Large-scale group decision-making (LSGDM), which involves dozens to hundreds of decision-makers (DMs), is attracting extensive attention and has become an interesting and hot topic in recent years. Because of various backgrounds and expression habits, DMs tend to elicit preferences with different preference representation structures. Moreover, due to various attitudes and interests, some DMs may adopt non-cooperative behaviors to further benefit themselves in LSGDM. To cope with these issues, this study develops an adaptive consensus framework to support heterogeneous LSGDM. A cosine similarity-based optimization model is constructed and its analytical solution is derived to directly obtain the collective priority vector of the group using heterogeneous preferences. An extended k-means algorithm is then used to classify DMs. Subsequently, a two-stage uninorm-based behavior management method is developed to generate personalized weight feedback to each DM and sub-cluster according to their cooperative and non-cooperative degrees in the consensus-reaching process. Finally, an illustrative example followed by simulation and comparative analyses is provided to reveal the advantages and features of the proposed approach.