arXivPaperNeeds Review
Beyond Scale and Generation: Understanding Language Model-based Entity Matching
A 1,215-run study finds entity-matching architecture and model variant matter more than scale alone; generative matchers mainly help under distribution shift.
arXiv · Jul 27, 2026
Source Summary
The study ran **1,215 fine-tuning experiments** across three matcher architectures, three Qwen3 variants, three sizes, and nine datasets. Cross-encoders consistently beat bi-encoders, while embedding-oriented variants gave bi-encoders better starting representations.
Practical Implication
For data-matching systems, choose architecture against deployment conditions rather than defaulting to the largest generative model. **Generative matchers** showed their advantage mainly under schema shifts and cross-dataset transfer; cross-encoders remained the stronger general baseline.
Agent-Ready Context
The study ran **1,215 fine-tuning experiments** across three matcher architectures, three Qwen3 variants, three sizes, and nine datasets. Cross-encoders consistently beat bi-encoders, while embedding-oriented variants gave bi-encoders better starting representations. For data-matching systems, choose architecture against deployment conditions rather than defaulting to the largest generative model. **Generative matchers** showed their advantage mainly under schema shifts and cross-dataset transfer; cross-encoders remained the stronger general baseline. Larger models sometimes relied more on shortcuts and did not reliably improve results. This is a preprint under review, and the supplied material gives no task-level scores or cost figures for judging the practical size of each tradeoff.
Context Map
benchmarkdata#model-selection#benchmark-integrityUncertainty
Larger models sometimes relied more on shortcuts and did not reliably improve results. This is a preprint under review, and the supplied material gives no task-level scores or cost figures for judging the practical size of each tradeoff.