Preprint Introduces Interpretable-by-Design Tabular Predictions Using In-Context Learning
A new arXiv preprint presents TICL³, which combines a shared linear model with local corrections guided by a tabular foundation model. This lets the same coefficients both…
- TICL³ learns a global linear model then fits a second one on residuals from similar samples retrieved via the foundation model, producing query-specific explanations from the prediction coefficients themselves.
- Across 386 classification benchmarks it trails TabICLv2 by 1.1 percentage points, outperforms CatBoost in 266 evaluations, and improves average support F1 over MAPLE by 4.7 points.
- The method is positioned as more reliable than post-hoc interpretability techniques that may not reflect what the model actually uses.
What the preprint proposes
An arXiv preprint dated October 29, 2026, by Ran Eisenberg, Steven Berger, and Ofir Lindenbaum introduces TICL³ (Tabular In-Context Local Linear Learning). The approach aims to give tabular foundation models built-in interpretability during in-context learning without any dataset-specific training or fine-tuning.
It first fits a shared linear model on the full dataset to capture broad feature effects. For each new query the method uses representations from the foundation model to identify similar training examples, then fits a second linear model on the residuals in that local neighborhood. The combined coefficients from both models determine the final prediction and the contribution of each feature for that specific query.
Reported benchmark performance
The abstract states that across 386 classification benchmark-dataset evaluations, TICL³ trails TabICLv2 by an average of 1.1 percentage points. It outperforms CatBoost in 266 of those evaluations and improves average support F1 over MAPLE by 4.7 points.
The authors also report that TICL³ achieves the best performance on real-data selected-feature utility metrics. They note potential uses in biology, such as identifying biomarkers or risk genes, where clear feature contributions are valuable. Because only the abstract is available, full experimental details, code, or exact conditions are not yet public.
Why it matters for uncensored AI
Our analysis: The preprint starts from the fact that post-hoc interpretability methods can assign feature importance scores that do not match what the model actually uses and can produce explanation errors. For people using tabular foundation models, an interpretable-by-design method that derives explanations directly from the prediction rule offers a practical way to understand individual feature influence without separate tools.
This does not change model refusals, content policies, or companion features. It does, however, add to work on making foundation-model predictions more transparent, which can help users who want to audit outputs themselves in private or locally run setups. Greater transparency in tabular tasks supports the broader goal of user control over AI behavior rather than relying on black-box systems.
