Adapting large language models (LLMs) to complex reasoning tasks often requires expensive fine-tuning efforts. We propose GAMeta-Prompt, a generative adversarial framework that streamlines prompt-based meta-learning through an iterative, team-based adversarial process. First, our approach organizes a role-based “expert team”. One member assumes the role of “team manager” acting as the sole conduit for all expert communications to maintain independent reasoning streams. Second, we engage an “adversarial team” that mirrors the expert structure but aims to refute the solution proposed by the expert team, creating a multi-round confrontation that targets logical gaps and domain oversights. Finally, we propose an adaptive evaluation mechanism based on multi-dimensional feedback, which dynamically adjusts the evaluation criteria by quantifying the consistency of expert scores and the uncertainty of evaluation. The method combines adaptive scoring integration, feedback adjustment and multi-round optimization to improve the robustness and accuracy of the evaluation and ensure the overall reliability of the scheme. Experiments demonstrate that GAMeta-Prompt outperforms conventional prompt engineering methods in few-shot and zero-shot scenarios, often enabling smaller models to match or exceed the performance of larger ones under resource constraints. By blending persona-based context, iterative adversarial checks, and global scoring, our framework offers a scalable, parameter-efficient pathway to refining LLM reasoning without fine-tuning.