Xinze Li

Xinze Li

Nanyang Technological University

I am a fourth-year Ph.D. student in College of Computing and Data Science (CCDS) at Nanyang Technological University (NTU). I am co-advised by Prof. Aixin Sun and Prof. Yixin Cao. Before that, I received my Bachelor of Computing Degree (with Honours) from National University of Singapore (NUS) in 2021, and worked as a project officer in S-lab, NTU until 2023.

My research focuses on how LLM/VLM agents use experience and external knowledge: episodic memory, retrieval-augmented generation (RAG), and long-context reasoning. My work has appeared at ACL, EMNLP, and NeurIPS, with 800+ Google Scholar citations.

🔥 News

  • 2026.08:  🎉🎉 Two papers accepted to EMNLP 2026 (Findings)!
  • 2025.07:  🎉🎉 Gave tutorial in SIGIR 2025 in Padova/Italy!
  • 2024.09:  🎉🎉 One paper accepted to NeurIPS 2024 as spotlight!
  • 2024.05:  🎉🎉 Two papers accepted to ACL 2024 (Findings)!
  • 2023.05:  🎉🎉 One paper accepted to ACL 2023 (Findings)!

🔎 Selected Research

Episodic memory EMNLP 2026 · Findings

EMemBench

Evaluating what agents remember from their own experience

I led the design and implementation of a benchmark generator that turns agent trajectories into balanced questions with verifiable, environment-derived answers.

Key finding. Across 15 text games and visual game environments, induction and spatial reasoning remain persistent memory bottlenecks.

Paper Code
Agent skills EMNLP 2026 · Findings

Skill-as-Pseudocode

Making skill libraries more effective for LLM agents

I designed a pipeline that refactors Markdown skills into typed pseudocode contracts with deterministic checks and concrete action templates.

Key finding. On ALFWorld with GPT-4o-mini across three seeds: 82/402 wins vs. 47/402 for Graph-of-Skills, with 22.8% fewer input tokens and 14.5% fewer LLM calls per game.

Paper Code
External knowledge Preprint

Long Context vs. RAG

Understanding when to retrieve and when to use the full context

I developed a controlled evaluation of long-context processing and RAG across tasks, retrieval methods, and context budgets.

Key finding. Long context leads on many QA tasks; summarization-based retrieval can perform comparably, while RAG has advantages on dialogue-based and general questions.

Paper Code SIGIR Tutorial

📝 Publications

📒 Preprints

🏆 Honors and Awards

  • 2023 - 2027, NTU Research Scholarship (Full Scholarship), Nanyang Technological University.
  • 2017 - 2021, NUS Science & Technology Undergraduate Scholarship (Full Scholarship), National University of Singapore.
  • 2012 - 2016, Singapore Government SM1 Full Scholarship.

📖 Education

  • 2023.08 - present, Ph.D. student, College of Computing and Data Science (CCDS), Nanyang Technological University (NTU).
  • 2017.08 - 2021.06, Bachelor of Computing with Honours (Distinction), School of Computing (SoC), National University of Singapore (NUS).

💬 Invited Talks

💻 Experience

  • 2025.08 - 2026.02, Research Intern, Shanghai AI Laboratory, Shanghai, China.
    Developed EMemBench from interactive trajectory logging and verifiable question generation to reproducible evaluation across text and visual games.
  • 2021.09 - 2023.07, Project Officer, S-Lab for Advanced Intelligence, NTU, Singapore.
    Led the Hierarchical Script Generation project and contributed to name disambiguation and causal-relation extraction for MMEKG.
  • 2019.05 - 2019.08, Full Stack Developer Intern, CGS-CIMB, Singapore.
    Built web and database workflows for an FX administration system using Python, Django, and SQL.

📚 Academic Services

  • Conference Reviewer: NeurIPS 2025, ACL ARR (ACL 2024, EMNLP 2024, NAACL 2024, ACL 2025, EMNLP 2025)
  • Journal Reviewer: TOIS