Faithfulness to Refusal: A Causal Audit of Neuron Selectors

2026-07-06 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

A recent causal audit, detailed in paper 2607.05355 by Seth, Avaiya, Sankarapu, and Eswar, directly tests the causal importance of neuron rows identified by attribution scores in large language models. The study employs two paired audits based on one-shot neuron-row zeroing. First, at the language-modeling level, attribution methods significantly surpassed activation and magnitude-based baselines in identifying dispensable rows across five distinct LLMs. Second, adapting this intervention into a behavior test with a contrastive harmful-versus-benign signal, the researchers found that attributed rows were sufficient to install refusal behaviors for hate and crime content while maintaining low benign over-refusal and preserving language model fluency. Crucially, layer-matched random controls failed to achieve this. The audit also revealed that highly rank-stable selectors can be among the least causally valid, and that refusal mechanisms reside in a redundant subspace, meaning different attribution methods achieve refusal through largely disjoint sets of neuron rows. This suggests that any recovered edit represents one sufficient set rather than a unique underlying mechanism.

Key takeaway

For Machine Learning Engineers evaluating neuron attribution methods or implementing safety mechanisms in LLMs, you should prioritize direct causal audits over proxy metrics like rank-stability. Your safety interventions, such as installing refusal behaviors, may achieve their goal through redundant neuron sets, meaning the specific "edit" is one of many sufficient realizations. This necessitates robust testing beyond simple attribution scores to ensure true causal efficacy and avoid over-reliance on potentially misleading metrics.

Key insights

Causal audits reveal attribution methods' true efficacy and the redundant nature of refusal mechanisms in LLMs.

Principles

• Rank-stability does not imply causal validity.
• Refusal mechanisms often reside in redundant subspaces.
• Direct causal audits are crucial for interpretability.

Method

The study employs one-shot neuron-row zeroing, applying it at both language-modeling and behavior-testing levels, driven by contrastive harmful-versus-benign signals.

In practice

• Evaluate neuron selectors with causal audits.
• Test refusal mechanisms using contrastive signals.
• Do not rely solely on rank-stability metrics.

Topics

• Neuron Attribution
• Causal Audit
• LLM Interpretability
• Model Safety
• Refusal Mechanisms
• Neuron Zeroing

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activations — LLM refusal behavior is linearly decodable from intermediate activations, offering an actionable signal for safety.
• Project Ariadne: A Structural Causal Framework for Auditing Faithfulness in LLM Agents — LLM agent reasoning traces often act as "Reasoning Theater," not faithful generative drivers, due to Causal Decoupling.
• Minimal, Local, Causal Explanations for Jailbreak Success in Large Language Models — LOCA provides minimal, local, and causal explanations for LLM jailbreak success by identifying key intermediate representation changes.
• Trust, but Don't Verify: Epistemic Blind Spots in LLM Source Evaluation — LLMs detect statistical fabrications in isolation but ignore them during multi-source synthesis, prioritizing stylistic credibility.
• Leverage Is Not Reach: A Control-Window Law for Single-Neuron Steering in Language Models — A "control-window law" predicts single-neuron steering coherence in LLMs, distinguishing effective interventions from output collapse.
• LLMs believe false statements even after explicit warnings that they're false — LLMs prioritize statistical patterns in training data over explicit negation, leading to "belief implantation."

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.