I’m Miikael Mändmets, a master’s student at the University of Turku. My thesis, “Auditing the VLM-as-Judge Protocol for Text-in-Image Style Editing: Sensitivity to Controlled Degradations,” audits whether the vision-language models used as automatic judges in text-in-image editing benchmarks can actually perceive the errors they’re asked to penalise.
Posts on this site cover that work and what building it taught me.