About
Songze Li (李松泽) is a master student in the School of Software Technology, Zhejiang University (ZJU). My research focuses on pre-trained language models, knowledge representation and reasoning, multimodal large language models, and GUI agents at the Knowledge Engine Laboratory, OpenKG, under the supervision of Prof. Wen Zhang. I completed my undergraduate studies at the School of Computer Science and Technology, Tongji University. I am seeking like-minded collaborators. If you are interested, please feel free to reach out via email.
Selected Publications
View All →Scaling LLM Knowledge Boundaries via Distribution-Optimized Synthesis
Songze Li, Yarong Lan, Zhongpu Bo, Zhaoyang Wang, Zhiqiang Liu, Yuan Yuan, Chengtao Gan, Menghao Qian, Enpei Niu, Xiaoke Guo, Yuanxiang Liu, Zhaoyan Gong, Xiangjin Hu, Liangyurui Liu, Jingdian Lu, Lei Liang, Jun Zhou, Huajun Chen, Wen Zhang†
EMNLP 2026 Submission
We propose Knowledge Distribution-Optimized Synthesis (KDoS), a synthetic data framework that controls knowledge distribution, and find that an optimal knowledge distribution consistently exists across model sizes, data scales, and backbone architectures, maximizing the efficiency of scaling up LLM knowledge boundaries.
Last Layer Logits to Logic: Empowering LLMs with Logic-Consistent Structured Knowledge Reasoning
Songze Li, Zhiqiang Liu, Zhaoyan Gong, Xiaoke Guo, Zhengke Gui, Huajun Chen, Wen Zhang†
EMNLP 2026 Submission
This paper presents Logits-to-Logic, a novel output-perspective approach addressing the Logic Drift of large language models (LLMs) in structured knowledge reasoning, with deeper insights into maintaining logic-consistent reasoning in LLMs.
What's Missing in Screen-to-Action? Towards a UI-in-the-Loop Paradigm for Multimodal GUI Reasoning
Songze Li, Xiaoke Guo, Tianqi Liu, Biao Yi, Zhaoyan Gong, Zhiqiang Liu, Huajun Chen, Wen Zhang†
ACL 2026 Findings
We propose UI‑in‑the‑Loop, which reformulates GUI reasoning from Screen‑to‑Action into a cyclic Screen-UI Elements-Action paradigm, and introduce the UI Comprehension task with three metrics and the 26K UI Comprehension-Bench benchmark.
Enrich-on-Graph: Query-Graph Alignment for Complex Reasoning with LLM Enriching
Songze Li, Zhiqiang Liu, Zhengke Gui, Huajun Chen, Wen Zhang†
EMNLP 2025 Main
We propose Enrich-on-Graph (EoG), a flexible framework that leverages LLMs' prior knowledge to enrich knowledge graphs and bridge the semantic gap between structured KGs and unstructured queries for complex KGQA reasoning.
News
🎉🎉 Four papers (UILoop, CoG, ASTRA, Temp-R1) have been accepted to ACL 2026.
🎉🎉 One paper (EoG) have been accepted to EMNLP 2025.
Starting my Master in Zhejiang University