--- title: "Attribution Patching Outperforms Automated Circuit Discovery" description: "The authors propose Edge Attribution Patching (EAP). 2 references, cited by 69 papers here." authors: ["Aaquib Syed", "Can Rager", "Arthur Conmy"] year: 2023 venue: "arXiv:2310.10348" source: "https://arxiv.org/abs/2310.10348" doi: "10.48550/arxiv.2310.10348" arxiv: "2310.10348" licence: "CC BY 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/syed-2023-attribution-patching-outperforms-automated-circuit/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/syed-2023-attribution-patching-outperforms-automated-circuit/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/9/) / Attribution Patching Outperforms Automated Circuit Discovery # Attribution Patching Outperforms Automated Circuit Discovery **Authors:** [Aaquib Syed](https://md.research.elamur.ai/free/en/authors/aaquib-syed/), [Can Rager](https://md.research.elamur.ai/free/en/authors/can-rager/) and [Arthur Conmy](https://md.research.elamur.ai/free/en/authors/arthur-conmy/) · **Year:** 2023 · **Venue:** arXiv:2310.10348 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2310.10348) · [PDF](https://arxiv.org/pdf/2310.10348) · [DOI 10.48550/arxiv.2310.10348](https://doi.org/10.48550/arxiv.2310.10348) · [arXiv 2310.10348](https://arxiv.org/abs/2310.10348) · [OpenAlex](https://openalex.org/W4404783717) **Citations:** 12 citations (OpenAlex, 2026-10-14) · [69](https://md.research.elamur.ai/free/en/papers/syed-2023-attribution-patching-outperforms-automated-circuit/cited-by/) in our library, of the 149,439 papers whose bibliographies we read ## Summary Mechanistic interpretability tries to find circuits, which are small subgraphs of a network that account for a specific behavior. Automated tools such as ACDC rely on activation patching. That method needs an extra forward pass for each edge it tests, so it scales poorly to large models. The authors propose Edge Attribution Patching (EAP). It replaces exact patching with a first-order Taylor estimate of how the task metric changes when an edge's clean activation is swapped for a corrupted one. Two forward passes and one backward pass yield scores for every edge. Absolute scores are averaged over clean/corrupted prompt pairs, edges are ranked, and the top k are kept. The authors benchmarked EAP against published circuits for IOI, Greater-Than and Docstring. EAP beat ACDC in ROC AUC on the first two tasks and was nearly tied on Docstring, at far lower cost. However, EAP scores match true patching effects only loosely. Running ACDC on the EAP-pruned graph improves results further, which makes cheap circuit discovery a realistic option for larger models. ## Key points - **ROC AUC results:** - IOI: EAP 0.904 vs. ACDC 0.868 (KL divergence) and 0.588 (task metric). - Greater-Than: EAP 0.896 vs. 0.849 and 0.458. - Docstring: EAP 0.976 vs. ACDC-KL 0.982, activation patching 0.978 and ACDC with task metric 0.972. - **Cost and setup:** - EAP needs a constant number of passes, whereas ACDC's cost grows rapidly with model size. Full activation patching was too expensive to run on GPT-2 small, which the IOI and Greater-Than tasks use. - EAP uses task-specific metrics such as logit difference rather than KL divergence, because the KL gradient is zero on the unablated model. - **Weak approximation:** Attribution and activation patching scores correlate weakly (R² = 0.27, fit slope 0.531). Concave interpolation curves cause EAP to overestimate some edges. - **Recommended pipeline:** Run EAP first, then ACDC on the pruned subgraph; this combination outperformed EAP alone on Docstring. Code is available at github.com/Aaquib111/acdcpp (NeurIPS 2023 ATTRIB Workshop). - **Evaluation caveat:** Benchmarking against human-found circuits is a noisy and limited measure of interpretability. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works · [💻 Software](https://md.research.elamur.ai/free/en/shelves/computer-science/software/) · 119 works 🧭 Topics (1): [🔬 Interpretability](https://md.research.elamur.ai/free/en/topics/interpretability/) · 149 works ## Cite Cite as: Aaquib Syed, Can Rager and Arthur Conmy. 2023. Attribution Patching Outperforms Automated Circuit Discovery. arXiv:2310.10348. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{syed2023attribution, title = {{Attribution Patching Outperforms Automated Circuit Discovery}}, author = {Aaquib Syed, Can Rager and Arthur Conmy}, year = {2023}, doi = {10.48550/arxiv.2310.10348}, eprint = {2310.10348}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2310.10348} } ``` - [← previous Analyzing the Performance of GPT-3.5 and GPT-4 in Grammatical Error Correction](https://md.research.elamur.ai/free/en/papers/coyne-2023-analyzing-performance-gpt-3-5/) - [next → D4: Improving LLM Pretraining via Document De-Duplication and Diversification](https://md.research.elamur.ai/free/en/papers/tirumala-2023-d4-improving-llm-pretraining-document/) - [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/9/) ## References 2 references, 1 of them with a card of ours. ↗ leads to the work. 1. Chris Olah. Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases. 2022. URL : — [How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model](https://md.research.elamur.ai/free/en/papers/hanna-2023-gpt-2-compute-greater-than/) [https://www.transformer-circuits.pub/2022/mech-interp-essay](https://www.transformer-circuits.pub/2022/mech-interp-essay). Chris Olah, Nick Cammarata, Ludwig Schubert, Gabriel Goh, Michael Petrov, and Shan Carter. “Zoom In: An Introduction to Circuits”.… [↗](https://www.transformer-circuits.pub/2022/mech-interp-essay) 2. Atticus Geiger, Hanson Lu, Thomas Icard, and Christopher Potts. Causal Abstractions of Neural Networks. 2021. URL : https://arxiv.org/abs/2106.02997. Kevin Meng, David Bau, Alex J Andonian, and Yonatan Belinkov. “Locating and editing factual associations in GPT”. In: Advances in Neural Information P… [↗](https://arxiv.org/abs/2106.02997) *Updated 2 November 2026*