The access of idiomatic expressions was investigated in three cross-modal priming experiments. When the idiomatic string was predictable, subjects were faster at performing a lexical decision to idiomatically related targets than to literally related targets (Experiment 1). When the string was not recognizable as idiomatic until its completion, then subjects were faster on the target literally related to the last word (Experiment 2). Lexical decision for the same materials used in Experiment 2, when target words were presented 300 ms after the end of idiom was heard, was faster both for targets related to the last word and for the idiomatic meaning of the string, relative to control targets (Experiment 3). Neither the direct access model nor the lexical representation model seems adequate to explain the present findings. A new hypothesis concerning the representation and the processing of idioms is presented. (Cristina Cacciari and Patrizia Tabossi)