Autism spectrum disorder (ASD) is a heterogenous disorder predominantly characterized by social and communicative differences, but increasingly recognized to also alter (multi)sensory function. To face the heterogeneity and ubiquity of ASD, researchers have proposed models of statistical inference operating at the level of computations. Here, we attempt to bridge both across domains, from social to sensory, and levels of description, from behavioral computations to neural ensemble activity to a biologically-plausible artificial neural network, in furthering our understanding of autism. We do so by mapping visuo-tactile peri-personal space (PPS), and examining its electroencephalography (EEG) correlates, in individuals with ASD and neurotypical individuals during both a social and non-social context given that (i) the sensory coding of PPS is well understood, (ii) this space is thought to distinguish between self and other, and (iii) PPS is known to remap during social interactions. In contrast to their neurotypical counterparts, psychophysical and EEG evidence suggested that PPS does not remap in ASD during a social context. To account for this observation, we then employed a neural network model of PPS and demonstrate that PPS remapping may be driven by changes in neural gain operating at the level of multisensory neurons. Critically, under the anomalous excitation-inhibition (E/I) regime of ASD, this gain modulation does not result in PPS resizing. Overall, our findings are in line with recent statistical inference accounts suggesting diminished flexibility in ASD, and further these accounts by demonstrating within an example relevant for social cognition that such inflexibility may be due to E/I imbalances.