Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

LongCrafter: Towards Diverse Long-Context Understanding via Evidence-Graph-Guided Instruction Synthesis

Source

arXiv

arxiv.org/abs/2607.06160v1

Read the source

Source

  • Source name: arXiv
  • Source kind: paper
  • Authors / organization: Chenhao Yuan, Yinhao Xu, Shuwen Xu, Xizhi Yang, Jiaxiang Liu, Chenxi Zhou, Shaoping Huang, Haolin Ren, Pengfei Cao, Jun Zhao, Kang Liu
  • Venue: arXiv: cs.CL
  • Published: 2026-07-07
  • Collected: 2026-07-08
  • URL: http://arxiv.org/abs/2607.06160v1
  • PDF: https://arxiv.org/pdf/2607.06160v1
  • DOI: UNKNOWN

Plain-Language Brief

LongCrafter is a framework designed to enhance the understanding capabilities of large language models by synthesizing long-context supervised fine-tuning data. It addresses three limitations of existing methods: narrow task coverage, insufficient instruction difficulty, and lack of faithful supervision. LongCrafter uses a hierarchical task taxonomy and an evidence-grounded pipeline to construct task-aligned long contexts, create explicit evidence graphs, and generate instruction-response pairs grounded in located evidence spans.

What the Document Says

  • LongCrafter is a structured synthesis framework that enhances the understanding of large language models (LLMs).
  • It addresses three limitations: narrow task coverage, insufficient instruction difficulty, and lack of faithful supervision in existing methods.
  • LongCrafter uses a hierarchical task taxonomy with 32 fine-grained task types as a global generative prior.
  • It constructs task-aligned long contexts, decomposes them into explicit evidence graphs modeling cross-paragraph dependencies.
  • Instruction-response pairs are generated based on the located evidence spans to ensure controllable difficulty and faithful, traceable reasoning.
  • LongCrafter outperforms all SFT baselines and official post-trained models across various datasets (LongBench, LongBench~v2, LooGLE) and model sizes (Qwen2.5-7B, LLaMA-3.1-8B).
  • The largest gains are on high-difficulty tasks.
  • Further analysis shows that LongCrafter data is more diverse and better spread across difficulty levels.
  • Trained models locate evidence robustly regardless of position, reducing the 'lost in the middle' problem.

Key Concepts Mentioned

  • LongCrafter
  • large language models (LLMs)
  • supervised fine-tuning data
  • hierarchical task taxonomy
  • evidence-grounded pipeline
  • explicit evidence graphs
  • instruction-response pairs
  • controllable difficulty
  • faithful, traceable reasoning
  • LongBench
  • LongBench~v2
  • LooGLE
  • Qwen2.5-7B
  • LLAMA-3.1-8B

Claims Made by the Document

  • LongCrafter addresses three limitations of existing SFT methods.
  • LongCrafter outperforms all SFT baselines and official post-trained models across various datasets and model sizes.
  • The largest gains are on high-difficulty tasks.
  • LongCrafter data is more diverse and better spread across difficulty levels.
  • Trained models locate evidence robustly regardless of position.

Evidence Used

  • The text states that LongCrafter addresses three limitations...
  • The text mentions that LongCrafter outperforms all SFT baselines...
  • The text indicates that the largest gains are on high-difficulty tasks...
  • The text reveals that LongCrafter data is more diverse and better spread across difficulty levels...
  • The text states that trained models locate evidence robustly regardless of position.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

Synthesizing long-context supervised fine-tuning (SFT) data is a scalable way to enhance the long-context understanding of large language models (LLMs), yet existing approaches share three limitations: narrow task coverage, insufficient instruction difficulty, and a lack of faithfulness supervision. We propose \textbf{LongCrafter}, a structured synthesis framework that couples a hierarchical task taxonomy with an evidence-grounded pipeline. The taxonomy organizes long-context understanding into local/shallow and global/deep levels and yields 32 fine-grained task types that serve as a global generative prior. Guided by this taxonomy, LongCrafter constructs task-aligned long contexts, decomposes them into explicit evidence graphs that model cross-paragraph dependencies, and generates instruction--response pairs strictly grounded in the located evidence spans, ensuring both controllable difficulty and faithful, traceable reasoning. Models fine-tuned on LongCrafter data outperform all SFT baselines and even the official post-trained models on LongBench, LongBench~v2, and LooGLE across both Qwen2.5-7B and LLaMA-3.1-8B, with the largest gains on high-difficulty tasks. Further analysis shows that LongCrafter data is more diverse and better spread across difficulty levels, and that the trained models locate evidence robustly regardless of position, effectively mitigating the ``lost in the middle'' problem.