Verbal short-term memory (STM) is highly sensitive to learning effects: digit sequences or nonword sequences which have been rendered more familiar via repeated exposure are recalled more accurately. In this study we show that sublist-level, incidental learning of item co-occurrence regularities affects immediate serial recall of words and nonwords, but not digits. In contrast, list-level chunk learning affects serial recall of digits. In a first series of experiments, participants heard a continuous sequence of digits in which the co-occurrence of digits was governed by an artificial grammar. In a subsequent STM test participants recalled lists that were legal or illegal according to the rules of the artificial grammar. No advantage for legal lists over illegal lists was observed. A second series of experiments used the same incidental learning procedure with nonwords or non-digit words. An advantage for legal "versus" illegal list recall was observed. A final experiment used an incidental learning task repeating whole lists of digits; this led to a substantial recall advantage for legal "versus" illegal digit lists. These data show that serial recall of non-digit words is supported by sublist-level probabilistic knowledge, whereas serial recall of digits is only supported by incidental learning of whole lists. (Contains 8 figures and 3 tables.)