Diversed Model Discovery via Structured Table Discovery

Researchers introduced StructuredSemanticSearch, a hybrid model search system that combines semantic retrieval with table discovery over model cards.

The approach uses structured evidence such as unionability and joinability to surface more diverse, comparable model candidates.

On 597 model-recommendation queries, it was evaluated with a nugget-based protocol focused on evidence coverage.

Results show better nugget coverage than a semantic-only baseline, improving the usefulness of model discovery.