Xinzhe Li
Postdoctoral Research Fellow, RMIT University
I am a Postdoctoral Research Fellow at RMIT University and hold a PhD in Information Technology from Deakin University. My research centers on large language model (LLM) multi-trajectory reasoning — how models explore and aggregate multiple reasoning paths at test time (including tree search), and how to close the loop by folding such trajectories back into training. I have first-author publications at ACL, TMLR, COLING, and IJCAI.
Research Interests
- Inference-time multi-trajectory reasoning. How LLMs explore and
aggregate multiple reasoning paths at test time — including tree search — while
maintaining both effectiveness and efficiency.
- Adaptive branching. Deciding when to branch during search rather than expanding at every step, substantially reducing compute with little or no loss in accuracy.
- Cross-trajectory agent memory. Transferring useful knowledge across reasoning attempts for tool-use agents, organized along the scope of transfer and the abstraction of content.
- Training-time: closing the loop between inference and learning. Folding search- and memory-generated trajectories back into training. (emerging direction)
News
- May 2026 Two papers accepted at ACL 2026: LiTS (Demo) and Chain-in-Tree (Findings).
- Oct 2025 Joined RMIT University as a Postdoctoral Research Fellow.
- Apr 2025 Survey on LLM test-time compute via search accepted at TMLR.
- Mar 2025 Completed PhD in Information Technology at Deakin University.
- Jan 2025 Review of LLM-based agent paradigms presented at COLING 2025.
Publications
All as first author.
- When Does Memory Help Multi-Trajectory Inference for Tool-Use LLM Agents? Under review arXiv · code
- LiTS: A Modular Framework for LLM Tree Search ACL Demo 2026 arXiv · code
- Chain-in-Tree: Back to Sequential Reasoning in LLM Tree Search ACL 2026 Findings arXiv · code
- A Survey on LLM Test-Time Compute via Search: Tasks, LLM Profiling, Search Algorithms, and Relevant Frameworks TMLR 2025 arXiv · code
- A Review of Prominent Paradigms for LLM-Based Agents: Tool Use (Including RAG), Planning, and Feedback Learning COLING 2025 paper · arXiv · code
- A Survey on Out-Of-Distribution Evaluation of Neural NLP Models IJCAI 2023 paper · arXiv
- Can Pretrained Language Models Derive Correct Semantics from Corrupt Subwords under Noise? ACL-SEM 2023 arXiv · code
- Make Text Unlearnable: Exploiting Effective Patterns to Protect Personal Data ACL-TrustNLP 2023 arXiv · code
- Exploring the Vulnerability of Natural Language Processing Models via Universal Adversarial Texts ALTA 2021 paper · code
- GRAMMAR: Grounded and Modular Methodology for Assessment of Closed-Domain Retrieval-Augmented Language Model Preprint arXiv · code
Teaching
- 2023–2025 SIT720 Machine Learning, Deakin University
- 2024 SIT744 Deep Learning, Deakin University
Academic Service
Program Committee Member / Reviewer: NeurIPS (2026), IJCAI (2024, 2025), COLING (2025).