Search papers, labs, and topics across Lattice.
This paper introduces RACE, a novel statistical framework designed to assess the functional consistency of Transformer neurons across entire domains, addressing the limitations of existing methods that either focus on instance-level estimates or are computationally intensive. By employing perturbation experiments, RACE demonstrates enhanced domain specificity compared to traditional gradient-based approaches, while also confirming the relevance of selected neurons to their target domains through token-distribution-level analysis. Notably, RACE operates with significantly reduced computational overhead, achieving efficiency two orders of magnitude better than gradient-based methods.
RACE reveals that a scalable statistical approach can vastly improve our understanding of neuron behavior in LLMs while slashing computational costs.
Discovering stable neuron behavior across entire domains remains a challenge in mechanistic interpretability. Existing methods often rely on instance-level point estimates or computationally expensive procedures, which either obscure population-level variability or limit scalable domain-wide analysis. We present RACE (Residual Alignment for Consistency Estimation), a forward-pass statistical framework that evaluates the domain-wide functional consistency of Transformer neurons. Perturbation experiments demonstrate that RACE achieves superior domain specificity compared to gradient-based point estimates. Meanwhile, token-distribution-level results verify the association between the selected neurons and the target domain. Furthermore, its computational overhead is two orders of magnitude lower than that of gradient-based methods.
