This contribution investigates the role of virtual space on social interactions during collaborative tasks. We previously observed that MUD users rely on spatial positions to refine the conversational context and thereby facilitate mutual understanding. Supporting mutual understanding is a main challenge of CSCL research. We explore how this may happen in a continuous space (VRML). Our first hypothesis was that the proximity of the emitter to the referred object clarifies the referential context. Our second hypothesis stated that the receiver uses gaze awareness (knowing what the emitter is looking at) in order to guess which object the emitter refers to. The experiment results confirmed the first hypothesis, surprisingly rejected the second hypothesis and reveal complex interactions between the two.