seq-JEPA: Autoregressive Predictive Learning of Invariant-Equivariant World Models

arXiv 25.5 2025 47 method

TLDR

seq-JEPA learns separate invariant and equivariant representations via autoregressive predictive learning of world models from sequences of views and actions.

Reasoning

The paper introduces a novel architectural inductive bias to resolve the invariance-equivariance trade-off in SSL, which is a strength. However, the abstract is cut off, limiting full assessment of empirical results and potential limitations such as domain generality.

Read-first score

Read-first score 47, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

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

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=none

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 421.

Keyword Scores

world model

world dynamics prediction

interactive world model

generative world model

model-based reinforcement learning world model

video world model

world simulator

Deep Analysis

Innovations

Resolves the invariance-equivariance trade-off in joint-embedding SSL without relying on dual predictors or loss terms
Introduces architectural inductive biases to learn separate representations for invariance and equivariance
Processes sequences of views with action embeddings through a transformer encoder for autoregressive predictive learning
Conditions the predictor head on upcoming actions to predict next observation representations

Methodology

seq-JEPA processes short sequences of different views of inputs, where each view is concatenated with an embedding of the relative transformation (action) that produces the next observation. These view-action pairs are passed through a transformer encoder to output an aggregate representation, and a predictor head conditions this aggregate representation on the upcoming action to predict the representation of the next observation. The model is trained via autoregressive predictive learning and evaluated on downstream tasks requiring invariance (e.g., classification) and equivariance (e.g., path integration).

Key Results

seq-JEPA achieves strong performance on both invariance- and equivariance-demanding downstream tasks without sacrificing one for the other, and excels at tasks requiring sequence aggregation such as path integration across actions and predictive learning across eye movements.

Resources

Paper