Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, short

Summary

Auto-Fill is a novel approach designed to accurately predict missing cell values in tabular data, addressing the high cost and overconfidence issues of existing state-of-the-art reasoning models. Developed by Yurong Liu and colleagues, this system post-trains three specialist small language models (SLMs), each focusing on a distinct capability: world knowledge, text-based reasoning, or code-based reasoning. Auto-Fill employs a calibrated ensemble mechanism that dynamically selects the most confident specialist or abstains from prediction to ensure high accuracy. Extensive experiments across 11 benchmarks, utilizing 2200 real tables from diverse domains, demonstrate Auto-Fill's superior accuracy compared to models such as o3-pro, Gemini 3 Pro, and DeepSeek R1. Crucially, it operates at less than 1% of the cost of these frontier models, highlighting the effectiveness of specialization and calibrated abstention in tabular data. The system's code is publicly available and was submitted on 22 Jul 2026 for VLDB 2026.

Key takeaway

For data scientists or ML engineers building robust data cleaning pipelines, Auto-Fill offers a compelling alternative to costly, overconfident frontier models for missing value prediction. You should consider integrating this specialist SLM ensemble, especially when high precision and cost efficiency are critical. Its calibrated abstention mechanism reduces hallucinated predictions, improving data quality and model reliability in your tabular data applications. Explore the publicly available code to assess its fit for your specific datasets.

Key insights

Auto-Fill uses specialized small language models and calibrated abstention to achieve high-precision, cost-effective missing value prediction in tabular data.

Principles

• High-precision missing-value prediction requires world, text, and code knowledge.
• Specialization of models enhances accuracy and cost-efficiency.
• Calibrated abstention improves prediction reliability.

Method

Auto-Fill post-trains three specialist SLMs, each for world knowledge, text-based, or code-based reasoning. A calibrated ensemble dynamically selects the most confident specialist or abstains.

In practice

• Deploy specialist SLMs for tabular data cleaning.
• Implement calibrated ensemble for prediction confidence.
• Utilize Auto-Fill for cost-effective data imputation.

Topics

• Missing Value Prediction
• Tabular Data
• Small Language Models
• Ensemble Learning
• Data Cleaning
• Auto-Fill

Code references

• lyrain2001/auto-fill

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, Data Scientist

Related on AIssential

• Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models — Specialized small language models with calibrated abstention significantly improve tabular data imputation accuracy and cost-efficiency.
• Adaptive Conformal Prediction for Improving Factuality of Generations by Large Language Models — Adaptive conformal prediction improves LLM factuality by calibrating uncertainty estimates based on specific prompts.
• Data Cleaning Techniques for Improved AI Predictions. — Effective data cleaning is fundamental for transforming raw data into reliable inputs for accurate analysis and machine learning models.
• Ensembles of Ensembles of Ensembles: A Guide to Stacking — Ensembling diverse models across multiple layers enhances predictive performance and robustness for tabular and time series data.
• LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction — LLMs serve best as complementary components to strong tabular machine learning baselines, not replacements, for industrial data with limited semantics.
• From prediction sets to calibrated class scores — Conformal classification simultaneously yields both prediction sets and formally calibrated p-values, resolving a critical gap in tabular ML.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.