Research
My research centers on reinforcement learning and multimodal reasoning, with data and compute efficiency as a shared goal.
Research Interests
- Reinforcement learning: efficient policy optimization for reasoning
- On-policy distillation: learning from token and attention signals
- Multimodal learning: vision-language models and medical AI
- Ongoing interests: world models, sampling, and inference-time scaling
Policy Optimization & Distillation
PrePO predicts a farther point on a same-batch optimization trajectory from two observed updates, then repositions and corrects. OPASD adds solution-conditioned attention supervision to on-policy self-distillation.
Undergraduate Thesis
AI-driven dermatological diagnosis using vision-language models
Supervised by Prof. Mohammad Ariful Haque, with collaborators Aranya Saha and Tanvir Ahmed Khan.
My undergraduate work studied dermatological reasoning through supervised fine-tuning, QLoRA, and DPO. Related work includes GRPO++, a stabilized policy-optimization pipeline for medical VLMs, and CLARIFY, a specialist–generalist framework for dermatological question answering.