KG-SFT团队 投稿 量子位 | 公众号 QbitAI 让大语言模型更懂特定领域知识，有新招了！ 来自中国科学技术大学MIRA实验室的王杰教授团队提出了提出了一个创新的框架—— 知识图谱驱动的监督微调（KG-SFT） ，该框架通过 引入知识图谱（KG）来提升大语言模型（LLMs）在特定领域的知识理解和处理能力。 实验结果表明，其在多个领域和多种语言的数据集上取得了显著的效果， 成功入选ICLR 2025 。 截至目前，LLMs在常识问答方面表现越来越出色，但它们对领域知识的理解和推理能力仍然有限。 由于难以深入理解专业领域问答背后所蕴含的复杂知识和逻辑关系，因此在面对这类问题时，往往无法准确地给出正确的答案和详细的推理过 程，这极大地限制了其在专业领域的应用价值。 尤其是在数据稀少和知识密集型的场景中， 如何让LLMs更好地理解和操纵知识，成为了研究的关键 。 而中科大MIRA实验室的这项工作即围绕此展开。 KG-SFT是如何工作的 KG-SFT针对LLMs难以理解领域问答背后的知识和逻辑，导致 推理能力弱 的问题，提出 基于知识图谱增强的大语言模型监督微调 技术。 KG-SFT首先通过解析领域知识图谱中的 ...

Core Insights - The company, Fenbi, reported a revenue of 2.79 billion yuan and a net profit of 240 million yuan for the fiscal year 2024, marking a 27% year-on-year increase in net profit [2] - Fenbi's strategy focuses on "AI + vocational education," with an average monthly active user count of 9.14 million during the reporting period [2] - The company plans to enhance its AI capabilities, with a research and development expenditure of 220 million yuan in 2024, and the introduction of the DeepSeek large model by the end of 2024 [2] AI Development and Product Performance - Fenbi's self-developed domain-specific large model outperformed general large models in various assessment categories, including language and reasoning [3] - The company has established a comprehensive AI product matrix, including AI teachers and interview evaluation systems, catering to diverse user needs through subscription and pay-per-use models [3] Industry Outlook - According to Guojin Securities, 2025 is expected to be a critical year for the application of AI in education, supported by new technologies such as knowledge graphs and multimodal interaction [4] - Fenbi's CEO emphasized the importance of integrating AI technology with industry data to enhance service relevance, with plans to increase AI R&D investment by 30% annually [4]

Group 1: Company Overview - The company initially conducted its own research to gather data due to limited data availability, but has since evolved to leverage data collection as a key service in the big data era [2] - The company has developed two analytical platforms: one for data aggregation and rapid analysis, and another for testing, which can operate independently or in conjunction [2] - The company aims to provide accurate data interpretations and responses to client inquiries, even in data-scarce situations, ensuring client satisfaction [2] Group 2: Model Development and Innovation - The company is transitioning from manual data analysis to automated analysis, enhancing its model-building capabilities to adapt to market changes [3] - The company is focusing on creating unique models tailored to specific problems, moving beyond traditional models to innovate in model construction [3] - The gaming industry serves as a reference for the company, which is working to improve its model-building capabilities to meet evolving market demands [3] Group 3: Short-term Development Strategy - The company is implementing internal growth strategies by encouraging younger management to share client expansion experiences, leading to business growth [3] - The company is pursuing external growth through strategic acquisitions to address research and development gaps and find complementary firms in vertical markets [3] - The company has invested in a knowledge graph company, aiming to leverage its technology for structured knowledge management and problem-solving within enterprises [3] - The company plans to market standard knowledge graph products to achieve scalable sales through client resource integration [3]