CV
PDF format of my CV is available: Curriculum Vitae, 个人简历.
Contact: aroma@mail.dlut.edu.cn · +86 15842668147 · recynie.github.io
Education
- Dalian University of Technology, School of Future Technology
- Undergraduate in Artificial Intelligence, 2023–2027
Academic performance
- Academic score: 87.0/100
- First five semesters academic rank: 13/64
- Second-year academic rank: 4/64
Language tests: CET-4 (618), CET-6 (591)
Selected coursework: Probability and Statistics (96), Deep Learning (96), Programming for Artificial Intelligence (92)
Research experience
Language-guided expertise evolution for protein optimization (Apr. 2025 – May 2026)
Advisor: Prof. Yanming Shen
- Developed a self-evolving agent system for protein optimization through explore–verify–refine cycles, accumulating executable prior knowledge in a code-based knowledge base.
- Represented knowledge as a code population, with gating mechanisms to match and optimize target protein sub-distributions.
- Used Thompson sampling during training and lower confidence bound (LCB) scoring for final knowledge selection to balance exploration and exploitation.
- Built an agent workspace containing parent-node strategies, validation results, and metrics to support feedback-driven exploration and analysis.
- Built a high-concurrency agent system from scratch, decoupling exploration from knowledge updates and supporting multiple servers for efficient resource use.
- An early version won the Best Innovation Award at the Bohr + SciMaster AI4S Agent Development Competition (only two teams awarded) and reached the MiraclePlus startup camp interview round (top 10%). This work led to co-first-author papers accepted at ICLR 2026 RSI Workshop and NeurIPS 2026.
Molecular representation learning (Oct. 2024 – Dec. 2024)
Advisor: Prof. Yanming Shen
- Integrated molecular motifs into GNNs and LLMs to improve representation learning beyond the limits of 1-WL testing and one-dimensional SMILES representations.
- For Uni-Motif, designed and evaluated three substructure feature fusion mechanisms and compared motif decomposition strategies.
- For S2Token, proposed and evaluated substructure-aware tokenization and reproduced leading baselines to validate motif-augmented performance gains.
- Co-authored papers published in Expert Systems with Applications and Information Processing & Management.
Protein hydration and crystallization condition prediction (Dec. 2023 – Present)
Advisor: Prof. Song Xue
- Modeled relationships among protein sequences, structures, and chemical environments to predict crystallization conditions.
- Built a crystallization condition database of roughly 7,000 proteins by extracting, filtering, and integrating records from the Protein Data Bank.
- Used Foldseek to cluster structurally similar proteins and analyze solution-component combinations and crystallization condition distributions.
- Fine-tuned SaProt, a pretrained structure-aware protein language model, on the crystallization condition task.
- Found that protein language models struggle to represent crystallization conditions. Current work adds hydration as a downstream modality and develops hydration-aware protein representations.
Publications
From Graphs to Tokens: Substructure-Aware Molecular Representation for Large Language Models
Runze Wang, Zijie Xing, Xingyue Liu, Mingqi Yang, Che He, Yanming Shen*, “From Graphs to Tokens: Substructure-Aware Molecular Representation for Large Language Models,” Information Processing and Management, vol. 63, no. 6, p. 104771, Sep. 2026
Language-guided expertise evolution for protein optimization
Xingyue Liu†, Zijie Xing†, Runze Wang, Luoming Hu, and Yanming Shen*, “Language-guided expertise evolution for protein optimization,” ICLR 2026 Workshop on AI with Recursive Self-Improvement, 2026
Motif-driven molecular graph representation learning
Runze Wang, Yuting Ma, Xingyue Liu, Zijie Xing, and Yanming Shen*, “Motif-driven molecular graph representation learning,” Expert Systems with Applications, vol. 269, p. 126484, 2025
Personal projects
Self-Aware Agent (Mar. 2026) — GitHub · Project details
- Explores agent-controlled context management by formalizing chat history, tool calls, skills, and other inputs as context items that the agent can manage and assemble.
pi-goal (Aug. 2026) — GitHub · Project details
- Developed a verifiable
/goalextension for long-horizon tasks, with interactive requirement clarification and explicit deliverables and termination criteria. - Uses an independent fresh-context verifier to check outputs against the agreed criteria, improving alignment and reducing premature completion.
Competitions and awards
- Sep. 2025: National First Prize, Mechanical Engineering Innovation & Creativity Competition – AI-based Non-Destructive Testing Image Evaluation Track (team leader). Detected weld defects in X-ray images; handled data preprocessing, model deployment, and fine-tuning of YOLOv8, RT-DETR, and DEIM.
- Jun. 2025: First Prize, Dalian University of Technology Information Security Competition (team leader). Designed a benchmark and testing platform for LLM text watermark attacks; proposed the project direction, built the code framework, and reproduced paraphrasing attacks. The team reproduced five attack papers.
Scholarships
- National Scholarship, 2024–2025
- Academic Excellence Scholarship, 2024–2025
- Academic Excellence Scholarship, 2023–2024
Research interests
- Autonomous scientific research: Improve AI’s long-horizon research ability through iterative idea development with experimental feedback to propose valuable, feasible ideas.
- Protein representation: Develop structure-aware and hydration-aware protein representations to improve downstream task performance.
Skills
- AI agents: Self-evolving agents, agent harness design
- AI for Science: Protein representation learning, molecular representation learning, protein optimization
