Why It Failed: A Benchmark to Evaluate Interpretability

Joel Mathew, Aditya Lagu, Anthony Tang, Prudhviraj Naidu

Keywords: Failure prediction, Interpretability, Benchmarking, AI governance

Corresponding Author Name: Joel Mathew

TL;DR: We introduce the Why It Failed benchmark to evaluate interpretability methods on model failure prediction

Corresponding Author Email: [e-mail]

Abstract: We introduce *Why It Failed*, a benchmark for evaluating whether interpretability methods can explain model failures. We test last token logistic probes on Gemma-2 2B across four basic reasoning tasks and find they fail to predict model failures, achieving near-chance performance across all tasks. Our benchmark provides a standardized framework to evaluate whether interpretability methods can explain model failures. Lastly, we motivate the AI community to move beyond reporting quantitative metrics and seek explanations of when and why models fail.

Submission Number: 33