Pando: Do Interpretability Methods Work When Models Won't Explain Themselves?
For researchers in mechanistic interpretability, this work highlights the need to control for elicitation confounders and provides a benchmark to evaluate methods under realistic conditions where models may not explain themselves.
The paper introduces Pando, a benchmark to evaluate mechanistic interpretability methods while controlling for the elicitation confounder. Results show that when model explanations are absent or misleading, gradient-based attribution improves accuracy by 3-5 percentage points and relevance patching (RelP) yields the largest gains, while logit lens, sparse autoencoders, and circuit tracing provide no reliable benefit.
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.