ARXIV:2604.11061 · AI INTERPRETABILITY · SUBMITTED 14 APR · 20:32 UTC · FRESHNESS STALE

VerifiedSource: PDF linkedVerifiedPaperPack: citation fields available

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

Ziqian Zhong · Aashiq Muhamed · Mona T. Diab · Virginia Smith · Aditi Raghunathan · arXiv

Pando is a model-organism benchmark that evaluates AI interpretability methods by controlling for model explanations, revealing that gradient-based attribution and relevance patching offer significant gains when models don't explain themselves.

Ship in 2-4 weeks›Score8.0Evidence verified

Opportunity summary

Pain Pando is a model-organism benchmark that evaluates AI interpretability methods by controlling for model explanations, revealing that gradient-based attribution and relevance patching offer significant gains when models don't explain themselves.

Evidence 0 refs | 4 sources | 67% coverage

Blocker Evidence verified

Open Build Read PDF Signal Canvas Track

PROBLEM

METHOD

Full 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.

RESULT

ScienceToStartup currently rates this 8.0/10 on the public viability pass. 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…

WHY NOW

AI Interpretability moved forward this cycle; last verified April 2026. Public score 8.0/10. Production flags indicate code availability.

Continue into Read for claims, analysis, references, and neighboring papers.

Opportunity summary

Score8.0

PainPando is a model-organism benchmark that evaluates AI interpretability methods by controlling for model explanations, revealing that gradient-based attribution and relevance patching offer significant gains when models don't explain themselves.