Theses and Dissertations
ORCID
https://orcid.org/0009-0008-7011-5133
Advisor
Zhiqian, Chen
Committee Member
Jingdao, Chen
Committee Member
Gudla, Charan
Date of Degree
5-15-2026
Original embargo terms
Immediate Worldwide Access
Document Type
Graduate Thesis - Open Access
Major
Computer Science (Artificial Intelligence & Robotics)
Degree Name
Master of Science (M.S.)
College
James Worth Bagley College of Engineering
Department
Department of Computer Science and Engineering
Abstract
This thesis extends data contamination auditing for multimodal large language models to multilingual settings. Using LLaVA 1.5 and a high-fidelity French parallel dataset derived from ScienceQA, the study evaluates how performance changes when identical image-question pairs are translated from English to French. The resultsshow a substantial cross-lingual performance decline and frequent flips from correct English predictions to incorrect French predictions, indicating that benchmark performance can depend heavily on memorized English-specific patterns rather than stable multimodal reasoning. To address this weakness, the thesis introduces an inference-time mitigation strategy based on perturbation ensembling and cross-lingual consistency aggregation. The proposed method reduces instance-level leakage without model retraining and offers a practical way to improve the robustness and trustworthiness of multimodal benchmark evaluation. The findings demonstrate the importance of cross-lingual auditing when assessing modern multimodal systems.
Recommended Citation
Adapa, Pavan Dharma, "Beyond English: Auditing and mitigating cross-lingual data contamination in multimodal large language models" (2026). Theses and Dissertations. 6851.
https://scholarsjunction.msstate.edu/td/6851
