Agentic Models
Building agentic models from scratch at Tencent Hunyuan through the Qingyun Program.
AgentsFrom-scratch training
Ph.D. student · Tsinghua IIIS
Hello, I’m Kexian Tang.
I am Kexian Tang, a Ph.D. student at Tsinghua University, advised by Prof. Kaifeng Lyu. I work on data-centric LLM training and multimodal reasoning. Since July 2026, I have also been a research intern at Tencent Hunyuan through the Qingyun Program, working on building agentic models from scratch.
Beijing, China
Curious about data, reasoning & vision.
01 / About
About me
My research asks a practical question: what should a model learn from, and in what order? I study synthetic data, data selection and mixing, reasoning-data construction, and multi-stage training—alongside spatial reasoning and visual editing in multimodal models.
02 / Research
Research interests
Data-Centric LLM Training
Efficient knowledge injection through synthetic data, data selection, and thoughtful data mixing.
Synthetic dataData selectionData mixing
Reasoning Models
Constructing high-quality reasoning data and designing multi-stage training recipes.
Reasoning dataMulti-stage training
Multimodal LLMs
Improving how vision-language models understand space, transformations, and visual intent.
Spatial reasoningVisual editing
03 / Publications
Selected work
* Equal contribution
SPA: A Simple but Tough-to-Beat Baseline for Knowledge Injection
Beyond Safe Data: Pretraining-Stage Alignment with Regular Safety Reflection
LEGO-Puzzles: How Good Are MLLMs at Multi-Step Spatial Reasoning?
Integrated into VLMEvalKit, an open-source evaluation toolkit.
Project page ↗Envisioning Beyond the Pixels: Benchmarking Reasoning-Informed Visual Editing
Oral presentation · Top 0.35%
Reliable Cosine Matching via Neighborhood Consensus for Fundus Image Registration
Oral presentation · Top 1.9%
More collaborative work +
2025
InternVL3.5
Data mixing strategies for cold-start training.
2025
Intern-s1
Synthetic data construction for cold-start training.
04 / Path
Experience & education
Jul 2026 — now
Tencent Hunyuan
Research Intern · Qingyun Program
Working on building agentic models from scratch.
2025 — now
Tsinghua University
Ph.D. Student, Institute for Interdisciplinary Information Sciences (IIIS)
Adviser: Prof. Kaifeng Lyu
2024 — 2025
Shanghai AI Laboratory
Research Intern, Intern Team
Supervisors: Haodong Duan and Yanhong Zeng
2021 — 2025
Tongji University
B.Eng. in Computer Science and Technology
GPA 4.97/5.00 · Ranked 1/226 (Top 0.44%)
05 / Recognition
Awards
- Outstanding Bachelor’s Degree Graduate of Shanghai
- National Scholarship
- Silver Medal, China Collegiate Programming Contest for Girls
- Outstanding Student Pioneer, Tongji University · Top 0.5%
- First Prize, iFLYTEK A.I. Developer Competition
Beyond the lab
Outside research, I play the piano.
I previously served as President of Tongji University’s CS College Student Union and Piano Association.
Say hello