arXiv:2605.03058v2 Announce Type: replace 
Abstract: A central goal of explainable AI is to express large language model (LLM) decision logic symbolically and ground it in internal mechanisms. Existing rule-extraction methods usually learn ungrounded symbolic surrogates, while mechanistic interpretability links behavior to neurons but often requires hand-crafted hypotheses and costly interventions. We introduce MechaRule, a pipeline that grounds rule extraction in LLM circuits by localizing sparse agonist activations whose ablation disrupts rule-related behavior. MechaRule rests on two findings. First, in a fixed baseline/flip regime, sparse agonist effects can exhibit overtopping: a few high-effect activations remain detectable within larger groups, dominate weaker ones, and flip many of the same examples. In such regimes, adaptive group testing with confidence-guided conservative pruning requires O(k log(N/k) + k) interventions over N candidates when k << N are agonists. Second, agonists are localized more reliably on data splits aligned with close-to-faithful rule behavior; spectral splits provide a rule-free fallback, whereas unfaithful splits degrade localization. Empirically, on arithmetic and jailbreaking, MechaRule recalls 97.0% of highest-effect agonists in matched brute-force validations at only 2.14% of exhaustive-ablation cost on average. Ablating the localized agonists eliminates 97.6--100.0% of eligible correct arithmetic answers and jailbreaks, and can correct arithmetic errors or induce jailbreaks by up to 72.8% and 32.5%.

تم تقديم نهج جديد يسمى MechaRule لاستخراج القواعد في نماذج اللغة الكبيرة (LLMs)، مع التركيز على ربط منطق القرار الرمزي بالآليات الداخلية من خلال تنشيطات أغونست المحلية. تهدف هذه الطريقة إلى تحسين القابلية للتفسير الميكانيكي من خلال تحديد التنشيطات العصبية التي تؤثر على السلوك المتعلق بالقواعد، مما يعالج قيود البدائل الرمزية غير المرتبطة.

Se ha introducido un nuevo enfoque llamado MechaRule para la extracción de reglas en grandes modelos de lenguaje (LLM), centrándose en anclar la lógica de decisión simbólica en mecanismos internos a través de activaciones agonistas localizadas. Este método busca mejorar la interpretabilidad mecanicista al identificar activaciones neuronales que influyen en el comportamiento relacionado con las reglas, abordando las limitaciones de los sustitutos simbólicos no anclados.

Une nouvelle approche appelée MechaRule a été introduite pour l'extraction de règles dans les grands modèles de langage (LLM), se concentrant sur l'ancrage de la logique décisionnelle symbolique dans les mécanismes internes grâce à des activations agonistes localisées. Cette méthode vise à améliorer l'interprétabilité mécaniste en identifiant les activations neuronales qui influencent le comportement lié aux règles, abordant les limitations des substituts symboliques non ancrés.

A new approach called MechaRule has been introduced for rule extraction in large language models (LLMs), focusing on grounding symbolic decision logic in internal mechanisms through localized agonist activations. This method aims to enhance mechanistic interpretability by identifying neuron activations that influence rule-related behavior, addressing limitations of existing ungrounded symbolic surrogates.