A study audits neuron selectors in language models using one-shot neuron-row zeroing to determine if attribution scores identify causally important rows. The research demonstrates that attribution methods substantially outperform activation and magnitude-based baselines at identifying dispensable rows across five LLMs.
- Attributed rows are sufficient to install refusal on hate and crime topics while keeping benign over-refusal low and preserving fluency.
- Specific layer-matched random controls at the same depths fail to produce this effect, confirming causal validity.
- Highly rank-stable selectors can be among the least causally valid according to the findings.
- Refusal resides in a redundant subspace where different attribution methods install it through largely disjoint row sets.
The authors conclude that rank-stability proxies miss selector failures that direct causal audits can surface, showing the recovered edit is one realization of a sufficient set rather than a unique mechanism.