LLM World Models

LLM world models represent environments as states and predict how actions change them, from web pages and APIs to games, spatial tasks, and research experiments. Common approaches fine-tune models on interaction traces, retrieve transition rules from memory, combine language models with symbolic constraints, or learn latent dynamics that agents can rehearse before acting. These models are used to generate training experience, improve long-horizon planning, and assess risks or candidate actions without repeated real-world interaction. A central challenge is that accurate local predictions or latent representations may still fail to support reliable reasoning over long horizons, hidden states, or rare but consequential events.