Active exploration helps humans overcome cognitive biases in causal reasoning, but LLMs don't automatically benefit the same way, suggesting they may lack human-like adaptive learning strategies.
This paper compares how humans and large language models learn causal rules through active exploration. Adults traditionally struggle with conjunctive rules (effects requiring multiple causes), but the study shows active exploration significantly improves human performance. LLMs show mixed results—some match human accuracy but use less efficient exploration strategies.