Large language models develop novel social biases through adaptive exploration 08/08 • openreview.net