Aleksandar (Suny) Shtedritski

Hi, I am Suny! I am joining Microsoft AI as Member of Technical Staff. Before that, I did a DPhil (PhD) at the Visual Geometry Group, Oxford, where was lucky to be advised by Prof. Andrea Vedaldi and Prof. Christian Rupprecht. During my PhD, I was a summer intern at Google DeepMind with Olivia Wiles.

During my PhD I started a student research group within the AI society in Oxford, OxAI Labs , where we did research on fairness and bias. Before PhD, I graduated with an MEng in Engineering Science from the Univesity of Oxford.

My nickname is to be pronounced like "Sunny" ☀️ despite missing an "n".

Email / Google Scholar / LinkedIn / Github

Research

I'm interested in (multimodal) LLMs, vision + language, un/self-supervised computer vision, 3D, and fairness.

	SynCity: Training-Free Generation of 3D Worlds. Paul Engslter^, Aleksandar Shtedritski^, Iro Laina, Christian Rupprecht, Andrea Vedaldi, ^* Equal contribution. ArXiv preprint, 2025 arXiv / website / code / demo We present a training-free and optimisation-free method for generating 3D worlds.
	Highlight: Learning Visual Prompts for Vision-Language Models. Jana Zeller Aleksandar Shtedritski, Christian Rupprecht, Coming soon website We automatically discover visual prompts for CLIP. Interestingly, they look like red circles.
	SHIC: Shape-Image Correspondences with no keypoint supervision. Aleksandar Shtedritski, Christian Rupprecht, Andrea Vedaldi ECCV, 2024 arXiv / website / code / demo We present an unsupervised method for shape-to-image correspondenes.
	HelloFresh: LLM Evaluations on Streams of Real-World Human Editorial Actions across X Community Notes and Wikipedia edits Tim Franzmeyer, Aleksandar Shtedritski, Samuel Albanie, Philip Torr, Joao F. Henriques, Jakob E. Foerster, ACL, 2024 arXiv / website / code A living benchmark for LLMs that addresses issues like test data contamination and benchmark overfitting.
	PaperQA: Retrieval-Augmented Generative Agent for Scientific Research Jakub Lala, Odhran O'Donoghue, Aleksandar Shtedritski, Sam Cox, Samuel G Rodriques, Andrew White Technical report, 2023 arXiv / code We present a RAG agent for scientific research, and a benchmark for information retrieval in sciences.
	BioPlanner: Automatic Evaluation of LLMs on Protocol Planning in Biology Odhran O'Donoghue, Aleksandar Shtedritski, John Ginger, Ralph Abboud, Ali Essa Ghareeb, Justin Booth, Samuel G Rodriques, EMNLP, 2023 arXiv / code A framework for evaluation of LLMs on long planning tasks, with an application in biology.
	What does CLIP know about a red circle? Visual prompt engineering for VLMs Aleksandar Shtedritski, Christian Rupprecht, Andrea Vedaldi ICCV, 2023 (Oral Presentation) arXiv / code We discover an emergent ability of CLIP, where drawing a red circle focuses the global image description to the region inside the circle.
	Learning Universal Semantic Correspondences with No Supervision and Automatic Data Curation Aleksandar Shtedritski, Andrea Vedaldi, Christian Rupprecht ICCV Workshop ion Representation Learning with Limited Data, 2023 (Oral Presentation) paper / We present a method for learning robust and generalizable semantic correspondences.
	VisoGender: A dataset for benchmarking gender bias in image-text pronoun resolution Siobhan Mackenzie Hall, Fernanda Gonçalves Abrantes, Hanwen Zhu, Grace Sodunke, Aleksandar Shtedritski, Hannah Rose Kirk NeurIPS Datasets and Benchmarks, 2023 arXiv / code A new dataset for benchmarking gender bias in vision-language models.
	Balancing the Picture: Debiasing Vision-Language Datasets with Synthetic Contrast Sets Brandon Smith, Miguel Farinha, Siobhan Mackenzie Hall, Hannah Rose Kirk, Aleksandar Shtedritski, Max Bain NeurIPS Workshop SyntheticData4ML, 2023 arXiv / code We demonstrate that the datasets used to evaluate the bias of VLMs are themselves biased. We propose to debias these datasets using synthetic contrast sets.
	A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning Hugo Bergh, Siobhan Mackenzie Hall, Wonsuk Yang, Yash Bhalgat, Hannah Rose Kirk, Aleksandar Shtedritski, Max Bain AACL-IJNCLP, 2022 arXiv / code We propose a lightweight method to debias CLIP.
	Bias out-of-the-box: An empirical analysis of intersectional occupational biases in popular generative language models Hannah Rose Kirk, Yennie Jun, Haider Iqbal, Elias Benussi, Frederic A. Dreyer, Aleksandar Shtedritski, Yuki M. Aasano NeurIPS, 2021 arXiv / code An in-depth analysis of instersectional biases of GPT-2.

Services

Reviewer

NeurIPS, ICLR, CVPR, ICCV, ECCV

Teaching Assistant

Information Engineering tutorial classes, Engineering Department, University of Oxford, 2021-22
Image and Signal Processing lab, Engineering Department, University of Oxford, 2022
Machine Learning tutorial classes, Department of Computer Science, University of Oxford, 2020-21
Machine Learning revision classes, Magdalen College, University of Oxford, 2020-21
Control Systems lab, Engineering Department, University of Oxford, 2021

This website template is borrowed from Jon Barron. Thanks!