This paper presents work on a new platform for producing realistic group conversation dynamics in shared virtual environments. An avatar, representing users, should perceive the surrounding social environment just as humans would, and use the perceptual information for driving low level reactive behaviors. Unconscious reactions serve as evidence of life, and can also signal social availability and spatial awareness to others. These behaviors get lost when avatar locomotion requires explicit user control. For automating such behaviors we propose a steering layer in the avatars that manages a set of prioritized behaviors executed at different frequencies, which can be activated or deactivated and combined together. This approach gives us enough flexibility to model the group dynamics of social interactions as a set of social norms that activate relevant steering behaviors. A basic set of behaviors is described for conversations, some of which generate a social force field that makes the formation of conversation groups fluidly adapt to external and internal noise, through avatar repositioning and reorientations. The resulting social group behavior appears relatively robust, but perhaps more importantly, it starts to bring a new sense of relevance and continuity to the virtual bodies that often get separated from the ongoing conversation in the chat window.