Do Language Models Show Structural Priming Across Different Domains?

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computational Linguistics · Depth: Advanced, short

Summary

A study by So Young Lee, Russell Scheinberg, and Ameeta Agrawal, presented at the 1st Workshop on Computational Developmental Linguistics (CDL) in July 2026, investigated whether large language models (LLMs) exhibit cross-domain structural priming. The research specifically tested if arithmetic expressions influence relative-clause attachment preferences. Experiment 1 examined English and French using materials from prior psycholinguistic studies, while Experiment 2 expanded this test to a larger multilingual dataset. Across both experiments, the authors found no robust priming effect. Instead, LLM responses primarily reflected baseline attachment preferences, which varied by language and only partially matched human patterns. These findings indicate that although LLMs possess some structural sensitivity, they demonstrate limited evidence of abstract structural generalization across different domains.

Key takeaway

For NLP Engineers evaluating LLM linguistic capabilities, this research suggests current models may not generalize structural knowledge abstractly across different domains. You should not assume LLMs will automatically transfer structural understanding from one task or domain to another, even if they show basic structural sensitivity. Consider designing specific training regimes or architectural modifications to explicitly foster abstract structural generalization, rather than relying solely on broad pre-training.

Key insights

LLMs exhibit structural sensitivity but lack robust cross-domain structural generalization.

Principles

• LLM responses reflect baseline attachment preferences.
• Baseline preferences vary across languages.
• LLM patterns partially align with human patterns.

Method

The study tested cross-domain structural priming in LLMs by assessing if arithmetic expressions influence relative-clause attachment preferences across English, French, and a larger multilingual dataset.

In practice

• Evaluate LLM structural sensitivity with psycholinguistic tasks.
• Compare LLM linguistic patterns to human baselines.

Topics

• Large Language Models
• Structural Priming
• Cross-Domain Generalization
• Relative-Clause Attachment
• Psycholinguistics
• Multilingual NLP

Best for: AI Scientist, NLP Engineer, Research Scientist

Related on AIssential

• Multilinguality of Large Language Models From a Structural Perspective — LLM multilinguality shows low-resource languages are structurally more distinct from English, with post-training altering structures but preserving inter-language relationships.
• Analysis of the Neglect-Zero Effect in Large Language Models — LLMs may not share the human "neglect-zero effect" cognitive bias regarding empty sets.
• Filling in the Mechanisms: How do LMs Learn Filler-Gap Dependencies under Developmental Constraints? — LLMs show shared, item-sensitive filler-gap dependency mechanisms, but require more data than humans.
• Hybrid Human-LLM Corpus Construction and LLM Evaluation for the Caused-Motion Construction — LLMs struggle with complex linguistic constructions like CMC, necessitating novel corpus creation methods.
• A family of large language models for materials research with insights into model adaptability in continued pretraining — Domain-adapted LLMs significantly enhance materials research, but overtraining can hinder further specialization.
• Large Language Models in NLP — LLMs integrate NLP advancements with massive scale to create general-purpose, prompt-driven language systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.