A dynamic grouping mechanism, developed through clustering human-AI interaction patterns, is devised to optimize learner collaboration by calibrating group settings in language learning contexts. A corpus-based comparative analysis of writing task performance and prompt-feedback data clusters is conducted across three division settings: random, self-organized and prompt-feedback mediated grouping. Seventy-five advanced English learners are organized into fifteen groups, rotating through the three division settings, and assigned three writing tasks: (1) product descriptions, (2) marketing proposals, and (3) holding statements. These tasks are completed with the support of AI, during which data on prompts and feedback are recorded, categorized, and analyzed alongside the final submissions which are graded through a blind peer-review process. Key metrics, including the frequency of human-AI interactions, the adoption ratio of AI assistances, the coverage ratio of AI references, and the duration of task completion, are examined and compared across group profiles and task grades, revealing distinct group performance patterns to collaboration dynamics. The integration of AI into language learning has expanded the possibilities for diverse and adaptive learning environments, which ensure the compatibility between task objectives, collaboration requirements and group dynamics to enhance collaboration potential and optimize learning efficacy.