一项系统性研究调查了大语言模型在表格数据预测任务中失败的原因，通过控制实验验证了多个假设[1]。研究针对纯推理模式下的前沿大语言模型进行测试，未采用工具、代理框架或微调等辅助手段[1]。
研究验证了五个主要假设：数据噪声处理能力、CSV格式线性化、数值分词、测试点数量及输入维度[1]。结果表明，前四个因素均非关键障碍，但维度成为决定性因素——大语言模型是九种测试方法中唯一在维度增加时准确率下降的方法，而传统基线模型则保持稳定或持续改进[1]。
研究基于31个基准数据集的随机线性投影进行测试[1]。在二维空间中，大语言模型的预测行为与局部距离方法高度一致，网格一致性达到91.6%；但在高维空间中，无任何经典学习模型能复现其行为[1]。研究团队对252个经典模型配置进行了对比分析[1]。
这一发现解释了为何大语言模型在其他领域表现卓越，却在表格任务上持续被五十年前的经典方法击败[1]。虽然研究证实了大语言模型的预测能力随维度增长而消退，且这一模式无任何加噪经典学习器能够复现，但其内部机制仍待进一步探索[1]。
A systematic investigation has identified dimensionality as the critical factor explaining why large language models consistently underperform on tabular prediction tasks [1]. Researchers tested state-of-the-art LLMs operating in pure inference mode without tools, agents, or fine-tuning, and examined five competing hypotheses to account for this gap [1]. While data noise, CSV formatting, numerical tokenization, and test set size proved insignificant, the research demonstrates that LLMs are unique among nine tested methods in experiencing accuracy degradation as input dimensionality increases [1].
The study employed random linear projections across 31 benchmark datasets to control experimental conditions [1]. In two-dimensional space, LLM prediction behavior aligned with local distance methods at 91.6% grid consistency, but no classical baseline model could replicate LLM behavior in higher dimensions [1]. This phenomenon creates a puzzling contrast: while LLMs excel across numerous domains, they are systematically defeated by classical methods developed over fifty years ago on tabular problems [1]. The researchers note that the performance decline pattern cannot be reproduced by any noise-robust classical learner, while the underlying mechanisms remain an open question [1].