The present study has investigated how mutual briefs about task executions are updated during collaborative tasks using the Japanese map task corpus. The results have shown that the current models describe only part of mutual brief updating processes, and that there exist other types of processes. According to the current models, a mutual belief is considered to be achieved when the instruction follower accepts instructions given by the instruction giver and gives some sign of actually completing the given task. However, the present study has shown that mutual beliefs could be achieved even when the follower neither follows the instruction nor gives any sign of completion. The analysis indicates that the conversations in map tasks do not necessarily require prior planning and conversations to obtain clear mutual briefs about the goals to achieve the way the current models expect. Rather, ambiguous mutual beliefs and somewhat independent actions, coupled with inference about mutual goals, are sufficient to achieve the map tasks. In order to explain these results, we have proposed more detailed mechanisms about how mutual beliefs update.