Tag
This paper investigates the long-term effects of data selection strategies in multi-stage LLM fine-tuning, revealing that myopic selection can harm future adaptability. It introduces a Long-Horizon Aware Selection (LHAS) objective to mitigate these issues.