Large Language Models as Universal Predictors? An Empirical Study on Small Tabular Datasets

Abstract: LLMs, originally developed for NLP, have demonstrated the potential to generalize across modalities and domains. With their in-context learning (ICL) capabilities, LLMs can perform predictive tasks over structured inputs without explicit fine-tuning on downstream tasks. In this work, we investigate the empirical function approximation capability of LLMs on small-scale structured datasets for classification, regression and clustering tasks. We evaluate the performance of state-of-the-art LLMs (GPT-5, GPT-4o, GPT-o3, Gemini-2.5-Flash, DeepSeek-R1) under few-shot prompting and compare them against established ML baselines, including linear models, ensemble methods and tabular foundation models (TFMs). Our results show that LLMs achieve strong performance in classification tasks under limited data availability, establishing practical zero-training baselines. In contrast, the performance in regression with continuous-valued outputs is poor compared to ML models, likely because regression demands outputs in a large (often infinite) space, and clustering results are similarly limited, which we attribute to the absence of genuine ICL in this setting. Nonetheless, this approach enables rapid, low-overhead data exploration and offers a viable alternative to traditional ML pipelines in business intelligence and exploratory analytics contexts. We further analyze the influence of context size and prompt structure on approximation quality, identifying trade-offs that affect predictive performance. Our findings suggest that LLMs can serve as general-purpose predictive engines for structured data, with clear strengths in classification and significant limitations in regression and clustering.