LLMs tend to cluster semantically related words more tightly than classical models, showing higher accuracy on the Bigger Analogy Test Set.
LLMs tend to cluster semantically related words more tightly than classical models, showing higher accuracy on analogy tasks.
The author compares Large Language Models (LLMs) with classical models like Sentence-BERT and Universal Sentence Encoder to determine if the performance improvement is due to scale or fundamentally different embeddings. The approach involves analyzing semantic clustering and accuracy on analogy tasks.