Learning procedures and reliability thresholds for LLM-based world models

Develop and evaluate training methodologies for large language model–based world models, and establish criteria that determine when these models are sufficiently reliable to improve performance of downstream agents.

Background

Prior efforts have used LLMs as simulators, generators, or planning interfaces, but the pathway to systematically training them as world models and certifying their reliability for agent improvement has not been concretely defined.

The paper proposes reframing language modeling as next-state prediction within a fixed interaction protocol and investigates supervised fine-tuning on multi-turn trajectories to internalize environment dynamics.

References

While prior work has explored LLMs as simulators, experience generators, or planning interfaces \citep{chen2025scalingagentlearningexperience,li2025simulatingenvironmentsreasoningmodels,wu2025rlvrworldtrainingworldmodels,gu2025llmsecretlyworldmodel,wang2025world,he2025pretrained}, it remains unclear how to learn a world model and when it is reliable enough to improve downstream agents.

— From Word to World: Can Large Language Models be Implicit Text-based World Models?  (2512.18832 - Li et al., 21 Dec 2025) in Introduction (Section 1)

This shift is important but not yet settled. Latent futures may improve planning, but they can also become inaccurate precisely in rare situations where planning most needs reliability.

— Planning-Oriented End-to-End Autonomous Driving: Architectures, Evaluation, and Emerging Paradigms  (2608.20111 - Guan et al., 20 Aug 2026) in Section 4.7.3, World-Model Supervision