我是清华大学深圳国际研究生院数据科学和信息技术专业2025级的硕士研究生，由陈鑫磊副教授、高宸助理研究员和李勇教授共同指导。研究关注具身智能与世界模型：如何让模型理解环境、预测动作带来的变化，推动物理智能发展。
I am a 2025 master's student in Data Science and Information Technology at Tsinghua Shenzhen International Graduate School, co-advised by Xinlei Chen, Chen Gao, and Yong Li. I work on embodied intelligence and world models: how a model understands its environment, predicts the consequences of actions, and advances physical intelligence.
目前在 Manifold AI（流形空间）主导世界模型训练，由高宸、武伟和李勇指导。课余我徒步、摄影、登山、攀岩，也坚持长跑，是多个户外协会领队，并且入选凯乐石-领攀签约摄影师。
I lead world-model training at Manifold AI, advised by Chen Gao, Wei Wu, and Yong Li. Outside research I hike, photograph, mountaineer, climb, and keep up distance running. I lead several outdoor clubs and am a KAILAS Lingpan signed photographer.
深圳国际研究生院，数据科学和信息技术，学术型硕士。院系内年级排名 1/1428。
Shenzhen International Graduate School, M.S. in Data Science and Information Technology. Ranked 1/1428 in the cohort.
计算机科学与技术，工学学士。专业综合 1/199。2023/2024 国家奖学金，2025 东北大学文浩命名奖学金（仅 1 位），十佳本科生，十佳双创之星。
B.Eng. in Computer Science and Technology. Comprehensive ranking 1/199. National Scholarship in 2023 and 2024, the 2025 Northeastern University Wenhao Named Scholarship (sole recipient), Top 10 Undergraduate, and Top 10 Innovation Star.
实习。主导内部 VLA 具身基座模型训练。
Intern. I lead training of an internal VLA embodied foundation model.
面向异源动作控制的 MoE 世界模型：共享专家学习跨控制的世界动态，专属专家吸收不同控制接口，并支持向新控制方式扩展。
A mixture-of-experts world model for heterogeneous action control. Shared experts learn dynamics that transfer across controls, while control-specific experts absorb each interface and extend to new ones.
用统一动作生成框架把异构动作接口放进同一评测协议，并提供数据、任务和公开榜单，使不同世界模型可以公平比较。
A unified action-generation framework that places heterogeneous control interfaces under one evaluation protocol, with data, tasks, and a public leaderboard.
按动作真正改变的区域重新分配训练监督，让世界模型学习动作如何改变场景，而不是被大量背景像素主导。
Reweights training toward the regions an action actually changes, so a world model learns action effects instead of fitting abundant background pixels.
不额外训练规划头，只把冻结世界模型的短期目标锚到自身轨迹上的近处观测，从而把长程规划做远。
A training-free method that aims a frozen world model at a nearby observation along its own trajectory, so short-horizon search can reach long-range goals.
用推理增强的长短时记忆和上下文学习，让世界动作模型既能按高层指令自主规划，也能按文本、目标图或视频细粒度执行。
A steerable world action model with reasoning-augmented memory and in-context learning, for planning from high-level instructions and control from text, goal images, or video.
用模型内部对操作物体的注意力作为交互图，把去噪监督集中到真正发生交互的区域，而不依赖外部标注。
Uses the model's own attention on manipulated objects as an interaction map, concentrating denoising supervision on interaction regions without external annotations.
按注意力键的局部密度压缩历史 KV 缓存，在存储不随生成长度增长的前提下保持长视频的主体和场景一致。
Compresses the historical KV cache by local key density, keeping long-video subjects and scenes consistent without storage that grows with generation length.
把空中视觉语言导航做成短时程世界状态预测，并把预测直接解码成可执行航点，形成闭环的世界动作模型。
Casts aerial vision-language navigation as short-horizon world-state prediction and decodes those predictions into executable waypoints in closed loop.
用 5037 条城市空域目标导航样本评测 17 个模型，衡量大模型离人类水平的空间行动还有多远。
A goal-oriented navigation benchmark in urban airspace, with 5,037 samples and an evaluation of 17 models against human-level spatial action.
用多视角角色扮演的大模型模拟人对事件的反应，预测社交媒体用户接下来的情绪，而不只是回顾已有发言。
Forecasts how social-media users will feel about unfolding events by role-playing multiple perspectives with large language models.
用大模型为异构空地搜索群做集中任务分配，并把推理与执行异步解开，已在仿真和真实空地平台上完成搜救部署。
An LLM task-allocation framework for a heterogeneous air-ground search swarm, with asynchronous reasoning and execution, deployed in simulation and on a physical platform.
大视觉语言模型负责感知、小语言模型负责推理，并用逻辑一致性奖励做强化学习；5k 条样本、3B 参数即可达到 o1 与 Gemini-2.5-Pro 的同等水平。
A large vision-language model perceives and a smaller language model reasons, trained with a logical-consistency reward. With 5k samples and 3B parameters it matches OpenAI-o1 and Gemini-2.5-Pro.
在同一套文本、图像和点云问答上比较三种模态，检验点云是否真能提升大模型的空间推理。
A bias-controlled study over text, vision, and point clouds that tests whether point clouds actually improve spatial reasoning in large language models.
从航拍视角评测多模态大模型的开放空间关系推理，题目来自真实与仿真城市场景，同时支持图像和点云。
An aerial-view benchmark for spatial-relation reasoning, with questions from real and simulated city scenes and support for both images and point clouds.
用真实城市与仿真环境中的第一人称无人机视频，评测视频大模型在回忆、感知、推理和导航上的具身能力。
Evaluates whether video language models can recall, perceive, reason, and navigate from first-person drone video in real and simulated cities.
基于真实城市搭建的室外具身智能仿真平台，提供行人与车流，以及感知、规划、行动等评测任务和统一接口。
A city-scale embodied platform built from a real urban environment, with pedestrian and traffic flow, evaluation tasks, and a unified agent interface.
我喜欢把研究以外的时间花在路上：徒步、登山、攀岩，也用相机记下这个美好的世界。如果你也喜欢徒步摄影，欢迎联系我！！！
Away from research I hike, climb, and photograph this beautiful world. If you also enjoy hiking and photography, please get in touch.
