Ontological knowledge underlies the performance of large language models (LLMs) on knowledge‑intensive tasks, yet existing evaluations focus on behavioral correctness without exposing the internal mechanisms. To address this gap we introduce MIOE (Mechanistic‑Abstraction Ontology Engineering), a causal‑abstraction benchmark that employs interchange interventions for two ontology engineering tasks: entity typing and subsumption prediction. MIOE requires that learned internal intervention targets modify the ontology‑level answer while preserving class‑specific auxiliary information. We evaluate SAE, DBM, and DAS‑based methods on Qwen3‑8B and Qwen3‑4B, including multi‑task and multi‑layer variants. Distributed subspace methods achieve the strongest results and generalize across prompt settings, but the induced changes remain only partially disentangled from auxiliary class information. Multi‑task objectives and contiguous layer‑window interventions improve this balance, and class‑level analyses reveal substantial variation across ontology classes. These findings suggest that ontological predictions are mechanistically accessible in LLMs, although current approaches have not yet isolated a cleanly disentangled causal mechanism for ontological knowledge.
Review