The widespread application of Generative Artificial Intelligence (Gen-AI) is transforming educational practices and driving pedagogical innovation. While cultivating higher-order thinking (HOT) represents a central educational goal, its achievement remains an ongoing challenge. Current evidence regarding the impact of Gen-AI on HOT is relatively fragmented, lacking systematic integration, particularly in the analysis of moderating variables. To address this gap, a meta-analysis approach was employed, integrating data from 29 experimental and quasi-experimental studies to quantitatively assess the overall impact of Gen-AI on learners’ HOT and to examine potential moderating factors. The analysis revealed that Gen-AI exerts a moderate positive effect on HOT, with the most significant improvement observed in problem-solving abilities, followed by critical thinking, while its effect on creativity is relatively limited. Moderation analyses further indicated that the impact of Gen-AI is significantly influenced by experimental duration and learners’ self-regulated learning (SRL) abilities: effects were strongest when interventions lasted 8–16 weeks, and learners with higher SRL capacities benefited more substantially. Based on the research findings, this study proposed that Gen-AI should be systematically integrated as a targeted instructional tool to foster HOT. Medium- to long-term interventions (8–16 weeks) are recommended to enhance learners’ problem-solving and critical thinking abilities. At the same time, effective approaches should also be explored to promote creative thinking through Gen-AI within existing pedagogical frameworks. Furthermore, individual learner differences should be accounted for by adopting dynamic and personalized scaffolding strategies to foster SRL, thereby maximizing the educational potential of Gen-AI in cultivating innovative talents.

Large language models (LLMs) have expanded the capabilities of intelligent tutoring systems, but adaptive regulation of instructional challenge and control of domain-specific factual drift remain difficult. We propose Cogent-Adversary, a dual-agent tutoring framework that combines cognitive-state-informed adaptive control (CODA) with Knowledge-Graph-Grounded Adversarial Reasoning and an adversarial consistency check (KGAR–ACCL). The framework was evaluated in a two-week stratified randomized pre-test–post-test study with a one-week delayed assessment at one Chinese university. A total of 78 students were randomized to four tutoring conditions, and 72 completed all sessions. In the primary T1 analyses, the CODA-adaptive condition was associated with higher adjusted knowledge and domain-specific reasoning scores than the single-agent condition (Hedges’ $g=2.02$ and $2.87$, respectively), while participant-clustered models indicated lower factual-error odds than the fixed and rule-based adaptive conditions. Given the short, single-site design and self-developed outcomes, these unusually large effects should be interpreted cautiously. Post hoc offline replays were exploratory only. The findings therefore provide preliminary proof-of-concept evidence for dynamically regulated, knowledge-grounded adversarial tutoring.