⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

JEDI: Latent End-to-end Diffusion Mitigates Agent-Human Performance Asymmetry in Model-Based Reinforcement Learning

arXiv 25.5 2025 70.6 method, benchmark

TLDR

Proposes JEDI, a latent diffusion world model to address performance asymmetry in MBRL on Atari100k, achieving balanced human-normalized scores.

Reasoning

Strengths: Identifies a novel performance asymmetry issue in MBRL and proposes a method to mitigate it. Weaknesses: Limited to Atari100k benchmark; generalization unclear.

Read-first score

Read-first score 70.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 56.

Methodology quality 25%
90

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

Recency 8%
86.7

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

Topical relevance 42%
80

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 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 88.

Keyword Scores

world model
10
model-based reinforcement learning world model
10
generative world model
9
world dynamics prediction
8
interactive world model
7
world simulator
6
video world model
6

Deep Analysis

Innovations

  • Identification and formalization of Performance Asymmetry in MBRL on Atari100k, where agents outperform humans on some tasks and underperform on others
  • Introduction of Sym-HNS, a more balanced aggregate metric that evenly weights Human-Optimal and Agent-Optimal subsets
  • Proposal of JEDI (Joint Embedding DIffusion) world model, a latent end-to-end diffusion model that operates in latent space to mitigate curse of dimensionality
  • Reversing the worsening Performance Asymmetry trend while improving computational efficiency and achieving SOTA on Sym-HNS, Human-Optimal tasks, and Breakout

Methodology

The authors partition the Atari100k benchmark into Human-Optimal and Agent-Optimal subsets based on performance asymmetry. They propose a latent end-to-end joint embedding diffusion (JEDI) world model that learns in a latent space to address the curse of dimensionality inherent in pixel-level diffusion models. The model is evaluated on Atari100k using standard MBRL training and evaluation, with baselines including prior SOTA pixel diffusion world models, and metrics include HNS, Sym-HNS, and per-task performance.

Key Results

JEDI achieves state-of-the-art results on Sym-HNS, Human-Optimal tasks, and the Breakout task, reversing the worsening Performance Asymmetry trend while remaining competitive on the full Atari100k benchmark and improving computational efficiency.

Resources

Paper

Tags