Fatemeh Pesaran Zadeh

Hello! I am an AI researcher in the Vision & Learning Lab at Seoul National University, advised by Prof. Gunhee Kim. My research centers on post-training methods, including reinforcement learning, preference optimization, and data-centric approaches, for building more capable, efficient, and autonomous AI systems.

My work spans large language models (LLMs), vision-language models (VLMs), and AI agents, with a focus on developing learning algorithms that enable these systems to reason, interact with complex environments, and continually improve through experience. Recently, I have been exploring multimodal agents for real-world workflows, building systems that can perceive, reason, and act across diverse modalities to complete complex, practical tasks. Looking forward, I am interested in extending this line of work to Vision-Language-Action (VLA) models and developing continual learning algorithms that allow agents to accumulate knowledge and skills over their lifetime, bridging AI agents and embodied intelligence.

My long-term goal is to develop frontier post-training algorithms that push the boundaries of what AI systems can do and bring frontier-level capabilities into real-world use. I believe advances in reinforcement learning, agent optimization, and continual learning are key to building AI that is not only powerful in benchmarks but genuinely capable, reliable, and useful in the real world.

CV   /  Github   /  Twitter  /  LinkedIn  /  Pledge

We introduce WebArena-Pro, a challenging benchmark for evaluating autonomous web agents on complex, realistic tasks.

We propose Weasel, an importance-diversity data selection method that improves out-of-domain generalization for offline-trained web agents while reducing training cost.

We present a Korean two-task dataset for document-description-based chart generation and chart reasoning question answering, comprising 10,000 chart construction instances and 30,000 reasoning QA pairs.

We propose LPOI, the first object-aware listwise preference optimization developed for reducing hallucinations in VLMs.

We propose a hierarchical pipeline and a new dataset for chart generation. Moreover, we introducea reinforcement learning-based instruction tuning technique for chart generation tasks without requiring human feedback.

We present mRedditSum, the first multimodal discussion summarization dataset. It consists of 3,033 discussion threads where a post solicits advice regarding an issue described with an image and text, and respective comments express diverse opinions.

We propose developing an AI-assisted CCTV system that enhances surveillance with anomaly detection and face tracking, using INNODEP-provided videos.

We developed a brand detection model using YOLOv5 to detect brands in South Korea. We collected a dataset of 8.4K images of brands in South Korea and trained a YOLOv5 model to detect them.

Template based on Jon Barron's website.