Parallelograms Strike Back: When AI Surpasses Humans at Generating Analogies

*A Chinese-language deep-dive into the paper 'Parallelograms Strike Back: LLMs Generate Better Analogies than People' (arXiv:2603.19066).*

Key points

• The classic puzzle: Four-term analogies (A:B::C:D), e.g., *king : queen :: man : woman*, have long been used to probe analogy — a cornerstone of human cognition emphasized by thinkers from Aristotle to Einstein.
• The parallelogram model: Proposed by Rumelhart & Abrahamson (1973), it holds that concepts are points in a geometric space and relations are vectors, so solving A:B::C:? means completing a parallelogram: vector(D) ≈ vector(C) + vector(B) - vector(A).
• word2vec revival (2013): Mikolov et al.'s word embeddings famously satisfied predictions like vector('king') - vector('man') + vector('woman') ≈ vector('queen'), seemingly validating the model. GloVe and FastText showed similar properties.
• The 2020 setback: Peterson, Smith & Griffiths (CogSci 2020) found that human answers fit a simple heuristic — 'pick the word most similar to C' — better than the parallelogram model, suggesting the geometric model fails to describe human analogy.

The new study

Researchers from Princeton and the University of Hong Kong asked: what if the model is right, but humans just underperform due to time pressure, cognitive load, or limited vocabulary? They tested six frontier LLMs (including GPT-4 and Claude) on the exact analogy items from the 2020 study, then had human judges blindly compare LLM vs. human answers.

Findings

• LLMs won: Human evaluators consistently rated LLM-generated analogies as better than human-generated ones, a statistically significant result across models.
• The parallelogram's revenge: LLM answers aligned more closely with parallelogram structure in GloVe vector space than human answers did, supporting the view that the model is valid but humans fail to realize it.
• Why LLMs excel:
• Less susceptibility to availability bias — they can choose relationally correct but less common words.
• More stable long-tail behavior — human answers had a noisy tail of poor responses; on modal (most common) answers, humans and LLMs performed comparably. LLMs' advantage comes from avoiding bad outliers.
• Rule comparison: For human answers, the C:D similarity heuristic outperformed the parallelogram model; for LLM answers, all three rules (parallelogram, C:D similarity, nearest neighbor) predicted better, and the gap between parallelogram and heuristic shrank by more than half (27,264 vs. 13,355 rank positions).
• Quality predictors: Higher parallelogram alignment (positive, significant) and lower word frequency (negative, significant) predicted better-rated analogies; local similarity sensitivity was not significant.

Implications

• For cognitive science: We may need to reconsider the 2020 interpretation — the parallelogram model may be a good normative model that humans fail to follow under lab conditions. LLMs serve as 'idealized' cognitive agents without fatigue, retrieval failures, or knowledge gaps, offering a new tool for testing cognitive theories.
• Philosophical question: If semantic relations fit geometric structure so well in LLMs, can semantics be fully captured by geometry — and what might such representations miss?
• For AI and education: Better analogy systems should balance local similarity with relational structure, counter availability bias, and possibly embed explicit geometric constraints. Applications include AI-generated teaching analogies and creativity training.

Conclusion

The 'parallelogram's revenge' is not humanity's defeat but a scientific advance: machine analogies exceeding human ones suggests the geometric model of analogy was right all along, and that humans — flexible but noisy reasoners — simply don't always live up to it.

References

1. Liu, Q. E., Marjieh, R., Zhu, J.-Q., Goldberg, A. E., & Griffiths, T. L. (2026). *Parallelograms Strike Back: LLMs Generate Better Analogies than People*. arXiv:2603.19066. 2. Peterson, J. C., Smith, K. A., & Griffiths, T. L. (2020). *Evaluating Vector-Space Models of Analogy*. Proceedings of the 42nd Annual Meeting of the Cognitive Science Society. 3. Rumelhart, D. E., & Abrahamson, A. A. (1973). *A Model for Analogical Reasoning*. Cognitive Psychology, 5(1), 1–28. 4. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). *Efficient Estimation of Word Representations in Vector Space*. arXiv:1301.3781. 5. Pennington, J., Socher, R., & Manning, C. D. (2014). *GloVe: Global Vectors for Word Representation*. EMNLP 2014, 1532–1543.