Pando: Do Interpretability Methods Work When Models Won't Explain Themselves?

TL;DR

Pando is a benchmark that tests interpretability methods on models with controlled explanations, revealing when white-box tools outperform black-box elicitation and providing insights into model decision processes.

Contribution

Introduces Pando, a benchmark with models trained to produce faithful, no, or unfaithful explanations, to evaluate interpretability methods under different explanation conditions.

Findings

When explanations are faithful, black-box elicitation matches white-box methods.

Gradient-based attribution improves accuracy when explanations are absent or misleading.

Relevance patching (RelP) yields the largest gains among interpretability tools.

Abstract

Mechanistic interpretability is often motivated for alignment auditing, where a model's verbal explanations can be absent, incomplete, or misleading. Yet many evaluations do not control whether black-box prompting alone can recover the target behavior, so apparent gains from white-box tools may reflect elicitation rather than internal signal; we call this the elicitation confounder. We introduce Pando, a model-organism benchmark that breaks this confound via an explanation axis: models are trained to produce either faithful explanations of the true rule, no explanation, or confident but unfaithful explanations of a disjoint distractor rule. Across 720 finetuned models implementing hidden decision-tree rules, agents predict held-out model decisions from $10$ labeled query-response pairs, optionally augmented with one interpretability tool output. When explanations are faithful, black-box elicitation matches or exceeds all white-box methods; when explanations are absent or misleading, gradient-based attribution improves accuracy by 3-5 percentage points, and relevance patching, RelP, gives the largest gains, while logit lens, sparse autoencoders, and circuit tracing provide no reliable benefit. Variance decomposition suggests gradients track decision computation, which fields causally drive the output, whereas other readouts are dominated by task representation, biases toward field identity and value. We release all models, code, and evaluation infrastructure.