Large language models develop novel social biases through adaptive exploration