Entity · paper

Faithfulness to Refusal: A Causal Audit of Neuron Selectors

paperactiveprovisionalfaithfulness-to-refusal-a-causal-audit-of-neuron-selectors-daec8261·1 events·first seen Jul 7, 2026

Aliases: Faithfulness to Refusal: A Causal Audit of Neuron Selectors

More like this (12)

Faithfulness Evaluation Neuron Populations Exhibit Divergent Selectivity with Scale Relaxing Faithfulness with Intervention-Only Causal Discovery Selection Without Signal, Recovery Through Expression: A Measurement Study of Post-Hoc Falsification Operators for Frozen Small Code Models The Blind Curator: How a Biased Judge Silently Disables Skill Retirement in Self-Evolving Agents Neuron-Aware Data Selection for Annotation-Free LLM Self-Distillation RAS: Measuring LLM Safety Through Refusal Alignment metacognitive data selection Neural Safety Filters Observe-and-Act Adaptive Context Selection Neuronal Circuit Policies (NCPs)Neuronal Stochastic Attention Circuit (NSAC)

Recent events (1)

6arXiv · cs.CL·Jul 7, 2026·source ↗

Causal audit of neuron selectors reveals refusal behavior lives in redundant subspace

A new arXiv preprint introduces a causal auditing framework for neuron-row attribution methods in LLMs, using one-shot zeroing interventions to test whether high-scoring neurons are genuinely causally important. The authors find that attribution methods outperform activation/magnitude baselines for identifying dispensable rows, and that refusal behavior (for hate/crime prompts) can be installed or removed via attributed rows while preserving fluency. A key finding is that refusal occupies a redundant subspace — different attribution methods identify largely disjoint sufficient row sets — meaning no single method recovers a unique causal mechanism. Rank-stability, a common proxy for selector quality, is shown to be a poor indicator of causal validity.

Evaluation and Benchmarking AI Safety Research Faithfulness to Refusal: A Causal Audit of Neuron Selectors +1 more