Auto-Fill Uses Specialist LLMs for Accurate Missing Data Prediction

This paper introduces Auto-Fill, a novel approach for accurately predicting missing cell values in tabular data, a critical task in data cleaning. While existing state-of-the-art reasoning models show promise, they are often costly to deploy at scale and prone to overconfidence, leading to hallucinations or false positives. Auto-Fill addresses these limitations by recognizing that high-precision missing-value prediction requires a blend of world knowledge, text-based reasoning, and code-based reasoning. The proposed system post-trains three specialist small language models (SLMs), each optimized for one of these distinct capabilities. A calibrated ensemble mechanism dynamically selects the most confident specialist or abstains from prediction, ensuring high accuracy and reliability. Extensive experiments across 11 benchmarks and 2200 real tables demonstrate that Auto-Fill achieves superior accuracy compared to frontier models like o3-pro, Gemini 3 Pro, and DeepSeek R1, all while operating at less than 1% of their cost. This highlights the effectiveness of specialization and calibrated abstention in tabular data imputation.