Jiaxin Yang

Agentic RL · Agent Design · Agent Evaluation · Embodied AI

Agent Systems Reinforcement Learning Embodied AI Evaluation

---

About Me

我目前的研究兴趣主要集中在 Agentic RL、Agent design、Agent evaluation 和 Embodied AI。除了研究方向外，我也很重视研究代码的可复现性、评测体系设计和工程化落地。

My primary interests are Agentic RL, agent design, agent evaluation, and embodied AI, with a strong emphasis on reproducible research workflows and evaluation-oriented engineering.

Focus

• Agentic Reinforcement Learning
• Agent design and agent systems
• Agent evaluation and benchmarking
• Embodied AI / embodied intelligence
• VLM-based 3D reconstruction

Tech Stack

Selected Projects

• shushu-internship-resume-optimizer: A practical project for improving internship resume content and presentation through structured optimization workflows.
• dense-physical-map-code: Benchmark-oriented code for dense physical map prediction and evaluation, with an emphasis on reproducibility and research workflow organization.

Competition

• MABe Challenge - Social Action Recognition in Mice: A Kaggle research competition on detecting social behaviors in mice from pose estimates, where we built single-mouse and pairwise interaction features from pose sequences and combined XGBoost, CatBoost, and LightGBM in an ensemble pipeline with threshold and weight optimization. Achieved a Bronze Medal, ranking 138 / 1412 teams (Top 10%).
• Second World Scientific Intelligence Competition - Logical Reasoning Track: Complex Reasoning Ability Evaluation, jointly hosted by SAIS and Fudan University. We built a lightweight LLM adaptation pipeline with LoRA, majority-voting ensembling, and vLLM-based inference acceleration, improving accuracy from 0.72 to 0.885 and earning the Meritorious Award, ranking 8 / 2815 teams (Top 1%).

Contact

• GitHub: @Sunanzhe2004
• Email: [email protected]

Collapse