Sign InOpen Brain
arXivPaperNeeds Review

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.

arXiv · Aug 31, 2026
Open Source Open MarkdownOpen JSON
Source Summary

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.

Practical Implication

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.

Agent-Ready Context
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.
Connected Context · Feed7 Judgment

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.

Program-as-Weights: A Programming Paradigm for Fuzzy FunctionsProgram-as-Weights likewise shows that parameter count can be a poor proxy for task performance, reinforcing selection by the exact operation rather than general model scale.Hy4 Preview now available on AI GatewayHy4 makes an open MoE route available, but this study shows that MoE advantage may depend on the ontology subtask; availability and context length therefore do not establish suitability.Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAIThe domain- and task-dependent scaling results reinforce the earlier argument that task fit can matter more than simply adding compute, although this study holds retrieval and datasets fixed rather than testing data curation.DeepSeek overtakes Google on volume, cost per token falls 13.6%Gateway routing can reduce blended spend, but the study omits cost and latency; combining its quality results with routing economics is necessary before choosing among the tested model classes.
Context Map
modeldataresearch#model-selection#open-models
Uncertainty
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.