Abstract--Focusing on the online processing of Negative polarity items (NPIs), in three studies I will address questions about the division of labor between semantic and pragmatic processes, the accessibility of different semantic features, and the potential individual differences in pragmatic processing. NPIs such as ever and any are known to be subject to particular licensing conditions, but the exact nature of the licensing condition is still under theoretical debate. Using ERP recording and behavioral measures, the current study argues that (i). there are two different kinds of licensing mechanisms: one licenses NPIs in the grammar proper, and the other licenses NPIs through pragmatic inferences; (ii). not all semantic properties are equally accessible in incremental parsing. In particular, although negation can be rapidly integrated into the current semantic representation, and hence facilitates the lexical access of an upcoming NPI word; licensing features that are not negative per se are not automatically activated; and (iii). electrophysiologically, N400 is best interpreted as an index of lexical access, rather than post-lexical semantic integration; P600, on the other hand, represents a post-lexical semantic process. Results from three experiments will be discussed. In Expt 1, ERP responses time-locked to NPIs licensed under different "negative" licensors (no, few, only and emotive factives such as surprised), are compared with an unlicensed ungrammatical condition. No and few are grammatical licensors under any notion of NPI licensing (e.g. negation, downward entailing or non-veridicality); whereas only and surprised type contribute some negative inferences to the semantic context, but they do not straightforwardly fit into a grammatical definition of an NPI licensor. In Expt 2, ERP response on NPIs licensed by no is compared to those licensed by every. Both are standard downward entailing licensors, but every is not negative. In Expt 3, using a self-paced reading and acceptability judgment task, NPI interference effect is compared to the well-known number agreement attraction effect. Individual subjects' autistic traits were correlated with the size of the interference/attraction effect.