When Does Bigger Help? A Controlled Study of LLM Scale for Ontology Learning
A controlled ontology-learning study finds model size is a weak selector on its own. Dense 27B models beat larger sparse models on one task, while MoE models led another.
The study controls retrieval, prompts, decoding, datasets, and metrics across **13 models** and four ontologies. Dense Qwen3.5 gains were largest from **9B to 27B**, mainly improving precision rather than recall.
Choose models against the exact extraction task, not parameter count. Dense 27B variants led larger sparse models on term typing, while **Mixture-of-Experts models** produced the strongest open-weight taxonomy-discovery results.
The study controls retrieval, prompts, decoding, datasets, and metrics across **13 models** and four ontologies. Dense Qwen3.5 gains were largest from **9B to 27B**, mainly improving precision rather than recall. Choose models against the exact extraction task, not parameter count. Dense 27B variants led larger sparse models on term typing, while **Mixture-of-Experts models** produced the strongest open-weight taxonomy-discovery results. Scaling was neither monotonic nor consistent across domains. Non-taxonomic relation extraction remained difficult, and the material gives no latency or cost comparison for the tested models.
This replaces a general “bigger is better” model-selection rule with task- and ontology-specific evidence: dense scaling mainly raised precision, larger sparse models did not always win, and MoE leadership was confined to taxonomy discovery. It reinforces prior calls for matched workload evaluation while leaving deployment decisions incomplete because latency and cost were not measured.