TabMI-Bench

A protocol benchmark for mechanistic interpretability (MI) of tabular foundation models (TFMs). NeurIPS 2026 Evaluations & Datasets Track submission.

What's in this dataset

This Hugging Face repository hosts the frozen aggregated artifacts that drive every numbered table and figure in the paper. Bundling these allows reviewers to verify the paper's key numerics without re-running 40 GPU-hours of experiments.

File	Source experiment	Used by
rd5_fullscale_aggregated.json	Phase 5 multi-seed core comparison (5 seeds × 3 models × 4 functions)	Table 1 (3 strategies), Figure 1
tabdpt_probing_3seed.json	TabDPT in-family holdout probing (3 seeds)	Table 1 TabDPT row, §4.3
tabdpt_causal_3seed.json	TabDPT noising-based causal tracing (3 seeds)	§4.4 TabDPT causal claim
nam_holdout.json	NAM out-of-family holdout (5 seeds)	§4.3 NAM holdout boundary case
lofo_primary_endpoint.json	Leave-one-function-out robustness on primary endpoint	LOFO appendix
c1c2_baselines.json	Shuffled-label and random-target negative controls	§4.6, Tables 21–22
scale_10k_multiseed.json	N=10K probing scale validation (3 seeds)	Appendix G
tabpfn25_fullscale_aggregated.json	TabPFN v2 vs v2.5 comparison (3 seeds)	Table 7, §G.1

Code & full benchmark suite

The hooks, synthetic probe generators, evaluation scripts, statistical analysis, figure generation, and tests are hosted at: https://github.com/evaldataset/TabMI-Bench

To regenerate paper-facing tables and figures from this dataset without GPU access:

git clone https://github.com/evaldataset/TabMI-Bench
cd TabMI-Bench
pip install -r requirements.txt
make reproduce-paper-frozen

What is TabMI-Bench?

TabMI-Bench provides:

1. Hook-based activation extraction for 5 TFMs spanning 3 architectural families (TabPFN v2/v2.5, TabICL v2, TabDPT, iLTM) plus NAM out-of-family holdout
2. 4 controlled synthetic probe families (bilinear, sinusoidal, polynomial, mixed) with known ground-truth intermediary variables
3. 4-step evaluation protocol (synthetic profile → causal validation → negative controls → real-world transfer)
4. Evidence-coded MI applicability matrix (8 techniques × 4 architectures with seed-count superscripts)
5. A primary diagnostic finding: whole-layer clean activation patching is uninformative on ICL-style TFMs due to deterministic cascading; corruption-based (noising) tracing is the informative alternative.

Three descriptive reference computation profiles emerge as calibration baselines:

• Staged (TabPFN): U-shaped intermediary recoverability with mid-layer concentration
• Distributed (TabICL, TabDPT): uniformly high recoverability across layers
• Preprocessing-dominant (iLTM): tree+PCA preprocessing performs the heavy lifting

Croissant 1.0 metadata

The repository includes croissant.json with 12 RAI fields (data limitations, biases, sensitive information, use cases, social impact, synthetic data flag, source datasets, provenance, collection, maintenance plan, etc.). See croissant.json in the repository.

Source datasets

Real-world evaluation uses public datasets only (no new collection):

Dataset	Source	License	Use
California Housing	OpenML 8092	CC0	causal tracing + steering
Diabetes	scikit-learn	BSD-3-Clause	causal tracing + steering
Wine Quality	OpenML 287	CC0	steering
Bike Sharing	OpenML 44063	CC0	steering
Abalone, Boston, Energy, Breast Cancer, Iris, Adult, Credit-G	OpenML / scikit-learn	CC0 / BSD-3-Clause	causal tracing

Citation

@inproceedings{anonymous2026tabmibench,
  title={TabMI-Bench: Evaluating Mechanistic Interpretability Methods Across Tabular Foundation Model Architectures},
  author={Anonymous},
  booktitle={Advances in Neural Information Processing Systems (NeurIPS) Evaluations \& Datasets Track},
  year={2026}
}

License

MIT. See LICENSE.