[Purpose/significance] The customized generation and targeted dissemination of disinformation driven by large language models have intensified the persistence and diversity of individual cognitive biases in fact-checking processes. To address this issue, this article aims to examine the debiasing effects of LLMs in dynamic human-LLMs interactions on multiple types of cognitive biases, as well as their behavioral implications. [Method/process] A sequential mixed-methods design was adopted. Firstly, grounded theory was applied to interaction data between fact-checkers and LLMs to identify common types of cognitive biases and construct experimental variables. Subsequently, based on pretest-posttest experimental data, nonparametric statistical tests were conducted to evaluate the effects of LLMs-driven dynamic debiasing interactions on different types of cognitive biases and disinformation engagement behaviors.[Result/conclusion] Cognitive biases in fact-checking can be categorized into three core dimensions: cognitive processing deficits, selective exposure, and defensive beliefs, which exhibit dynamic and interwoven characteristics.LLMs, through personalized and evidence-based interactive communication, can effectively identify and mitigate various cognitive biases in fact-checking. Moreover, such interactions significantly reduce individuals′ willingness to "like" disinformation, while showing limited impact on promoting high-effort corrective behaviors such as commenting.