Learning to Predict Missing Values Accurately with Specialist Language Models
Official repository for the paper accepted at VLDB 2026.
Predicting missing cell values is a fundamental problem in data cleaning. Auto-Fill post-trains three small language model (SLM) specialists — one each for knowledge, reasoning, and coding — and combines them with a calibrated ensemble that selects the most confident specialist or abstains. This yields high-precision predictions at a fraction (< 1%) of the cost of frontier reasoning models such as o3-pro, Gemini 3 Pro, and DeepSeek-R1, across 11 benchmarks spanning 2,200 real tables.
Note
Code, models, and benchmark are coming soon. We are preparing the artifacts for release; this repository currently serves as a placeholder.
| Component | Status |
|---|---|
| Training & inference code | 🚧 Coming soon |
| Model checkpoints | 🚧 Coming soon |
| Benchmark | 🚧 Coming soon |
Star or watch this repository to be notified when the release is available!
@article{autofill2026,
author = {Yurong Liu and Yeye He and Haoyu Dong and Junjie Xing and
Shi Han and Dongmei Zhang and Surajit Chaudhuri},
title = {Auto-Fill: Learning to Predict Missing Values Accurately with Specialist Language Models},
journal = {Proceedings of the VLDB Endowment},
volume = {19},
number = {11},
year = {2026}
}Released under the Apache License 2.0.