Human language efficiently conveys information in context, with listeners inferring meaning from utterances that convey only relevant information. This study models pragmatic reasoning in referential communication games, assuming speakers aim to be informative and listeners use Bayesian inference to recover intended referents. The model uses information-theoretic tools to predict pragmatic reasoning, showing close agreement with human judgments. The model assumes that speakers choose words to be informative, with informativeness measured by surprisal. Listeners use Bayesian inference to determine the most likely referent, combining prior knowledge (contextual salience) with likelihood (informativeness). The model was tested with three groups: speaker, salience, and listener. Results showed strong correlations between model predictions and human judgments, especially when combining salience and informativeness terms. The model integrates insights from various traditions, including disambiguation models, game-theoretic signaling models, and systems for generating referring expressions. It uses an information-theoretic definition of informativeness and empirical measurements of common knowledge to capture the richness of human pragmatic inference. Participants saw sets of objects with varying features and bet on which object a word referred to. The speaker condition tested likelihood, the salience condition tested prior knowledge, and the listener condition tested posterior predictions. The model accurately predicted human behavior in all conditions, demonstrating its effectiveness in modeling communication. The study highlights the importance of combining information-theoretic principles with empirical data to understand pragmatic reasoning in context. It provides a framework for formal models of communication, emphasizing the role of informativeness and shared knowledge in human language use.Human language efficiently conveys information in context, with listeners inferring meaning from utterances that convey only relevant information. This study models pragmatic reasoning in referential communication games, assuming speakers aim to be informative and listeners use Bayesian inference to recover intended referents. The model uses information-theoretic tools to predict pragmatic reasoning, showing close agreement with human judgments. The model assumes that speakers choose words to be informative, with informativeness measured by surprisal. Listeners use Bayesian inference to determine the most likely referent, combining prior knowledge (contextual salience) with likelihood (informativeness). The model was tested with three groups: speaker, salience, and listener. Results showed strong correlations between model predictions and human judgments, especially when combining salience and informativeness terms. The model integrates insights from various traditions, including disambiguation models, game-theoretic signaling models, and systems for generating referring expressions. It uses an information-theoretic definition of informativeness and empirical measurements of common knowledge to capture the richness of human pragmatic inference. Participants saw sets of objects with varying features and bet on which object a word referred to. The speaker condition tested likelihood, the salience condition tested prior knowledge, and the listener condition tested posterior predictions. The model accurately predicted human behavior in all conditions, demonstrating its effectiveness in modeling communication. The study highlights the importance of combining information-theoretic principles with empirical data to understand pragmatic reasoning in context. It provides a framework for formal models of communication, emphasizing the role of informativeness and shared knowledge in human language use.