Jinqi Jiang

Zihan Luo, Xiran Song, Hong Huang et al.

Improving the general capabilities of large language models (LLMs) is an active research topic. As a common data structure in many real-world domains, understanding graph data is a crucial part of advancing general intelligence. To this end, we propose a dynamic benchmark named GraphInstruct in this paper, which comprehensively includes 21 classical graph reasoning tasks, providing diverse graph generation pipelines and detailed intermediate reasoning steps for each sample. Based on GraphInstruct, we develop GraphSolver via efficient instruction-tuning, which demonstrates prominent graph understanding capability compared to other open-sourced LLMs. To further endow LLMs with multi-step graph reasoning capability, we propose a label-mask training strategy and build GraphSolver+, which leverages masked supervision on intermediate reasoning tokens to emphasize crucial node-identification signals. As one of the pioneering efforts to enhance the graph understanding and reasoning abilities of LLMs, extensive experiments have demonstrated the superiority of GraphSolver and GraphSolver+ over other LLMs. We sincerely hope GraphInstruct will facilitate further research on applying LLMs to graph-structured data. Our code and data are released publicly at: https://github.com/CGCL-codes/GraphInstruct.

Zhaoyang Wang, Jinqi Jiang, Tian Qiu et al.

Recent large reasoning models such as DeepSeek-R1 exhibit strong complex problems solving abilities by generating long chain-of-thought (CoT) reasoning steps. It is challenging to directly train small language models (SLMs) to emerge long CoT. Thus, distillation becomes a practical method to enable SLMs for such reasoning ability. However, the long CoT often contains a lot of redundant contents (e.g., overthinking steps) which may make SLMs hard to learn considering their relatively poor capacity and generalization. To address this issue, we propose a simple-yet-effective method to prune unnecessary steps in long CoT, and then employ an on-policy method for the SLM itself to curate valid and useful long CoT training data. In this way, SLMs can effectively learn efficient long CoT reasoning and preserve competitive performance at the same time. Experimental results across a series of mathematical reasoning benchmarks demonstrate the effectiveness of the proposed method in distilling long CoT reasoning ability into SLMs which maintains the competitive performance but significantly reduces generating redundant reasoning steps.