Reasoning or Memorization: Can LLMs Understand and Generate Chinese Xiehouyu Riddles?

This paper evaluates LLMs on Chinese xiehouyu, a form of two-part riddle, using linguist-created expressions that had never previously existed to reduce contamination. It tests multiple-choice understanding, free-form explanations, and new-riddle creation. The reported accuracy gap between existing low-frequency and novel xiehouyu averages 23.6% for frontier Chinese models, versus 5.1% for English-centric models. Gemini 3.1 Pro reaches 92.6% accuracy on novel xiehouyu, reportedly 24% above human accuracy. However, LLM-generated xiehouyu receive substantially worse ratings than human-created ones.