Back
3770855.3818091_compressed
Conference paper (in proceedings)

Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation

  • 2026
Published in:
  • Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’26), August 09–13, 2026, Jeju Island, Republic of Korea. - New York : ACM - Association for Computing Machinery. - 2026, vol. 2, p. 4369-4380
Explainable AI
Rule extraction
Mechanistic interpretability
Contrastive hierarchical ablation
Neuron activation analysis
Large Language Models
English 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%.
Collections
Language
  • English
Classification
Computer science and technology
Notes
  • KDD '26: The 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining
  • Jeju Island Republic of Korea
  • August 9-13, 2026
License
CC BY
Open access status
gold
Identifiers
Persistent URL
https://n2t.net/ark:/12658/srd1336637
Statistics
Document views: 1 File downloads:
  • 3770855.3818091_compressed: 0