[Purpose/significance] While Large Language Model (LLM) demonstrates substantial potential in academic peer review, its application has also sparked widespread controversy regarding academic ethics and norms. Focusing on reviewers, this study investigates the behavioral processes and mechanisms underlying their collaboration with LLMs in real-world workflows. [Method/process] This study conducted in-depth interviews with 20 reviewers experienced in LLM-assisted peer review and employed thematic analysis for data analysis. [Result/conclusion] This study develops a human-AI collaborative behavior model in the context of peer review, consisting of five elements: collaborative conditions, collaborative motivations, collaborative strategies, collaborative feedback evaluation, and collaborative adaptation. Reviewers' behavioral decisions are moderated by subjective norms and capability foundations, and are driv⁃en by both instrumental and avoidance motivations. Based on the level of cognitive offloading and the degree of impact on the final review opinion, task allocation strategies between reviewers and LLMs can be structured as a continuum of "assistance–augmentation–co-creation agent." This continuum is accompanied by risk management strategies to ensureda⁃tasecurity and information quality. The collaboration between reviewers and LLMs is a continuous process of learning and adaptation. Driven by collaborative feedback evaluation, reviewers make corresponding behavioral adaptations, including the enhancement of LLM usage skills and the adjustment of collaborative strategies. This behavioral model provides new insights into the elements, pathways, and boundaries of human-AI collaboration in the context of peer review tasks.