Mechanistic interpretability asks which components of a language model matter for a given behavior. Activation patching answers this by changing one component's activation and measuring the effect, but the cost grows with the number of components, making it expensive for large models.
The new paper, "Correcting Attribution Patching with a Few Exact Measurements," proposes a correction method that uses only a few exact measurements to improve attribution patching. This could reduce the computational burden while maintaining accuracy.
The abstract is brief, so details of the correction technique are not yet available. The paper is posted on arXiv and appears to target a key bottleneck in interpretability research.