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

Learning an Adversarial World Model for Automated Curriculum Generation in MARL

arXiv 25.9 2025 51.8 method, system

TLDR

Proposes adversarial world model for automated curriculum generation in MARL, where attacker generates challenging worlds and defenders learn cooperative strategies.

Reasoning

Strengths include a novel adversarial co-evolutionary framework for infinite curriculum and emergence of complex behaviors like flanking and focus-fire. Weaknesses are the lack of quantitative results, real-world benchmarks, or detailed empirical evaluation in the abstract.

Read-first score

Read-first score 51.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 8%
86.7

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

Topical relevance 42%
54.3

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

Methodology quality 25%
50

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

Reproducibility 25%
38

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 413.

Keyword Scores

world model
9
generative world model
9
interactive world model
7
world simulator
6
model-based reinforcement learning world model
4
world dynamics prediction
3
video world model
0

Deep Analysis

Innovations

  • Framing environment generation as an adversarial game for automated curriculum generation in multi-agent reinforcement learning (MARL)
  • Co-evolutionary dynamic between a procedurally generative attacker and a cooperative defender team, creating a self-scaling environment
  • Emergence of complex intelligent behaviors (flanking, shielding, focus-fire, spreading) from minimal training

Methodology

The system involves a team of cooperative multi-agent defenders learning to survive against a procedurally generative attacker. The attacker learns to produce increasingly challenging configurations of enemy units, dynamically creating novel worlds tailored to exploit the defenders' current weaknesses. The defender team learns cooperative strategies to overcome these threats, forming a co-evolutionary dynamic that generates a self-scaling environment.

Key Results

With minimal training, the approach leads to the emergence of complex intelligent behaviors such as flanking and shielding by the attacker, and focus-fire and spreading by the defenders.

Tags