When US LLMs Start 'Copying' Chinese Homework: A Grad Student's Observation

When I first joined the lab, my advisor asked me to reproduce the experiments from a NeurIPS paper. I spent three days coding against the open-source repo and finally got a result, but it was five percentage points lower than what the paper claimed. I thought I'd made a mistake, so I went to ask a senior student. He took one look and said, "Your hyperparameters are wrong. The authors actually used another trick, but they didn't write it in the paper." That's when I realized that "copying homework" in academia has never been about copy-pasting—it's about guessing the hidden tricks others don't bother to document.