EvoSCM: Scientific Belief Revision Through Causal Model Evolution and Experimentation
The paper turns an AI scientist’s “beliefs” into testable causal models that are revised after experiments.
EvoSCM keeps a population of structural causal model hypotheses instead of leaving hypotheses as free-form text. In each loop, the agent proposes hidden mechanisms, chooses interventions, makes falsifiable predictions, and updates the models when observations disagree. The authors test it on DiscoverPhysics, where agents must infer hidden rules in altered physical worlds, and report better explanations, predictions, and use of experiments than baselines. ArXiv · AI/CL/LG's note
EvoSCM keeps a population of structural causal model hypotheses instead of leaving hypotheses as free-form text. In each loop, the agent proposes hidden mechanisms, chooses interventions, makes falsifiable predictions, and updates the models when observations disagree. The authors test it on DiscoverPhysics, where agents must infer hidden rules in altered physical worlds, and report better explanations, predictions, and use of experiments than baselines. ArXiv · AI/CL/LG's note
score 5