Does an AI that guides science need a real model of how the world works, or are pattern-matching and outside support enough?

This explores whether an AI steering scientific discovery needs a real internal model of how the world works, or whether strong pattern-matching can do the job. No paper in the corpus tests this head-on. Read together, though, the papers point to a split answer: models seem to get by without true world models when generating and predicting, but reliable steering so far comes from structure built around the model, not from structure inside it.

Start with what a world model is supposed to be. A useful one lets you ask "what happens if I intervene?" and not only "what usually comes next?" What makes a world model actually useful for reasoning?. On that test, current models often fall short. Transformers trained on orbital mechanics predict trajectories well but, when probed, turn out to have learned patchy rules that hold only for particular slices of the data, not anything like Newton's laws Do foundation models learn world models or task-specific shortcuts?. Other work argues that LLMs still absorb some real causal structure secondhand, because the text they learn from was written by people who were in contact with the world, but that this chain has gaps when it comes to checking and updating Can large language models develop genuine world models without direct environmental contact?.

The surprise is how far heuristics alone can go. Fine-tuned LLMs beat neuroscience experts at predicting which experimental results actually happened. The same tendency to blend patterns that causes hallucination when you ask about the past becomes useful generalization when you ask about the future Can LLMs predict novel scientific results better than experts?. LLM-generated research ideas get rated as more novel than experts' ideas, though somewhat less feasible Do language models generate more novel research ideas than experts?. Models fine-tuned on psychology experiments predict human choices better than theory-driven cognitive models Can language models learn to model human decision making?. So for proposing and forecasting, a missing world model doesn't stop the model from being useful.

Steering is where the weakness shows. Steering means judging which hypotheses deserve pursuit, and models tend to judge a claim as supported when it sounds familiar from training data, even when the stated premises don't back it up Do LLMs predict entailment based on what they memorized?. A discovery loop built on that habit will drift toward hypotheses that are already familiar. The systems that steer well bring in causal structure from outside. Giving an LLM an explicit structural causal model lets it propose and test social-science hypotheses in simulation, and it reliably gets the direction of effects right but not their size Can structural causal models automate social science with language models?. Robin goes further and uses the real world as the world model: literature agents and a data agent propose hypotheses, human-run wet-lab experiments answer them, and the results revise the next round Can multi-agent systems guide wet-lab discovery through iterative cycles?.

One more lateral point: when researchers do build explicit world models, the ones that work model more than physics. Predicting what people will do requires tracking their beliefs and intentions along with the physical scene Can world models predict human action from physics alone?. Language world models trained on agent trajectories can partly stand in for real environments Can language models learn to simulate agent environments?. The practical takeaway: for now, the world model in AI-driven discovery usually sits outside the AI, as a causal graph, a simulator, or a lab. The open question is whether learned world models will get good enough to bring that structure inside.