Beyond Scale and Generation: Understanding Language Model-based Entity Matching

A controlled arXiv study isolated how architecture, Qwen3 variant, model size, and dataset conditions affect language-model-based entity matching.

Embedding-oriented variants help bi-encoders, but cross-encoders stay stronger overall because they jointly encode record pairs.

Generative matchers are especially useful under distribution shift and cross-dataset transfer.

The researchers found that bigger models do not always perform better, likely due to shortcut learning, so variant choice and task conditions matter more than scale.