Spatiotemporal Forecasting as Planning: A Model-Based Reinforcement Learning Approach with Generative World Models

arXiv 25.10 2025 54 method

TLDR

Proposes SFP, a model-based RL approach using generative world models for spatiotemporal forecasting, improving prediction via beam search planning and self-training.

Reasoning

The paper introduces a novel paradigm combining model-based reinforcement learning with generative world models to address stochasticity and non-differentiable metrics in spatiotemporal forecasting. Strengths include the innovative use of beam search planning and self-training, but the abstract lacks explicit real-world validation and details on scalability or limitations.

Read-first score

Read-first score 54, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

Recency 8%

86.7

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

Topical relevance 42%

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 465.

Keyword Scores

world model

generative world model

model-based reinforcement learning world model

world simulator

world dynamics prediction

video world model

interactive world model

Deep Analysis

Innovations

Formulating spatiotemporal forecasting as a planning problem using model-based reinforcement learning (MBRL)
Constructing a Generative World Model to simulate diverse, high-fidelity future states for imagination-based environmental simulation
Using beam search-based planning with non-differentiable domain metrics as reward signals to explore high-return future sequences
Iterative self-training where high-reward candidates serve as pseudo-labels to continuously optimize the agent's policy

Methodology

SFP constructs a Generative World Model to simulate diverse, high-fidelity future states, enabling imagination-based environmental simulation. A base forecasting model acts as an agent, guided by a beam search-based planning algorithm that leverages non-differentiable domain metrics as reward signals to explore high-return future sequences. These identified high-reward candidates then serve as pseudo-labels to continuously optimize the agent's policy through iterative self-training.

Key Results

The approach significantly reduces prediction error and demonstrates exceptional performance on critical domain metrics, particularly in capturing extreme events.

Resources

Paper