Behavioral Analysis of Information Salience in Large Language Models