Reproduce “Interpretable by Design Tabular Prediction via In-Context Local Linear Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ran Eisenberg, Steven Berger, Ofir Lindenbaum
TICL³ combines a shared linear rule with a query-specific residual model, using a tabular foundation model’s representation to select neighbors. Their coefficients form the same rule used for class scores and feature contributions. On four controlled tasks, support F1 averaged 0.539 ± 0.010, 4.7 percentage points above MAPLE, though MAPLE led on E4. Contributions explain linear class scores, not final probabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.