Parallelograms Strike Back: LLMs Generate Better Analogies than People

This study compares how large language models (LLMs) and humans complete word analogies (A:B::C:D problems). The researchers found that LLMs generate better analogies than humans because they more consistently follow the 'parallelogram' geometric structure—where analogies are represented as equal relationships between word pairs in semantic space—while humans often rely on simpler shortcuts like picking familiar words. The results suggest the parallelogram model is actually a good way to understand analogies; humans just struggle to follow it consistently, whereas LLMs naturally do.