Live streaming platforms have become popular media for young people in China to interact with others online. While the synchronized interaction with live streamers is said to make users feel more engaged than simply watching a video, it brings the risk of online violence when the words become highly aggressive and hurtful. This research analyzes such extremized situations by modeling opinion evolution during live streaming. An agent-based system is built to mimic audiences’ opinion interactions with socio-cognitive theories, including social-judgment theory, information overload, and entertainment motivation. The simulation results reveal that opinion polarization could happen even without the engagement of extremists, if audiences easily repulse others’ opinions. When positive and negative extremists are involved, the results suggest that their proportions are responsible for the polarization among ordinary audiences. The impacts of socio-cognitive factors are also discussed in the reseach.