Turns out teaching an AI to predict what happens next works better than letting it look things up.
Researchers compared two ways of building "world models" - language models that simulate how an environment will change in response to an agent's action, so the agent can plan ahead without actually trying the action first. One approach fine-tunes the model on collected experience. The other uses retrieval-augmented generation (RAG), pulling similar past transitions from a memory store instead of baking them into the model's weights. Testing both across five environments spanning embodied tasks, web navigation, and social settings, the fine-tuned world models won in 15 of 20 test configurations. RAG's one clear advantage was data efficiency - it got useful results with far less experience, while fine-tuning only paid off once given a lot more of it.
That cuts against the current enthusiasm for retrieval as a cheap stand-in for retraining. For agents that need to reason about consequences before acting - the kind of planning behind everything from robotics to software agents - the expensive, data-hungry approach to building a simulator still mostly beats the lookup-table approach. The researchers also pinned down why RAG struggles: its retriever regularly surfaces suboptimal past transitions, not just imperfect ones, which drags down the simulation.
Their fix was a hybrid system - a fine-tuned core model backed by a better-curated retrieval memory - which outperformed both pure approaches. The real answer isn't either-or. It's do the expensive thing, and keep a good notebook too.