivity

Stability: volatile; rank range: 232.

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 causal dynamics information collapse as a failure mode in joint embedding predictive architectures, where visual information is preserved but action-relevant physical consequence information is discarded.
Proposes action-grounded vision-invariance latent (AVL) to prevent causal dynamics information collapse.
Uses executed actions as auxiliary dynamics anchors to encourage preservation of dynamics information.
Introduces a vision-invariance pathway that aligns perturbed and clean latent predictions without discarding dynamics information.
Evaluates preservation of causal dynamics through physical consequence alignment, clean-noisy dynamics consistency, and targeted transition subspace erasure.

Methodology

AVL extends joint embedding predictive architecture world models with an action-grounded auxiliary dynamics anchor and a vision-invariance pathway that aligns perturbed and clean latent predictions. The method is evaluated on four robotic control tasks: TwoRoom, PushT, OGBench Cube, and Reacher. Evaluation includes success rates under visual perturbations, clean-environment performance, physical consequence alignment, clean-noisy dynamics consistency, and causal effect of targeted transition subspace erasure.

Key Results

AVL substantially improves success rates under visual perturbations while preserving clean-environment performance on four robotic control tasks. Further evaluations show improved physical consequence alignment, clean-noisy dynamics consistency, and causal evidence that dynamics information relevant to planning is preserved from collapse.

Resources

Paper