Greetings! I am Yunze Song, a research intern at Tencent and a graduate student at the National University of Singapore.
Publications
TrustJudge: Inconsistencies of LLM-as-a-Judge and How to Alleviate Them
Paper
·
Code
·
Yidong Wang*, Yunze Song*, Tingyuan Zhu, Xuanwang Zhang...
We identify two fundamental inconsistencies in LLM-as-a-judge evaluation: Score-Comparison inconsistency and Pairwise
Transitivity inconsistency. We propose TrustJudge, a probabilistic framework with distribution-sensitive scoring and
likelihood-aware aggregation to alleviate these issues and improve reliability of automated assessment.
RAGLAB: A Modular and Research-Oriented Unified Framework for Retrieval-Augmented Generation
Paper
·
Code
·
Xuanwang Zhang*, Yunze Song*, Yidong Wang, Shuyun Tang...
Two major obstacles hinder RAG's progress: (1) lack of thorough and fair comparisons between emerging RAG algorithms;
(2) high-level abstractions in open-source tools reduce transparency and limit innovation. We introduce RAGLAB, a modular,
research-focused open-source library that replicates existing algorithms and provides an ecosystem for exploring RAG techniques,
enabling fair comparisons across benchmarks.
Internships
- 2025.12 – now Research Intern, Tencent
- 2023.11 – 2024.11 Research Intern, Peking University
- 2023.05 – 2023.09 Research Intern, Westlake University
