Discovering Transformer Circuits via a Hybrid Attribution and Pruning Framework

2025-09-28

AI Summary

Existing circuit discovery methods face a trade-off between efficiency and fidelity: attribution patching is efficient but deviates from the full model, whereas edge pruning is faithful yet computationally expensive. This paper proposes the Hybrid Attribution-and-Pruning (HAP) framework, which first employs attribution patching to rapidly identify high-contribution subgraphs, then applies edge pruning to extract faithful, sparse Transformer circuits from them. HAP jointly leverages both techniques—marking the first such integration in circuit discovery—thereby preserving cooperative components (e.g., S-inhibition attention heads) while balancing speed and fidelity. Experiments on indirect object identification demonstrate that HAP achieves a 46% speedup over baseline methods while retaining critical functional circuit structures. The implementation is publicly available.

Technology Category

• Machine Learning: Probabilistic Circuits and Graphical Models
• Computer Vision: Adversarial Attacks & Robustness
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Data transparency and provenance

Abstract

Interpreting language models often involves circuit analysis, which aims to identify sparse subnetworks, or circuits, that accomplish specific tasks. Existing circuit discovery algorithms face a fundamental trade-off: attribution patching is fast but unfaithful to the full model, while edge pruning is faithful but computationally expensive. This research proposes a hybrid attribution and pruning (HAP) framework that uses attribution patching to identify a high-potential subgraph, then applies edge pruning to extract a faithful circuit from it. We show that HAP is 46% faster than baseline algorithms without sacrificing circuit faithfulness. Furthermore, we present a case study on the Indirect Object Identification task, showing that our method preserves cooperative circuit components (e.g. S-inhibition heads) that attribution patching methods prune at high sparsity. Our results show that HAP could be an effective approach for improving the scalability of mechanistic interpretability research to larger models. Our code is available at https://anonymous.4open.science/r/HAP-circuit-discovery.

Problem

• Develops hybrid framework for faster faithful circuit discovery in transformers
• Resolves trade-off between speed and faithfulness in model interpretation
• Improves scalability of mechanistic interpretability for larger language models

Innovation

• Hybrid framework combines attribution and pruning
• Uses attribution patching to identify high-potential subgraph
• Applies edge pruning to extract faithful circuits