StructuredSemanticSearch uses model card tables for diverse model discovery

Model cards—documents describing AI model behavior—contain a mix of text and structured artifacts such as performance, configuration, and dataset tables. Traditional model search systems rely on semantic similarity over text, often producing homogeneous results that limit exploration of alternatives. Researchers Zhengyuan Dong and Renée J. Miller argue that model search is inherently comparative: users want models that are task-aligned yet differentiated in measurable ways. They hypothesize that retrieving condensed, high-quality evidence from structured tables is key, rather than relying on verbose descriptions.

To test this, they built StructuredSemanticSearch, a table-driven model search framework built on the new ModelTables benchmark. Given a query, it combines a semantic baseline for task alignment with a structure-aware pipeline that discovers query-related model-card tables using operators such as unionability, joinability, and keyword search. Retrieved tables are mapped back to model cards under a controlled top-k budget. Beyond retrieval, it adapts table integration to the model-table domain through orientation-aware integration, producing compact views from partially overlapping and transposed evidence tables.

For evaluation, the authors introduced a nugget-based protocol that extracts compact evidence items from model cards, matches queries to condition- or intent-specific nuggets, and measures evidence coverage and diversity. Experiments on 597 model-recommendation queries showed that the structure-aware pipeline achieved improved nugget coverage compared to the semantic baseline. This work provides a scalable path toward approximate, evidence-based labeling in dynamic model lakes, enabling practitioners to discover relevant yet diverse models more effectively.