Closing the Train-Test Gap in World Models for Gradient-Based Planning

arXiv 25.12 2025 47.7 method

TLDR

Proposes train-time data synthesis to close the train-test gap in world models for gradient-based planning, matching CEM in 10% time.

Reasoning

The paper identifies a key mismatch between training and test objectives for world models and introduces data synthesis techniques to improve gradient-based planning. Strengths include a clear problem formulation and strong empirical results on manipulation and navigation tasks. Weaknesses are the lack of real-world validation and limited comparison to other gradient-based methods.

Read-first score

Read-first score 47.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 31.

Recency 8%

86.7

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

Methodology quality 25%

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

Topical relevance 42%

44.3

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 25%

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: 389.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Identifying the train-test gap in world models for gradient-based planning, where models are trained on next-state prediction but used for action sequence estimation at test time.
Proposing train-time data synthesis techniques to close this gap and improve gradient-based planning performance.
Achieving performance that outperforms or matches the classical gradient-free cross-entropy method (CEM) across object manipulation and navigation tasks in only 10% of the time budget.

Methodology

The authors propose train-time data synthesis techniques to close the train-test gap in world models used for gradient-based planning. The world model is trained on a next-state prediction objective but used at test time for action sequence estimation. The method is evaluated on object manipulation and navigation tasks, comparing against the gradient-free cross-entropy method (CEM).

Key Results

The proposed approach outperforms or matches the classical gradient-free cross-entropy method (CEM) across a variety of object manipulation and navigation tasks, while using only 10% of the time budget.

Resources

Paper