X-Intelligence 3.0：面向半导体显示器的推理LLM训练与评估
计算与语言
2025-07-23 v2
摘要
大型语言模型 (LLM) 最近在推理方面取得显著进展，已展现出解决复杂问题的优势。然而，其在半导体显示行业的有效性仍受限于缺乏领域特定训练和专业知识。为弥合这一差距，我们提出X-Intelligence 3.0，这是第一个为半导体显示行业开发的高性能推理模型。该模型旨在为行业的复杂挑战提供专业水平的理解与推理能力。我们构建了精心策划的行业知识库，使模型经过监督微调和强化学习以提升推理与理解能力。为进一步加速开发，我们实现了一个模拟专家水平评估的自动化评估框架。我们还集成了领域特定的检索增强生成 (RAG) 机制，显著提升了在基准数据集上的性能。尽管其规模相对较小（仅320亿参数），X-Intelligence 3.0 在多个评估中超越了SOTA DeepSeek-R1-671B。这表明其卓越的效率，并使其成为半导体显示行业长期解决推理挑战的强大解决方案。
引用
@article{arxiv.2507.14430,
title = {X-Intelligence 3.0: Training and Evaluating Reasoning LLM for Semiconductor Display},
author = {Xiaolin Yan and Yangxing Liu and Jiazhang Zheng and Chi Liu and Mingyu Du and Caisheng Chen and Haoyang Liu and Ming Ding and Yuan Li and Qiuping Liao and Linfeng Li and Zhili Mei and Siyu Wan and Li Li and Ruyi Zhong and Jiangling Yu and Xule Liu and Huihui Hu and Jiameng Yue and Ruohui Cheng and Qi Yang and Liangqing Wu and Ke Zhu and Chi Zhang and Chufei Jing and Yifan Zhou and Yan Liang and Dongdong Li and Zhaohui Wang and Bin Zhao and Mingzhou Wu and Mingzhong Zhou and Peng Du and Zuomin Liao and Chao Dai and Pengfei Liang and Xiaoguang Zhu and Yu Zhang and Yu Gu and Kun Pan and Yuan Wu and Yanqing Guan and Shaojing Wu and Zikang Feng and Xianze Ma and Peishan Cheng and Wenjuan Jiang and Jing Ba and Huihao Yu and Zeping Hu and Yuan Xu and Zhiwei Liu and He Wang and Zhenguo Lin and Ming Liu and Yanhong Meng},
journal= {arXiv preprint arXiv:2507.14430},
year = {2025}
}
备注
Technical Report