Recent Event Related Potential (ERP) studies in sentence comprehension suggest that the predictability of a word in a sentence does not straightforwardly corresponds to its processing difficulty reflected by the N400 amplitude (e.g., Chow et al., submitted; Hoeks et al., 2003; Kim & Osterhout, 2005). For instance, the N400 to the verb in the hearty meal was devouring and in the hearty meal was devoured was shown to be nearly identical. In this talk, I will argue, following Chow et al. (submitted) and based on our new experimental evidence from Japanese, that this insensitivity of N400 to the word predictability reflects the failure to successfully incorporate top-down information in prediction on time. I will then propose a model of lexical prediction that combines the fast-acting spreading-activation mechanism that generates a potential search space for up-coming words and relatively slow-acting top-down mechanism that successively narrows down the generated search space. Some implications of the model, along with a potential experimental idea to test some of them, will also be discussed.