Ph.D. student · Peking University
I work with Prof. Houfeng Wang at the Institute of Computational Linguistics, Peking University.
I study LLM evaluation and agent reliability, alongside iterative coding and multimodal benchmarks. My current focus is trace-based diagnosis of long-horizon agent failures. I also collaborate on post-training refinement of LoRA adapters.
Overconfidence in LLM-as-a-Judge published at NLPCC 2026.
Fragility Under Pressure accepted to COLM 2026.
Sheetpedia at NeurIPS 2025, Datasets & Benchmarks Spotlight.
I study when tool-using agents report task progress that is not supported by the available evidence. I develop trace-based methods to diagnose mismatches between observations, actions, and completion claims, and to distinguish these from tool or execution-environment failures.
I lead the research design, implementation, experiments, analysis, and writing.
I collaborate on post-training refinement of LoRA adapters, studying how redistributing adaptation gains affects task specialization and the retention of broader capabilities.
My contribution focuses on running a subset of the experiments.
* Equal contribution
STRIDE evaluates whether LLMs preserve instructions and functionality through repeated edit, verify, and debug cycles, rather than only on their first attempt.
I led the benchmark design, data construction, evaluation implementation, experiments, analysis, and paper writing.
A study of nine LLM judges distinguishes confidence calibration from its usefulness for ranking judgments. LLM-as-a-Fuser combines decisions, rationales, and confidence signals.
I co-designed the evaluation methodology and conducted experiments on confidence estimation and confidence-aware judgment fusion.
A corpus of more than 290,000 real-world spreadsheets, with benchmarks for semantic-range understanding and formula generation.
I worked on data cleaning and standardization, language and quality filtering, and LSH-based deduplication. I also contributed to fine-tuning experiments, analysis, and writing.
A benchmark that examines whether vision-language models understand the text they recognize. I contributed to scene-text data collection, the annotation workflow, and English question-answer construction.
Ph.D. student in Natural Language Processing
Institute of Computational Linguistics
Advisor: Prof. Houfeng Wang
B.S. in Computer Science and Technology
Yuanpei College
Research intern · FlagEval
Worked on evaluation datasets for vision-language models, focusing on text recognition, text understanding, and annotation quality.