Evidence Absence Is Not Evidence Insufficiency: Diagnosing NEI Construction Artifacts in Fact Verification
Jingxi Qiu, Zeyu Han, Cheng Huang
Findings of EMNLP, 2026
Jingxi Qiu
Independent Researcher · Shanghai, China
I study when a language model should trust its evidence — and when it should admit it does not have enough.
I work on the reliability of large language models: how to verify whether retrieved evidence actually supports an answer, how to detect when it is insufficient, and how to make retrieval-augmented generation selective rather than credulous. My recent work (SURE-RAG, NEI-CAP) shows that common fact-verification benchmarks carry construction artifacts that inflate "not enough information" performance, and proposes ways to audit and mitigate them.
Before that I completed an M.S. in Data Science & Analytics at Georgetown University, where I built offline reinforcement-learning pipelines over large-scale longitudinal patient records and analysed disparities in access to CAR-T therapy across 125,000+ patients. I care about models whose confidence can be trusted — in language and in medicine alike.
Email · CV · Google Scholar · GitHub
Selected publicationsall publications →
SURE-RAG: Sufficiency and Uncertainty-Aware Evidence Verification for Selective Retrieval-Augmented Generation
Jingxi Qiu, Zeyu Han, Cheng Huang
IEEE PRAI, 2026