Pando Eval

pando-eval

Pando: Do Interpretability Methods Work When Models Won't Explain Themselves? — Zhong et al. (2026) (arXiv:2604.11061, 2026)

What this evaluates

Evaluates whether mechanistic interpretability methods can recover decision-relevant signals from black-box models that lack faithful explanations. It probes the ability of gradient-based, representation-based, and black-box elicitation agents to predict held-out outcomes and identify correct decision-rule fields across varying explanation qualities and model complexities.

Datasets

• Pando — total 720; splits: test (-1)

Metrics

• Held-out accuracy (%) (primary) — range: percent
  • Percentage of correctly predicted held-out outcomes across the test set.
• Decision-rule field F1 (%) — range: percent
  • F1 score computed over the set of correctly identified decision-rule fields (features) used by the model.

Input / output format

Input: A decision-tree model instance (or its internal states/activations) along with a scenario prompt (e.g., car purchase, movie selection, policy violation) and optional explanation text (faithful, unfaithful, or none).

Output: A predicted decision rule (set of field-value pairs) or a classification label for the held-out instance.

Scoring recipe

def compute_accuracy(preds, golds):
    return sum(p == g for p, g in zip(preds, golds)) / len(golds) * 100

def compute_field_f1(pred_fields, gold_fields):
    tp = len(pred_fields & gold_fields)
    fp = len(pred_fields - gold_fields)
    fn = len(gold_fields - pred_fields)
    prec = tp / (tp + fp) if (tp + fp) > 0 else 0
    rec = tp / (tp + fn) if (tp + fn) > 0 else 0
    return (2 * prec * rec / (prec + rec)) * 100 if (prec + rec) > 0 else 0

Common pitfalls

• Assuming white-box interpretability methods always outperform black-box elicitation; the benchmark shows prefill baselines actually exceed white-box methods when explanations are faithful.
• Relying solely on end-to-end accuracy to judge interpretability quality; field F1 reveals that gradient methods capture significantly stronger feature-level signals that accuracy masks due to threshold recovery noise.
• Ignoring task representation bias; many methods (logit lens, SAE) score well on accuracy but are dominated by field identity/value rather than actual decision relevance.

Evidence (verbatim from paper)

We report held-out accuracy (Table 3(a)) and decision-rule field F1 (Table 3(b)) aggregated across three scenarios (car purchase, movie selection, and policy violation) and three explanation setups (no explanation, faithful, and unfaithful).

Citation

@misc{zhong2026pando,
  title={Pando: Do Interpretability Methods Work When Models Won't Explain Themselves?},
  author={Zhong et al. (2026)},
  year={2026},
  note={arXiv:2604.11061}
}

• arXiv: 2604.11061