📄 中文摘要
过去十年，应用人工智能研究的主要指标是模型的参数数量。随着每一代新模型的推出，研究者们越来越关注模型的规模，而忽视了模型结果的可重复性。这种趋势导致了可重复性危机的加剧，许多研究结果在不同实验中无法得到验证。模型的复杂性和参数的增加并未必能带来更可靠的结果，反而可能引发对模型性能的误解。研究者们需要重新审视评估标准，关注模型的可解释性和结果的可靠性，以推动人工智能领域的健康发展。
更大的模型，不可靠的结果：人工智能的可重复性危机
出处: Bigger Models, Unreliable Results: The Reproducibility Crisis in AI
发布: 2026年3月29日
📄 English Summary
Bigger Models, Unreliable Results: The Reproducibility Crisis in AI
The past decade of applied AI research has been dominated by a singular focus on parameter count as the primary metric for model evaluation. Each new generation of models has emphasized scale, often at the expense of reproducibility. This trend has exacerbated the reproducibility crisis, with many research findings failing to be validated across different experiments. The increasing complexity and parameterization of models do not necessarily correlate with more reliable outcomes, potentially leading to misconceptions about model performance. A reevaluation of evaluation standards is necessary, with a focus on model interpretability and result reliability to foster a healthier advancement in the AI field.