I’m Haz 👋

I’m a CS PhD student at the University of California, Riverside, advised by Yue Dong. I study how multimodal and long-context models actually use what they see and read — and when that behavior becomes unsafe.

My current work spans multimodal interpretability and safety, long-context safety, and AI for science. That includes figuring out when vision-language models skip visual detail in favor of language shortcuts, how models fail to refuse harmful goals they infer over long contexts, and how generative models can be used for problems like antibody design and RNA modeling.

I spent the summers of 2025 and 2026 as an Applied Scientist intern at Amazon in San Diego, working on buyer-risk prevention. Before UCR, I completed my BSc in Computer Science and Engineering at Bangladesh University of Engineering and Technology.

News

• [10/2026] VLMs Need Words was accepted at COLM 2026.
• [09/2026] Two papers accepted at NeurIPS 2026: AbFlowNet and Do Reasoning LLMs Refuse What They Infer in Long Contexts?.
• [07/2026] D-Judge was accepted at ICML 2026.
• [02/2026] Harnessing the Unseen was accepted at AAAI 2026.
• [11/2025] BiRNA-BERT was published in Nature Communications Biology.
• [11/2025] ExpertGenQA was accepted at Findings of EMNLP 2025.
• [09/2025] Our OCT biomarker detection paper was accepted at ICIP 2025.
• [08/2024] Co-presented the ACL 2024 tutorial Vulnerabilities of Large Language Models to Adversarial Attacks.