Original title: EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation
Qing Zhao, Haowei Li, Weijian Deng, Pengxu Wei +1 · September 2026
Most AI tools write scientific ideas as plain text. This makes it hard to see what the AI actually believes, test its ideas, or fix mistakes. To solve this, researchers built EvoSCM. Instead of plain text, it gives the AI explicit structural causal modelsDiagrams that show how different factors cause changes in a system.. These models are clear maps that show how causes link to effects. EvoSCM tracks a whole group of competing models at once. During each round, it guesses hidden mechanisms, designs focused experiments, makes testable predictions, and runs the tests. If the results do not match the predictions, EvoSCM turns the errors into rules to update the models. When tested on DiscoverPhysics, EvoSCM beat standard baselines. It produced more accurate explanations, better predictions, and used experiments more effectively.
Imagine baking a cookie without a recipe, trying to figure out why it tastes bad. Each time you change an ingredient, you don't know if it fixed the problem or made it worse, because you lack a clear model of what's happening. In this paper, the AI agent is like a baker with multiple evolving recipes, testing each one through experiments and updating them based on what tastes right, so it learns the true recipe faster.
Scientists using AI can have agents that learn from experiments in a structured way, making discoveries more reliable. This changes how AI assists research by providing clear, updatable beliefs instead of vague hypotheses.
Evaluated only on DiscoverPhysics benchmark for noncanonical physical worlds; generalization to other scientific domains is not established.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation — Qing Zhao, Haowei Li, Weijian Deng et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.