Spreadsheets and database tables are still where most business data lives, and three jobs keep coming up with them: predicting a missing target such as churn, filling in blank cells, and generating synthetic rows that can be shared without exposing real customers. Each job has its own tools and benchmarks. A new preprint asks whether one model can now do all three, and its answer is a careful "partly".
The One Table Benchmark, or OTBench, from researchers at StableAI, Tsinghua University and other Chinese universities, covers 56 datasets, 46 methods and 20 capability axes. The key design choice is that all three tasks use the same data splits, so a method's results on prediction, imputation and generation can be compared directly rather than across unrelated test sets. The team also trained five model backbones under every combination of the three training objectives, on five datasets with three random seeds, to see whether learning one task helps another.
The findings split along a clear line. Prediction and imputation tend to rank methods similarly, and generation lines up with both when synthetic data is judged by how useful it is for training a downstream model. But when synthetic data is judged by how faithfully it reproduces the original distribution, the rankings diverge and sometimes reverse. Some generalist models beat portfolios of specialist tools on usefulness, while specialists still win on fidelity. Recent tabular foundation models, used frozen with no per-dataset fine-tuning, did well at prediction and at filling in numbers: LimiX-2 led every prediction axis shown, and TabPFN-3 led numerical point accuracy for imputation. No single model was best across all three profiles.
The training experiments add a practical warning. Training on a task's own objective gave the biggest gains, as expected, and generation had the widest positive side effects on other tasks. But benefits from extra objectives did not simply add up, so combining them is not a free win.
For data teams, the takeaway is concrete: a single foundation model can be a sensible default when you need predictions or useful synthetic data, but if your synthetic data must stand in statistically for the original, for example in a privacy-sensitive release, a specialist generator is still the safer choice. Note that the lead authors are affiliated with StableAI and the code sits under a GitHub organisation named after LimiX, the model that topped prediction, so independent replication would help. It is a preprint, not peer reviewed.