Intelligent assistant foundation智能助手 AI-Agent 基座

Trained a foundation LLM and tightened intent recognition and procedure modules with RLHF and DPO. Adaptive workflows and VLA-style models power a classified, private-domain RAG assistant. A daily insight agent cleans external knowledge, scores relevance, and drafts the briefing deck plus a live view. For the China Film Science Research Institute text-to-film project, ReAct splits scripts so generation stays efficient and character control stays consistent. 训练基础 LLM，用 RLHF 和 DPO 加强意图识别与规程模块。自适应工作流和 VLA 等模型支撑内部分类的私域 RAG。洞察 Agent 按天清洗外部知识、计算关联度，并生成简报大纲和动态可视化。在中国电影科学研究所的文生电影项目里，用 ReAct 拆分剧本，提高生成效率，并保持人物规控一致。

Realtime affective agent实时多模态情感交互

A VLM drives proactive multimodal interaction. Grammar checking cut scheduled-task latency, and the loop runs on low-compute edge devices. Face and emotion detectors, tuned together with the VLM, let the dialogue inject emotion and control its intensity. 用 VLM 驱动主动的多模态交互。文法检测降低了定时任务时延，交互循环可以跑在小算力端侧设备上。人脸与情绪检测和 VLM 一起优化后，对话可以按情感注入并调节强度。

Data-flow foundation数据流转基座

Following a data-generation paradigm from Carnegie Mellon University, I trained a synthesis model for natural dialogue and classification data, then used RL and DPO to improve naturalness and usability. An agent pipeline self-corrects real-scene ASR annotation and cut about ¥800,000 in manual correction cost on the related project. 参考卡内基梅隆大学的数据生成范式，自训练了面向自然对话和分类数据的合成模型，再用 RL 与 DPO 提高自然度和可用性。另一条 AI-Agent 管线会自主纠正真实场景的 ASR 标注，相关项目大约减少了 80 万人工纠错成本。

CUHK-Shenzhen × Huawei Singapore香港中文大学（深圳）× 华为（新加坡）

Research lead · Spoken misinformation研究组长 · 语音虚假信息检测

Proposed the task, built the dataset with TTS, and assembled a three-stage detection pipeline. I cleaned the text and speech, generated the audio, and wrote the paper. 提出该任务，用 TTS 建成数据集，并搭建三阶段检测流程。负责文本与语音清洗、音频生成和论文撰写。

Tongji Psychology × Toyota × FORVIA同济心理 × 丰田 × FORVIA

Algorithm engineer · Personality model and text-to-video算法工程师 · 人格大模型与文生视频

Fine-tuned Qwen, built the data pipeline, and tuned classification and dialogue models with automated prompts. I trained the psychological-mode and intervention-mode classifiers, and the personality model. 微调 Qwen，完成数据管线，以及分类模型和对话模型。负责心理模式、干预模式分类，和人格大模型训练。

NIO蔚来汽车

Algorithm engineer · Test-case generation算法工程师 · 测试用例生成

Turned requirement documents into test cases with a large model, RAG, and a database. I fine-tuned LLaMA, designed the prompts and chain-of-thought, and built the retrieval stack and frontend APIs. 用大模型、RAG 和数据库，把需求文档生成测试用例。负责 LLaMA 微调、多级 prompt、思维链、检索系统，以及前端接口。

IEEE SLT 2024 · Best Student Paper · Equal contribution最佳学生论文 · 共同一作

SpMis: An Investigation of Synthetic Spoken Misinformation Detection

arXivDataset数据集

Computer Science · CCF T2 · Student first author《计算机科学》· CCF T2 · 学生一作

A Single-Stage Unsupervised Visible-Infrared Person Re-Identification Method一种单阶段无监督可见光-红外跨模态行人重识别方法

Open source开源

Contributions to OpenMMLab and Amphion. Those projects have accumulated more than 10k stars. 参与 OpenMMLab 与 Amphion，相关项目累计超过 10k stars。

OpenMMLabAmphion