Faithfulness to Refusal: A Causal Audit of Neuron Selectors
Ananth Eswar, Pratinav Seth, Utsav Avaiya, Vinay Kumar Sankarapu
Read on arXiv →
Key claim
Rank-stability proxies can misidentify causally important neuron rows.
In plain English
Language models often rely on neuron rows for tasks like pruning and safety editing, but current methods may not accurately identify which rows are truly important. This paper tests attribution scores against direct causal audits, revealing that some highly ranked selectors are not causally valid. The findings suggest that different methods can achieve similar safety edits, indicating a need for more nuanced approaches in model interventions. Builders should be aware that relying solely on rank-stability could lead to ineffective or misleading results.
Introduces a direct causal audit for neuron-row importance in LLMs.
Demonstrates findings across multiple LLMs with robust comparisons.
Deep reliability assessment
The methodology supports the claim that attribution methods can identify dispensable neuron rows more accurately than baselines, but the causal importance of these rows in real-world applications is not fully validated.
Reproducibility
No open source code or dataset is mentioned in the paper.
Key figure
Figure 1 likely illustrates the neuron-row zeroing process or the architecture of the transformer model used for the causal audit.