paper / taesung-lee / May 9 / failedLatent Traits and Cross-Task Transfer: Deconstructing Dataset Interactions in LLM Fine-tuning