Hi, I’m Ahmet Fırat Gamsız, a senior computer engineering student at Boğaziçi University. My research interests span the interaction between language models and the real world, focusing on robotics, interpretability, generative AI, and applications in education and creativity. I’m currently applying for funded MSc and PhD positions in these areas.

For the past year, I’ve been working in the CoLoRs lab, led by Prof. Emre Uğur, where I study the intersection of robotics and large language models. My main project involves implementing a framework for task detection and inverse task planning using vision-language models. See our recent paper: Backwards Planning from Onward Task Demonstrations via Vision-Language Models (see here). Currently, my focus is on how different robotic embodiments—such as navigation or manipulation platforms—affect the behavior of language models in grounded planning tasks.

Previously, I completed a summer internship at CIS@LMU Munich, under the supervision of Prof. Hinrich Schütze. Together with my PhD advisor, Abdullatif Köksal, I developed a framework to generate a large-scale educational conversation dataset with 40,000 samples. The resulting paper is currently under preparation: May I Ask a Question? MIA40K: A Large-Scale Educational Conversation Dataset and Generation Pipeline (see here).

Beyond my ongoing work, I’m also interested in the internal structure of language models—particularly in interpretability research that investigates low-level mechanisms and representations. I have research interests in synthetic data generation, creativity in language models, and generative AI, and I’m looking to gain hands-on experience with diffusion models as well. I also have a background in game development, and I’ve been exploring how AI can be used in interactive storytelling—most recently experimenting with using LLMs as dungeon masters.

Earlier in the year, I worked on a senior thesis project focused on step-by-step solution generation for geometry problems using vision-language and diffusion models, co-supervised by Prof. Suzan Üsküdarlı and Abdullatif Köksal.