95% Accuracy with Modular Prompting (vs. Near-Failure Without)

On the GeoLink complex ontology alignment benchmark:

  • Without modularity: LLM produced “essentially unusable” results when given full ontologies (40-156 classes)
  • With modularity: 95% accuracy (104 of 109 target mappings correct) using two-stage modular prompting

The benchmark involved creating complex mapping rules between two oceanographic ontologies, a task that had “defied automation for years” with traditional approaches.

The only change was how the task was structured: first prompting to identify relevant modules from a list of 20 named options, then prompting for the actual mapping using only those modules.

Related: 05-atom—modularity-unlocks-llm-performance, 05-molecule—two-stage-modular-prompting