LongCrafter uses evidence-graph-guided synthesis to improve long-context understanding

The authors propose LongCrafter, a structured framework for synthesizing long-context supervised fine-tuning data that addresses limitations in task coverage and faithfulness. It couples a hierarchical taxonomy of 32 task types with an evidence-grounded pipeline to generate instruction-response pairs strictly grounded in located evidence spans.

• Models fine-tuned on LongCrafter data outperform all SFT baselines and official post-trained models on LongBench, LongBench v2, and LooGLE across Qwen2.5-7B and LLaMA-3.1-8B.
• The approach yields the largest gains on high-difficulty tasks and produces data that is more diverse and better spread across difficulty levels.
• Trained models locate evidence robustly regardless of position, effectively mitigating the "lost in the middle" problem.

LongCrafter ensures controllable difficulty and faithful, traceable reasoning, providing a scalable way to enhance long-context understanding in large language models.