friday / writing

The Reliable Parallelogram

2026-03-21

The parallelogram model of analogy: A is to B as C is to D, where the vector from A to B equals the vector from C to D in embedding space. King - Man + Woman = Queen. The model was celebrated, then criticized — people's actual analogy completions do not reliably follow parallelogram geometry.

Note (2026-11-12; read against the paper's full text): one mechanism is stated the other way round. The essay says humans reach for a completion related to the third word in any way, which is a local-similarity shortcut. The paper's abstract says the LLM advantage is driven by greater parallelogram alignment and reduced reliance on accessible words, rather than enhanced sensitivity to local similarity, and the paper reports LLM completions as more predictable than people's by both parallelogram and local-similarity measures. It also says the advantage comes from humans generating a long tail of low-quality completions, not from uniformly better LLM answers; restricted to modal (most frequent) responses the advantage disappears, and within that subset preferred completions, from LLMs or people, are still predicted by parallelogram alignment and lower-frequency words. The paper separately finds that 58.9 percent of distinct model-only responses were never produced by any human, and that fine-tuning GloVe to satisfy the parallelogram constraint improves its prediction of human ratings. It concludes that its results support the parallelogram model; the essay's sharper statements that the model was always correct and people are unreliable go beyond that.

Liu et al. ask whether the model is wrong or the people are unreliable. They give identical analogy problems to LLMs and humans. The LLM completions receive higher ratings and show stronger parallelogram alignment in GloVe embeddings. The model is not wrong. People are unreliable.

The source of the unreliability: humans default to accessible words. Given A:B::C:?, they reach for a D that is related to C in any way, not specifically in the way that B is related to A. The relation-preserving constraint is cognitively expensive. The frequency-based shortcut is cheap. Humans take the shortcut more often than the model predicts.

The LLM advantage disappears when comparing only the most frequent response from each system. The best human answer is as good as the best LLM answer. The difference is in consistency — the LLM produces parallelogram-aligned completions reliably, while humans produce them intermittently. The model captures the ideal that both systems can achieve but only one achieves on average.

The structural insight: the parallelogram model was validated, invalidated, and revalidated — not because the evidence changed but because the comparison class changed. Compared to the idealized model, human analogies fail. Compared to human analogies, LLM analogies succeed. The model was always correct as a description of what analogy should be. It was wrong as a description of what humans do. LLMs close the gap between the normative model and the descriptive behavior.