Research map / Large language models / Reasoning and chain of thought
CGMIS: Concept-Graph Based Multi-Hop Instructions Synthesis for Enhancing Long-Context Reasoning
Zechen Sun, Zecheng Tang, Juntao Li, Wenpeng Hu, Wenliang Chen, Zhunchen Luo, Qiaoming Zhu
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CGMIS, a concept-graph-based pipeline that synthesizes multi-hop instruction data with explicit intermediate reasoning steps to improve LLM long-context reasoning.
Problem
Scarcity of high-quality multi-hop instruction data for long-context reasoning, and existing synthesis methods do not explicitly model intermediate reasoning steps, making the data unverifiable.
Approach
Builds a concept graph to structure multi-hop instruction synthesis so that intermediate reasoning steps are explicitly modeled and the resulting instructions are verifiable.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune on 1K CGMIS multi-hop data | LLaMA-3-8B-Instruct · LongBench avg | average score: 43.7 -> 48.1 | Improved |
| fine-tune on 1K CGMIS multi-hop data | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 52.0 -> 53.1 | Improved |
| fine-tune on 1K CGMIS multi-hop data | Qwen-2.5-14B-Instruct · LongBench avg | average score: 48.0 -> 53.1 | Improved |
| fine-tune on LongAlign (10K) | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 52.0 -> 45.9 | Hurt |
| fine-tune on LongMIT-en (18K) | Qwen-2.5-7B-Instruct · LongBench avg | average score: 39.8 -> 33.9 | Hurt |
| fine-tune on CGMIS | Qwen-2.5-14B-Instruct · LongBench Code | average score: 35.3 -> 53.3 | Improved |
| fine-tune on CGMIS | LLaMA-3 models · LongBench Multi-Document QA | average score: LLaMA-3: HQA 48.2 -> 52.3 but MuSiQue-type tasks vary; slightly underperforms vs LongAlpaca | Mixed |
| regenerate 4-hop data with CGMIS framework instead of LongMIT from same documents (ablation) | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 46.4 -> 50.8 | Improved |
| regenerate 4-hop data with CGMIS framework instead of LongMIT from same documents (ablation) | Qwen-2.5-7B-Instruct · LongBench avg | average score: 39.4 -> 40.6 | Improved |
| increase hops of CGMIS training data from 1 to 4 (ablation) | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 50.2 -> 53.1 | Improved |
| increase hops of LongMIT training data from 1 to 4 (ablation) | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 45.4 -> 46.4 | Improved |
| scale CGMIS data from 1,000 to 2,000 samples (ablation) | LLaMA-3.1-8B-Instruct · LongBench avg | average score: 53.1 -> 51.7 | Hurt |
Closest papers by meaning
- Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models (ICLR 2026 · similarity 0.89)
- Information Re-Organization Improves Reasoning in Large Language Models (NeurIPS 2024 · similarity 0.89)
- Finding the Minimal Parameter Budget for Implicit Reasoning: A Data Complexity Driven Scaling Law for Language Models (ICML 2026 · similarity 0.89)
- GRIP: A Graph-Based Reasoning Instruction Producer (NeurIPS 2025 · similarity 0.89)
- SCOUT: Teaching Pre-trained Language Models to Enhance Reasoning via Flow Chain-of-Thought (NeurIPS 2025 · similarity 0.89)
- MAmmoTH2: Scaling Instructions from the Web (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
