This research investigates how multilingual identity influences communication tool preferences and personal self-efficacy in global virtual teams, examining the mediating role of conflicts. Drawing on Media Richness Theory, Social Identity Theory, and Conflict Management Theory, we hypothesized that (H1) individuals speaking more languages prefer richer communication tools (video calls) over leaner channels (email); (H2) richer tool preferences reduce both task and relationship conflict; (H3) task conflict is positively related to self-efficacy while relationship conflict is negatively related to self-efficacy; and (H4) tool selection and conflicts sequentially mediate individual outcomes.Analyzing data from 2,823 students across 629 teams in the 2022 Spring X-Culture project using SPSS PROCESS, we examined ten communication scenarios. Our findings challenge the hypothesized conflict-centered framework: the direct pathway (multilingual identity → tool preferences → self-efficacy) proved significantly stronger than the conflict-mediated pathway. In “introduction” and “member missing” scenarios, multilingual identity increased preference for high-richness tools, enhancing self-efficacy directly, with conflict playing minor mediating roles. The relationship between multilingual identity and tool preferences proved context-dependent: significant positive correlations in three scenarios, non-significant correlations in five routine scenarios, and slight negative correlations in two explanatory contexts. While information-rich tool preferences consistently correlated negatively with both conflict types (especially relationship conflict), these reductions played surprisingly modest mediating roles. Both conflict types negatively impacted self-efficacy (especially task conflict), contradicting H3. This reversal occurs because virtual environments lack the non-verbal cues that help teams navigate disagreements productively.The critical exception emerged when addressing the “member missing” question: language number significantly predicted tool preferences (β=0.0530, p
