SkepticalSam

TabPFN Shapley Values — Speedup Irrelevance

Computing Conditional Shapley Values Using Tabular Foundation Models

Read the original source

What the paper says

TabPFN enables fast conditional Shapley value estimation with significant speedup.

The Critique

The speedup claim is compelling but they don't address whether it matters in practice. Shapley values are typically computed once per model, not in a tight loop. For tabular datasets, even slow Monte Carlo methods finish quickly. They also don't discuss the fundamental limitation: TabPFN has quadratic complexity in context length, limiting it to small datasets.

Why It Matters

If the speedup is irrelevant for practical use cases and the approach doesn't scale, this is an academic exercise rather than a practical improvement. The field needs methods that work for real-world dataset sizes.

What They Missed

For large tabular datasets, the approach may fail entirely. No validation that explanations are actually useful for model debugging.