Large language models (LLMs) excel at generating text, but how do they fare at classification tasks? A new study tackles this question by examining "edit intent classification" (EIC) – the process of identifying the purpose behind edits in a document. This involves understanding not just *what* changed, but *why*. Think about revising a scientific paper. You might correct grammar, clarify language, add supporting evidence, strengthen claims, or make other miscellaneous tweaks. EIC seeks to label each edit with its underlying purpose. Researchers created a framework to thoroughly test LLMs on EIC, comparing various approaches and training strategies. Surprisingly, LLMs fine-tuned for EIC proved highly effective, outperforming even instruction-tuned behemoths like Llama2-70B. They even bested smaller, fully fine-tuned models, setting a new state-of-the-art for EIC. But the most exciting outcome? The researchers used their top-performing EIC model to build a massive new dataset, "Re3-Sci2.0," containing 1,780 scientific papers and over 94,000 labeled edits across various disciplines. This treasure trove opens doors for deeper explorations into how scientists revise their work. Initial findings suggest that successful revisions often focus on improving clarity and adding evidence – valuable insights for any researcher. This work pushes the boundaries of LLMs in classification and provides powerful tools for studying human editing behavior in scientific writing and beyond.