Kai Bian

Kai Bian

Undergraduate Researcher in LLM Agents and Efficient Inference

I am an undergraduate student in Computer Science and Technology at Harbin Institute of Technology, Shenzhen (HITSZ), advised by Prof. Xuebo Liu. My current research focuses on LLM vocabulary optimization, LLM agents, multi-agent systems, and efficient LLM inference.

My recent work, AgentVocab, studies how vocabulary adaptation can reduce tokenization overhead in structured tool-calling scenarios for LLM agents, improving inference efficiency while preserving tool-calling performance.

🔥 News

  • 2026.07: Launched this personal homepage.
  • 2026: AgentVocab: Structure-Aware Vocabulary Adaptation for Efficient LLM Agents was accepted to ICML 2026.

📝 Publications

ICML 2026
AgentVocab framework overview

AgentVocab: Structure-Aware Vocabulary Adaptation for Efficient LLM Agents

Kai Bian, Haosi Mo, Xuebo Liu, Shuangyong Song, Jing Li, Yongxiang Li, Min Zhang, Xuelong Li

OpenReview / PDF / Code

  • Introduces a structure-aware vocabulary adaptation framework for efficient LLM agents.
  • Derives specialized vocabulary entries from real tool-calling traces without task-specific schema engineering.
  • Reduces decoding latency relative to the vanilla baseline by 17.7% on τ-bench and 19.5% on τ²-bench while preserving tool-calling performance.

💻 Research & Projects

  • Research Assistant, Research on LLM Vocabulary Optimization Mechanisms for Single-Agents, Harbin Institute of Technology, Shenzhen.
    Sep. 2025 - Present. Advisor: Prof. Xuebo Liu.

  • Technical Leader, Resource-Efficient and Trustworthy Text Analysis System Based on LLMs.
    Built a four-stage LLM processing pipeline for large-scale financial text analysis, reducing processing time from 33.8 hours to 0.5 hours and achieving 99.74% accuracy with an anti-hallucination mechanism.

  • Project Leader, Multi-Agent System for Automated Digital Human Video Generation.
    Designed a multi-agent collaboration framework for text, speech, visual rendering, and composition, reducing the production cycle from 7-10 days to within 1 hour.

  • Project Leader, End-to-End LLM-Driven Automated Digital Human Generation System.
    Built a modular web-based generation pipeline integrating Qwen, GPT-SoVITS, SadTalker/MuseTalk, and FFmpeg.

📖 Education

  • Harbin Institute of Technology, Shenzhen (HITSZ), B.Eng. in Computer Science and Technology.
    Sep. 2023 - Jun. 2027. CGPA: 3.751/4.0.

🎖 Honors and Awards

  • 2025.10: Second Prize of Guangdong Province, China Undergraduate Mathematical Contest in Modeling.
  • 2025.05: Second Prize of HITSZ, National Collegiate Statistical Modeling Contest.
  • 2025.03: Third Prize and Excellent Award of HITSZ, “Challenge Cup” National College Student Extracurricular Academic and Technological Works Competition.

📄 CV