From Specialist to Generalist: A Comprehensive Survey on World Models
Description
World models endow artificial agents with the internal predictive capabilities necessary to anticipate future states so as to act purposefully. While theoretical results underscore the necessity of world models for general tasking capability, implementing them involves navigating complex challenges in high-dimensional dynamics and compounding errors over long horizons. Currently, no existing approach simultaneously attains both precision and generalization, creating a divide between specialist and generalist models. In this survey, we systematically review the rapidly evolving field of world models through a distinct lens: Specialist versus Generalist. Unlike existing reviews, we frame the literature as a technical continuum spanning explicit physics-based priors to implicit data-driven learning. A key insight from our analysis is that despite the evolutionary trend toward generalist architectures, specialist and generalist paradigms are destined to coexist. We demonstrate that this persistence stems from a fundamental trade-off between the high precision required for control tasks and the broad adaptability needed for open-ended environments. By critically analyzing the strengths, limitations, and practical applications across this spectrum, we identify the open challenges hindering widespread deployment and propose a research roadmap to reconcile accuracy with transferability—a synergy essential for realizing Artificial General Intelligence (AGI).
Files
WM_Survey.pdf
Files
(30.3 MB)
| Name | Size | Download all |
|---|---|---|
|
md5:551bdb40e8867c8d2bb0c6f46b6ca0cf
|
30.3 MB | Preview Download |