LLMs Learn Constructional Semantics for Rare English Paired-Focus Phrases
A new study from arXiv (2605.31586) investigates whether large language models understand the semantics of rare English constructions like "let alone" and "much less". The researchers created a novel dataset testing Paired-Focus constructions using scalar adjectival semantics and world knowledge. They found that several modestly-sized models are sensitive to both form and meaning, while models trained on human-scale data fail all meaning evaluations. The study also examines training dynamics for open-checkpoint models.
Key facts
- Study focuses on rare Paired-Focus constructions in English
- Examples include 'let alone' and 'much less'
- Novel dataset tests meanings using scalar adjectival semantics and world knowledge
- Several modestly-sized models show sensitivity to form and meaning
- Models trained on human-scale data fail all meaning evaluations
- Study examines training dynamics for open-checkpoint models
- arXiv paper number: 2605.31586
- Research addresses open question about open-source model constructional understanding
Entities
Institutions
- arXiv