Puzzle it Out: Local-to-Global World Model for Offline Multi-Agent Reinforcement Learning

arXiv 26.1 2026 56.6 method

TLDR

Proposes a local-to-global world model for offline multi-agent RL to generate synthetic data and improve policy generalization.

Reasoning

The paper introduces a novel framework (LOGO) that uses local predictions to infer global dynamics, addressing the challenge of modeling joint dynamics in offline MARL. Strengths include a principled approach to uncertainty-aware data augmentation and reduced computational overhead. Weaknesses: no real-world validation, and the abstract lacks explicit comparison to baselines or empirical results.

Read-first score

Read-first score 56.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 24.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

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

Reproducibility 25%

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

Topical relevance 42%

34.3

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 587.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

generative world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Local-to-global (LOGO) world model that leverages local predictions to infer global state dynamics, improving prediction accuracy while implicitly capturing agent-wise dependencies.
Uncertainty-aware sampling mechanism that adaptively weights synthetic data by prediction uncertainty, reducing approximation error propagation to policies.
Requires only an additional encoder for uncertainty estimation, significantly reducing computational overhead compared to conventional ensemble-based methods.

Methodology

The paper proposes a local-to-global world model for offline multi-agent reinforcement learning. It uses local predictions, which are easier to estimate, to infer global state dynamics and generate synthetic data to augment the original dataset. An uncertainty-aware sampling mechanism adaptively weights the synthetic data by prediction uncertainty to ensure reliable policy learning, and the method is evaluated across 8 scenarios against 8 baselines.

Key Results

The method surpasses state-of-the-art baselines on standard offline MARL benchmarks, establishing a new model-based baseline for generalizable offline multi-agent learning.

Resources

Paper