Zhihao Lin

Zhihao Lin

AI Systems Research Assistant

Research

Research interests: Large Language Models, LLM Inference and Serving, AI Systems, AI Agents, and Representation Learning.

Current work

  • AI Systems Research Assistant, University of Virginia (Sep 2026 – present).
    • Profiling vLLM-based agentic workloads to identify concurrency and inference bottlenecks, focusing on KV-cache management and serving efficiency.
  • Independent research on large language models (Jan 2026 – Sep 2026): adaptive LLM generation, representation learning, and inference-time methods.
    • Co-developed Scenario-Actor, a scenario-adaptive decoding framework built on Llama-3.1-8B-Instruct that learns scenario representations from hidden states and selects a sampling temperature for each prompt.
    • Led all experiments in PyTorch, implementing hidden-state extraction, scenario-representation training, and prompt-level temperature selection.
    • Built a Python LLM pipeline using OpenAI APIs for GPT-4o annotation of 37K+ prompts and GPT-4-turbo evaluation, with confidence-based filtering and validation against human-labeled gold sets.
    • Paper: Read the Room Before Your Model Generates. Muyao Kong*, Zhihao Lin* (*equal contribution). Accepted to the AACL-IJCNLP 2026 Main Conference. [arXiv]

Prior experience

  • Research Assistant, Santa Clara University (Jan 2022 – Apr 2022), advised by Prof. Yi Fang.
    • Sentiment Analysis for Email Interaction
      • Contributed to an NLP research project on sentiment analysis for email interaction through literature review, research discussions, and experimental model deployment on AWS EC2.
      • Built a Chrome/Gmail sentiment-analysis plugin prototype that integrated NLP inference into email workflows for lightweight sentiment interpretation.
    • Reporter Image Analysis for Fake-News Research
      • Developed reusable research software that crawled web images associated with news reporters and combined Amazon Rekognition with DeepFace for face attribute analysis, producing structured reporter data for downstream fake-news detection research; packaged the software for use by a PhD researcher.