夏 伟 ☕️
夏 伟 Xia Wei

Senior Applied Researcher

👋 Hello and welcome to my homepage!

My name is Wei Xia (Xwell, 夏 伟). I am a Senior Applied Researcher at Tencent, specializing in the research and practical application of large language models for Coding. From 2017 to 2024, I worked as a researcher at Huawei Noah’s Ark Lab, where I delved into the study of AI systems (algorithms), including autonomous driving, personalized recommendations, and large language models.

I received my Master's degree in Pattern Recognition and Intelligent Systems from the University of the Chinese Academy of Sciences (UCAS) in 2017 and my Bachelor‘s degree in Automation from the University of South China (USC) in 2014. During my academic journey, I won national awards in competitions like the National Undergraduate Electronic Design Contest and the Freescale Intelligent Car Competition.

Currently, I am seeking interns who are passionate about coding and this specialized field. If you are interested, please feel free to contact me. Thank you!

把书放回床头:我和大模型的向内求与向外看
把书放回床头:我和大模型的向内求与向外看

通勤从地铁换成自驾后,深度阅读的习惯悄悄丢了;把书放回床头,它又慢慢回来了。从这件小事出发,本文把《非暴力沟通》和《助推》看作向内与向外的一体两面,再对照大模型的两层演进:模型自身的推理能力是向内,模型之外的 harness 是向外,讨论两者各自的边界,以及在向外变得太容易的今天,如何把它们组合起来。

Sep 23, 2026

从 Vibe Coding 到 AI-Native Game:生成式游戏技术栈拆解
从 Vibe Coding 到 AI-Native Game:生成式游戏技术栈拆解

生成式游戏 AI 并不是一条单一赛道。本文从生成对象与运行时机出发,拆解 Vibe Coding、专业资产管线、Agent NPC 和神经世界模型四条技术路线,并讨论它们各自真正的工程瓶颈。

May 4, 2026

长程 Agent 为什么会失效:数据、训练与评估
长程 Agent 为什么会失效:数据、训练与评估

Agent 单步很强,任务一长却容易迷路。本文以代码 Agent 为例,分析误差累积、上下文膨胀和奖励稀疏,并梳理轨迹数据、后训练、安全与评估中的关键方法。

Apr 11, 2026

Agent 如何积累能力:RAG、Memory 与 Skills
Agent 如何积累能力:RAG、Memory 与 Skills

RAG、Memory 和 Skills 分别解决信息获取、经验保持与过程复用问题。本文从读写路径、上下文成本和工程边界出发,解释三者如何组合成一套可维护的 Agent 能力系统。

Feb 14, 2026

Code Agent 的中控系统:搜索、验证与长期运行
Code Agent 的中控系统:搜索、验证与长期运行

Code Agent 的能力不只来自模型,也来自规划、工具、验证、上下文和权限组成的控制闭环。本文从搜索、验证、RLVR、上下文与长期运行几个角度拆解这套中控系统。

Jan 2, 2026