---
title: "Beyond Scale and Generation: Understanding Language Model-based Entity Matching"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.24688v1"
permalink: "https://feed.astralor.com/items/itm_307f53e8701825e8f07a99bc"
published: "2026-07-27T17:29:18.000Z"
score: 81
channel: "ai"
tags: ["Qwen3","Cross-Encoder","Bi-Encoder","Entity Matching","Shortcut Learning","generative matcher"]
---

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

## AI commentary

> This paper is noteworthy for providing the most rigorous comparison of entity matching architectures to date, especially the counterintuitive conclusion that larger models are not necessarily better. With open-sourced resources, it is highly valuable for practitioners. Teams working on data integration, knowledge graph construction, or structured data processing with LLMs will find this paper helps avoid common pitfalls.

## At a glance

This paper conducts a controlled factorial study on three matcher architectures (bi-encoder, cross-encoder, generative), three model variants, and three model sizes (from Qwen3 family) with 1,215 fine-tuning runs across nine datasets. Key findings: model variant is critical for bi-encoders; cross-encoders consistently outperform bi-encoders but larger models can narrow the gap; generative matchers do not universally outperform cross-encoders, only excelling under distribution shift or cross-dataset transfer; larger models rely more on shortcut learning and do not necessarily perform better. The study clarifies the interplay of architecture, model variant, and scale in entity matching.

## Key points

- Controlled 1,215 fine-tuning runs across architecture, model variant, and size
- Model variant is critical for bi-encoder initialization and representation geometry
- Cross-encoders consistently outperform bi-encoders, but larger models narrow the gap
- Generative matchers only excel under distribution shift and cross-dataset transfer; larger models show shortcut learning

## Full text

Entity matching identifies records that refer to the same real-world entity. Language models can be adapted to this task through bi-encoder, cross-encoder, and generative matcher architectures. However, prior studies often conflate matcher architecture with differences in model backbone, model variant(reflecting different pretraining objectives), and model size, making it difficult to isolate the sources of performance gains. We address this issue through a controlled factorial study spanning three matcher architectures, three model variants and three model sizes from the Qwen3 family, and nine datasets, totaling 1,215 fine-tuning runs. We also evaluate cross-dataset transferability and computational cost. Our results show that model variant is critical for bi-encoders: embedding-oriented variants provide stronger initialization and more favorable representation geometry predictive of downstream matching performance. Cross-encoders retain a consistent advantage over bi-encoders because they jointly encode record pairs rather than representing each record independently, although larger models partially narrow this gap. Generative matchers do not universally outperform cross-encoders. Instead, their advantages concentrate under distribution shift, including subtle unseen differences in record schemas and cross-dataset transfer. We further find that larger models rely more heavily on shortcut learning and therefore do not necessarily perform better. These findings clarify the factors underlying performance differences across matcher architectures and motivate future research and benchmark designs that better disentangle architectural choices from model-level factors while explicitly evaluating distribution shift and cross-dataset transferability. We release our experimental results, code, training scripts, and evaluation data at https://github.com/Jantory/llm-trained-matcher.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_307f53e8701825e8f07a99bc