Deputy Chief Scientist & Head of NLP · Huawei Xiaoyi 华为小艺 副首席科学家 · NLP 负责人
Building personal AI experts for everyone — self-evolving agents that make life and work better.
为每个人打造专属的 AI 专家顾问，
用自我进化的智能体，让生活与工作更美好。
About关于我
I am Deputy Chief Scientist and Head of NLP at Huawei Xiaoyi, where I lead research and development on LLM agents, question answering, and search. My current focus is building and scaling self-evolving AI systems for deep research, AI scientists, and office automation.
我现任华为小艺副首席科学家、NLP 负责人，带领团队从事大模型智能体、智能问答与搜索的研究与研发。目前专注于构建并规模化可自我进化的 AI 系统，应用于深度研究（Deep Research）、AI 科学家与办公自动化。
Before Huawei, I was a research scientist at Didi AI Lab and an assistant professor at the Institute of Computing Technology, Chinese Academy of Sciences, where I received my Ph.D. and created the Biterm Topic Model (BTM), cited over 2,400 times, for short-text topic modeling.
加入华为之前，我曾任滴滴 AI Lab 研究科学家、中国科学院计算技术研究所助理研究员。博士毕业于中国科学院计算技术研究所，期间提出了短文本主题模型 BTM（Biterm Topic Model），被引用 2,400 余次。
News动态
NEWReleased our survey From Trajectories to Experience: A Survey of Experience-Driven LLM Agents, with a curated list of 271 works.
NEW发布综述 From Trajectories to Experience: A Survey of Experience-Driven LLM Agents，并开放配套论文清单（271 篇）。
NEWSearchJev, a fast and calibrated System-1 model for search agents, is on arXiv. Code, models, and SearchDecision-Bench are open-sourced.
NEW发布 SearchJev：面向搜索智能体的快速、可校准 System-1 决策模型。代码、模型与 SearchDecision-Bench 均已开源。
RubricReviewer is accepted to Findings of EMNLP 2026. It builds a rubric for each manuscript before writing the review, making LLM peer review more objective and comprehensive.
RubricReviewer 被 EMNLP 2026 Findings 接收：先为每篇稿件构建专属评审细则（rubric），再据此生成更客观、全面的审稿意见。
Open-sourced EvoScientist, a vibe research framework for AI-driven autonomous scientific exploration.
开源 vibe research 框架 EvoScientist，用于 AI 驱动的自主科学探索。
Xiaoyi Deep Research ranked #1 on both Deep Research Bench and Deep Research Bench II.
小艺深度研究在 Deep Research Bench 与 Deep Research Bench II 两个榜单均位列第一。
Research研究
From fast decisions inside search agents, to agents that accumulate experience, to systems that do research end to end.
从搜索智能体内部的快速决策，到能够积累经验的智能体，再到端到端完成研究的系统。
A fast, calibrated System-1 model for search agents. It makes each search decision in a single forward pass: 5.2× faster than same-size autoregressive models, with 41–74% lower calibration error.
面向搜索智能体的快速、可校准 System-1 模型，一次前向即可完成每个搜索决策：比同规模自回归模型快 5.2 倍，校准误差降低 41–74%。
A survey of LLM agents that learn from their own trajectories. It organizes 271 works along the experience lifecycle: formation, organization, utilization, and maintenance.
综述：系统梳理大模型智能体如何从自身轨迹中学习，沿经验的形成、组织、利用与维护四个阶段整理了 271 篇工作。
A multi-agent, self-evolving framework for end-to-end scientific discovery. Agents form hypotheses, design and run experiments, and synthesize findings, and they keep improving their own research strategies along the way.
面向端到端科学发现的多智能体自进化框架：智能体自主提出假设、设计并执行实验、总结发现，并在过程中持续改进自身的研究策略。
A large-scale deep research system built on a multi-agent, self-evolving architecture. It decomposes complex questions and synthesizes evidence from many sources. Ranked #1 on Deep Research Bench I & II in 2026.
基于多智能体自进化架构的大规模深度研究系统，可自主拆解复杂问题、检索并综合多源证据。2026 年在 Deep Research Bench I 与 II 双榜位列第一。
The production QA system behind Huawei's Xiaoyi assistant. It covers open-domain question answering, knowledge-grounded dialogue, and intent-aware retrieval across hundreds of millions of queries.
