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

DevFeed: [When Attribution Patching Lies: Diagnosis and a Second-Order Correction](<https://devfeed.tech/articles/when-attribution-patching-lies-diagnosis-and-a-second-order-correction-120049.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2606.09899>)

Author: Luyang Zhang, Jialu Wang

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>), [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Language models](<https://devfeed.tech/topics/language-models.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>)

Tags: [approximation](<https://devfeed.tech/tags/approximation.md>), [circuit](<https://devfeed.tech/tags/circuit.md>), [compute](<https://devfeed.tech/tags/compute.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [error](<https://devfeed.tech/tags/error.md>), [gradient](<https://devfeed.tech/tags/gradient.md>), [mechanistic-interpretability](<https://devfeed.tech/tags/mechanistic-interpretability.md>), [patching](<https://devfeed.tech/tags/patching.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [scale](<https://devfeed.tech/tags/scale.md>), [tools](<https://devfeed.tech/tags/tools.md>), [vector](<https://devfeed.tech/tags/vector.md>), [workflow](<https://devfeed.tech/tags/workflow.md>)

## AI overview

Attribution patching can misidentify causal components because its main error comes from nonlinearities in the downstream network, rather than local curvature at the patched component. The work proposes a reliability score and error bounds to identify risky estimates, plus a Hessian-vector-product (HVP) correction that removes the leading-order error with one additional backward pass. Across five model families (124M-9B parameters) and random-token and name-swap perturbations, HVP is reported as the only second-order correction feasible at larger scale. A multi-step HVP variant matches or exceeds Integrated Gradients accuracy at lower compute and outperforms prior second-order baselines. The resulting Screen-Flag-Fix workflow directs extra computation to components flagged as unreliable.

## Source excerpt

arXiv:2606.09899v2 Announce Type: replace-cross Abstract: A central goal of mechanistic interpretability is to identify which internal components causally drive a language model's behavior. Because these importance estimates serve as the evidence for identifying circuits, systematic errors can lead to the misidentification of the underlying mechanisms. While activation patching provides a gold-standard causal metric, its computational cost is prohibitive at scale. Practitioners instead rely on attribution patching, a gradient-based, first-order approximation whose reliability remains poorly understood. In this work, we characterize the source of this unreliability, demonstrating that the dominant error stems from the non-linearities in the downstream network rather than local curvature at the patched component. This insight yields three practical tools: (i) a reliability score to detect untrustworthy estimates, (ii) error bounds quantifying potential attribution mis-specifications, and (iii) a Hessian-vector-product (HVP) correction that eliminates the leading-order error with only one additional backward pass. In evaluations across five model families (124M--9B parameters) and both random-token and naturalistic (name-swap) perturbations, HVP is the only second-order correction feasible at larger scale, where standard baselines like Integrated Gradients become computationally prohibitive. In comparative experiments, a multi-step HVP variant matches or exceeds the accuracy of Integrated Gradients at significantly lower compute, outperforming prior second-order baselines. These improvements lead to higher-fidelity circuit recovery on standard benchmarks and support a Screen-Flag-Fix workflow that targets computational effort only toward the components flagged as unreliable.