Prune, Interpret, Evaluate: A Cross-Layer Transcoder-Native Framework for Efficient Circuit Discovery via Feature Attribution
For mechanistic interpretability researchers, this provides an efficient method to reduce the cost of interpreting CLT features by pruning irrelevant ones without sacrificing fidelity.
The paper introduces PIE, a framework for pruning cross-layer transcoder (CLT) features using Feature Attribution Patching (FAP), achieving ~40x compression (100 vs 4000 features) while maintaining behavioral fidelity on IOI and Doc-String tasks with Llama-3.2-1B and Gemma-2-2B models.
Existing feature-interpretation pipelines typically operate on uniformly sampled units, but only a small fraction of cross-layer transcoder (CLT) features matter for a target behavior, with the rest resulting in expensive feature explaining and evaluating costs. We introduce the first CLT-native end-to-end framework, PIE, connecting Pruning, automatic Interpretation, and interpretation Evaluation, enabling systematic measurement of behavioral fidelity and downstream interpretability under pruning. To achieve this, we propose Feature Attribution Patching (FAP), a patch-grounded attribution method that scores CLT features by aggregating gradient-weighted write contributions, and FAP-Synergy, a synergy-aware reranking procedure. We evaluate pruning using KL-divergence behavior retention and assess interpretation quality with FADE-style metrics. Across IOI and Doc-String, across budgets $K \in \{50, 100, 200, 400, 800\}$, and across FAP, FAP-Synergy, Activation-Magnitude, and ACDC-style pruning, the FAP family consistently achieves the best or near-best fidelity, with FAP-Synergy providing its clearest gains in strict-budget regimes. On IOI with CLTs for Llama-3.2-1B and Gemma-2-2B, pruning to $K=100$ features matches the KL fidelity that random selection from the active feature set requires $\approx 4$k features to achieve ($\approx 40\times$ compression), enabling $\approx 40\times$ fewer interpretation/evaluation calls while substantially reducing low-quality features.