About Me
I am currently a master's student at Beijing University of Posts and Telecommunications, advised by Prof. Quan Fang and Prof. Jun Hu. In 2024, I earned dual bachelor's degrees in Internet of Things Engineering through the joint programme between BUPT and Queen Mary University of London (QMUL).
In March 2026, I joined Kuaishou as a research intern working on generative recommendation, mentored by Peng Jiang.
Research interests
- Embodied World Models — Learning physical relationships through observation and action, so an agent can anticipate the consequences of its actions when objects or conditions change.
- Physical Representation Learning — Understanding how physical knowledge is represented and organized inside models, why encoded information sometimes goes unused, and how that organization can improve generalization.
- Sim-to-Real Transfer (Sim2Real) — Predictions and simulations can be internally consistent yet diverge from physical reality. I am interested in the assumptions and omissions behind this gap, and what it reveals about our own reasoning.
I am seeking Ph.D. opportunities for Fall 2027. If you are interested in working with me, please feel free to email me.
I am guided by Bertrand Russell's three passions: the longing for love, the pursuit of knowledge, and compassion for human suffering.
News
- The FUV manuscript on decodability, model usage, and targeted adaptation is now available.
- SPRII, our work on persistent representations across independent interactions, is available on arXiv.
- Joined Kuaishou as a research intern.
- CRANE was published in ACM TOMM.
Education
|
M.S. in Intelligent Science and TechnologyBeijing University of Posts and Telecommunications (BUPT)
Expected graduation: June 2027
|
Dual Bachelor's Degrees in Internet of Things EngineeringBeijing University of Posts and Telecommunications (BUPT) & Queen Mary University of London (QMUL)
Joint undergraduate programme
Research Experience
|
KuaishouCommercialization Algorithms DepartmentResearch Intern in Generative Recommendation · Mentor: Peng Jiang
- Worked on generative recommendation for video-ad retrieval, including structured tool use, KV-cache distillation, and compressing image–text histories into fixed-length memory tokens.
- Contributed to SinkRec's separation of recurring behavior patterns from changing interests, and developed RoFiGR's collaborative semantic identifiers and level-adaptive distillation. Both methods were deployed and evaluated online.
Selected Research Projects
Representation Analysis and Planning with Visual World Models
Evaluated the released DINO-WM models in closed-loop planning on PointMaze and Wall, using zero-input and random-action controls to distinguish prediction, planning, and control performance.
Trained patch- and CLS-based models on PushT. In development-set candidate-action tests, CLS predictions better matched future representation-distance rankings, while patch predictions better matched rankings by actual task cost.
Action-Delay Robustness of Robot Policies
Tested Octo-small and Octo-base on four simulated manipulation tasks in SIMPLER, applying a 200 ms delay to all actions and comparing against the baseline with matched initial states and policy random seeds.
In held-out scenes, Octo-small's spoon-task success rate fell from about 31% to 17%; the effect varied across tasks and seeds.
Selected Publications & Manuscripts
Google Scholar
Shaping Persistent Representations from Independent Interactions
ICLR 2027Under review
Learns reusable system information by aligning context across separate interactions with the same system and predicting one interaction from another. Tests whether that information is encoded, used in prediction, and useful for downstream tasks.
Decodability, Usage, and the Limits of Targeted Adaptation
TMLRUnder review
Separates information formation, model usage, and task value, then tests whether these diagnostics guide targeted adaptation. Similar decodability can yield different adaptation gains; even useful decoded information does not guarantee that limited-budget retraining will repair the model.
Modality-Guided Mixture of Structured Experts with Entropy-Triggered Routing for Multimodal Recommendation
ACM TOISConditionally accepted · 2026
RoFiGR: Coarse Routing, Fine Refinement via Level-Adaptive Distillation in Multimodal Generative Recommendation
WWW 2027Under review
Combines collaborative semantic identifiers with level-adaptive distillation, translating item preferences into supervision for coarse-to-fine routing.
SinkRec: Mitigating Semantic State Sink in Long Sequence Recommendation with Memory-Conditioned Gated Delta Networks
SinkRecarXiv preprint · 2026
Stores recurring behavior patterns in memory so the recurrent state can focus on changing interests, suppressing repeated writes and subtracting already-covered information at readout.
Beyond Late Fusion: Reciprocal Critique and Revocable Decoding for Multimodal Generative Recommendation
RECAPManuscript · 2026
Using cross-modal evidence to locate and repair decoding errors while preserving the original recommendation backbone.
Awards
| First-Class Outstanding Student Scholarship
Beijing University of Posts and Telecommunications
| First-Class Outstanding Student Scholarship
Beijing University of Posts and Telecommunications
| Outstanding Graduate (Undergraduate)
Beijing University of Posts and Telecommunications
| Second-Class Outstanding Student Scholarship
Beijing University of Posts and Telecommunications
Academic Services
Journal Reviewer
- ACM Transactions on Information Systems (TOIS)
Ecommerce & Business
Internships
|
TAL Education GroupAnalyzed competing learning tablets to refine product positioning and develop Douyin creator partnerships for launch.
|
Wanwu XinxuanSelected home products, identified their selling points, and worked with suppliers on product information and supply arrangements.
Independent Stores
|
TaobaoRan a Taobao store, managing product selection, promotion, and order fulfillment, with total sales exceeding RMB 100,000.
|
PinduoduoSold kitchenware and pet products, adjusting product selection and advertising spend based on returns, with daily sales exceeding RMB 500.
|
1688Sourced products from factories and handled custom and wholesale orders. The store reached over 500 daily organic visitors within a month.