MoP-JEPA: Hard-Assigned Predictor Mixtures for Stochastic JEPA World Models

arXiv 2026 43.9 method

TLDR

MoP-JEPA uses hard-assigned predictor mixtures to avoid collapse in stochastic JEPA world models, achieving high planning success on OGBench.

Reasoning

The paper provides strong theoretical proofs and empirical results showing that deterministic predictors fail in stochastic environments, while MoP-JEPA's hard-assigned predictors quantize transitions effectively. However, the evaluation is limited to offline data and the verification protocol adds complexity.

Read-first score

Read-first score 43.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

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=baseline,evaluation,result

Topical relevance 29%

64.3

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: 335.

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

Hard-assigned predictor mixtures for JEPA world models to handle stochastic transitions
Proof that deterministic and gated mixture-of-experts predictors collapse to the conditional mean in stochastic environments
Proof that hard-assigned predictors converge to a quantizer of the transition distribution, with one head per successor mode
Verification protocol to prevent coverage freeloading in multi-prediction evaluation (input-agnostic codebook control, shuffled-context test, router-gated readouts, transition-precision guards, verified-route criterion)
Demonstration that planning over predicted modes significantly outperforms single-predictor rollouts and other predictors on OGBench tasks

Methodology

MoP-JEPA extends JEPA world models with a mixture of hard-assigned predictors. The paper theoretically proves the collapse of deterministic and gated MoE predictors in stochastic environments and the convergence of hard-assigned predictors to a quantizer. It evaluates planning success on OGBench offline data with leak-free evaluation, comparing against deterministic, gated-MoE, and variational predictors, and introduces a verification protocol to ensure fair multi-prediction evaluation.

Key Results

Planning over single-predictor rollouts achieves only 0.02-0.09 success, while MoP-JEPA's predicted modes reach up to 0.85, outperforming all baselines on every task. Under the verified-route criterion, MoP-JEPA outperforms the strongest soft alternative by 2-5x on all three mazes, and the same model places second of seven on the hardest maze in real execution.

Resources

Paper