AIME 24/25

Data repetition outperforms scaling in LLM fine-tuning, study finds

A new research paper suggests that for supervised fine-tuning of reasoning language models, data repetition can be more effective than data scaling. The study found that training for more epochs on smaller datasets, eve…