At IIT Bhubaneswar, I am currently working on VLM evaluation, reasoning benchmarks, and new multimodal LLM applications.

• Vision-Language Model Evaluation: Recently, we pioneered a math and science multimodal and bilingual benchmark that tests VLMs in an exam-style evaluation setting. Through extensive testing, we found interesting metacognitive behavior patterns and reasoning gaps between open and closed models. (Accepted to IJCNLP-AACL 2025 and NeurIPS 2025 MATH-AI workshop)

• Multimodal Cultural Competence: Created the first systematic evaluation framework for VLM cultural competence through multimodal story generation. Analyzed 5 contemporary VLMs with novel evaluation metrics. (Accepted to ICCV 2025 ASI workshop)

• Modality Translation Framework: Designed UNITE, a VLM-in-the-loop framework that achieved state-of-the-art FakeNewsNet, Fakeddit, and Hateful Memes performance.