← News

Models & toolsSources: arXiv · 6 Nov 2026

Study maps which rows to label when adapting tabular foundation models

An arXiv paper examines active in-context learning for tabular foundation models, testing how diversity, uncertainty and target-aware information gain guide which rows are worth labeling.

In plain words

Tabular foundation models are machine-learning models that adapt to a new task using only labeled examples placed in their input context, without retraining. Because labels are costly, the authors study which unlabeled rows a model should ask to have labeled next. Their answer: pick rows near the target inputs and those with high expected information gain.

What was announced

The paper, posted as an abstract on arXiv, addresses a practical constraint: tabular foundation models (TFMs) adapt to a new task only through labeled rows in their context, and acquiring those labels is expensive because high-quality data requires specific domain knowledge. The authors transfer classical acquisition rules to this setting, where the model guiding acquisition changes with every label added, and ask how diversity, predictive uncertainty and target-aware information gain change in value as a frozen TFM acquires more context.

They introduce two acquisition rules, Target-Aware Nearest Neighbor (TANN) and Expected Predictive Information Gain in Context (EPIC), and evaluate them against classical rules across twenty benchmarks using two established TFMs and gradient-boosted decision trees (GBDTs). Three findings are reported: diversity-based acquisition works best in the first iterations; selecting pool rows that serve the target inputs via a simple nearest-neighbor rule matters substantially; and EPIC gives the best models at the label budget, retaining on average 80–87% of chance-corrected full-data performance with only 5–10% of the pool labeled on the eight benchmarks with the largest pools. EPIC-selected labels also improved GBDT performance.

What it does not show

  • The evidence is the paper's abstract only; full methods, experimental details and caveats are not available for verification.
  • The reported performance figures come from the authors' own benchmarks and have not been independently confirmed.
  • The 80–87% retention figure applies only to the eight benchmarks with the largest pools, not all twenty.

For security and IT teams

Scope
The study covers tabular foundation models adapting via labeled context rows, evaluated on twenty benchmarks alongside GBDTs, according to the abstract.

Sources

Source-support review completed 13 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 13 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.