支撑华为小艺助手的线上智能问答系统，覆盖开放域问答、知识增强对话与意图感知检索，服务数亿次用户请求。
LLM-powered search for Huawei's Xiaoyi assistant. It retrieves from the web and authoritative sources and returns direct answers with cited references, turning data into charts when helpful.
华为小艺的大模型搜索：联网检索网页与权威信源，直接生成带引用来源的答案，并可将数据整理为图表。
From Trajectories to Experience: A Survey of Experience-Driven LLM Agents
Preprint coming soon预印本即将发布
Exploring Cross-lingual Latent Transplantation: Mutual Opportunities and Open Challenges
IEEE Transactions on Audio, Speech and Language Processing · earlier arXiv title: XTransplantarXiv 早期版本名为 XTransplant
Knowledge-Augmented Question Error Correction for Chinese Question Answer System with QuestionRAG
CC-Tuning: A Cross-Lingual Connection Mechanism for Improving Joint Multilingual Supervised Fine-Tuning
A topic model for short texts such as tweets, queries, and titles. Instead of modeling sparse documents one by one, BTM models word co-occurrence patterns (biterms) over the whole corpus. It has been widely adopted for short-text mining and was extended to bursty topic discovery (BBTM) and online inference.
面向微博、查询、标题等短文本的主题模型。BTM 不再逐篇建模稀疏的短文档，而是直接在整个语料上建模词对（biterm）共现，有效缓解了短文本的稀疏性问题。该模型被广泛用于短文本挖掘，并扩展至突发话题发现（BBTM）与在线推断。
Graph LSTM with Context-Gated Mechanism for Spoken Language Understanding
Convolutional Hierarchical Attention Network for Query-Focused Video Summarization
Generative Question Refinement with Deep Reinforcement Learning in Retrieval-based QA System
Long-Form Video Question Answering via Dynamic Hierarchical Reinforced Networks
Learning Representation Mapping for Relation Detection in Knowledge Base Question Answering
Abstractive Meeting Summarization by Hierarchical Adaptive Segmental Network Learning with Multiple Revising Steps
Hashtag Recommendation for Photo Sharing Services
An Integral Tag Recommendation Model for Textual Content
Zero-shot User Intent Detection via Capsule Neural Networks
Abstract Meaning Representation for Paraphrase Detection
Ease the Process of Machine Learning with Dataflow
A Probabilistic Model for Bursty Topic Discovery in Microblogs
BTM: Topic Modeling over Short Texts
Topic Modeling over Short Texts
Probabilistic Transaction Model for Recommending Data of Offline Shopping Mall
A Biterm Topic Model for Short Texts
Learning Topics in Short Texts by Non-negative Matrix Factorization on Term Correlation Matrix
Clustering Short Text Using Ncut-weighted Non-negative Matrix Factorization
Context-aware Query Recommendation by Learning High-order Relation in Query Logs
Experience经历
2017.02 — Present至今
Huawei · Xiaoyi华为 · 小艺
Leading research on deep research agents, search-augmented LLMs, large-scale QA, and conversational AI. Xiaoyi Deep Research ranked #1 on Deep Research Bench I & II in 2026. Earlier, led the dialog and QA group at Huawei Poisson Lab.
负责深度研究智能体、搜索增强大模型、大规模问答与对话系统的研究。小艺深度研究于 2026 年登顶 Deep Research Bench I 与 II 双榜。此前负责华为泊松实验室对话与问答团队。
2016.04 — 2017.02
Didi AI Lab滴滴 AI Lab
Worked on user profiling and intelligent car dispatch at scale.
从事大规模用户画像与智能派单研究。
2014.07 — 2016.04
Institute of Computing Technology, Chinese Academy of Sciences中国科学院计算技术研究所
Built BDA (Big Data Analysis), a cloud-based machine learning platform with a web interface for building and managing ML pipelines (CIKM'16 demo).
研发云端机器学习平台 BDA（Big Data Analysis），通过 Web 界面构建与管理机器学习流程（CIKM'16 Demo）。
— 2014
Institute of Computing Technology, Chinese Academy of Sciences中国科学院计算技术研究所
Thesis: 博士论文：Topic Modeling over Short Texts