When Attribution Patching Lies: Diagnosis and a Second-Order Correction

A new study finds that attribution patching, a widely used shortcut for identifying which components drive a language model’s behavior, is systematically unreliable due to downstream network non-linearities, and proposes a computationally cheap second-order correction to fix it [1]. Mechanistic interpretability researchers rely on importance estimates to map the internal circuits of language models. The gold-standard method, activation patching, is computationally prohibitive at scale, so practitioners have adopted attribution patching, a gradient-based, first-order approximation [1]. The reliability of this shortcut has remained poorly understood until now [2]. The authors of the paper, posted to arXiv on June 5, 2026, demonstrate that the dominant error in attribution patching does not come from local curvature at the component being patched, but from non-linearities in the downstream network [1]. This misdiagnosis has allowed systematic errors to propagate through circuit-identification work, potentially leading to the misidentification of underlying mechanisms [2]. To address this, the researchers introduce three tools. The first is a reliability score that flags untrustworthy estimates. The second provides error bounds that quantify potential attribution mis-specifications. The third is a Hessian-vector-product correction, which eliminates the leading-order error with only one additional backward pass [1]. The HVP correction is described as the only second-order correction feasible at larger scale, where standard baselines such as Integrated Gradients become computationally prohibitive [2]. The evaluation covered five model families ranging from 124 million to 9 billion parameters, tested under both random-token and naturalistic name-swap perturbations [1]. In comparative experiments, a multi-step variant of the HVP correction matched or exceeded the accuracy of Integrated Gradients while using significantly lower compute, outperforming prior second-order baselines [2]. The improvements translated into higher-fidelity circuit recovery on standard benchmarks. The authors propose a “Screen-Flag-Fix” workflow that targets computational effort only toward components flagged as unreliable, rather than applying expensive corrections uniformly across a model [1]. The work was developed within arXivLabs, a framework for community collaborators, and associated code and media were made available through Hugging Face [1].