About
The short version
Show a vision-language model a wedding in Ohio and it will describe the dress, the cake, the first dance. Show it a Barong dance in Bali and ask which figure represents good, and it will still answer: fluently, immediately, and often wrongly. It has learned what a confident answer sounds like, not what it is looking at. The models learn from the internet, and the internet is not the world.
My work is to measure that gap precisely enough to fix it. Seeing Culture (EMNLP 2025) makes a model show its work: answer a question about a cultural artifact, then point to it in the image. GPT-o3 answers 91% of across-culture questions correctly, yet its grounding never rises above 32.5 mean IoU. It says the right word and points at the wrong thing.
Cultural Moment (EMNLP 2026, oral) carries the test into video, where culture unfolds in time. A model must name a Southeast Asian concept, recognize it among unlabeled video moments, and locate its sub-events. Even the strongest models clear all three stages for fewer than 30% of concepts. You can try the three stages yourself.
How I got here
Bursa → Siena and Naples → Valencia → Bursa → İstanbul → London → Singapore → Bristol → Singapore
I grew up in Bursa, Türkiye, and studied electronics engineering at Uludağ University, with exchange semesters in Siena and Naples. Then came a software traineeship in Valencia, a master’s on vehicle detection back in Bursa, machine learning engineering at Turkish Technic in İstanbul, and mentoring at a developer festival in London.
Since 2020 I have been in Singapore: a PhD at NTU and A*STAR’s Institute for Infocomm Research on a SINGA scholarship, three months in Dima Damen’s group at the University of Bristol, and, since March 2025, a Research Scientist at SMU.
I am an outsider to every culture in these benchmarks, which is why they were built with people who are not: co-authors, students and annotators from each country. Cultural AI built only by people who visited is the problem, not the fix.
What I think the field should do
- Source data from the region, with the people of the region. Representation is a data collection decision, not a fine-tuning trick.
- Test concepts, not translations. If the concept inventory is Western, translating the language changes nothing about what is measured.
- Demand evidence, not just answers. Grounding or segmentation separates understanding from a lucky guess.
- Report the gap, not the average. A single accuracy number hides exactly the disparity that matters.
- Treat cultural knowledge and cultural perception as different problems. Fixing one does not fix the other.
Common questions
Are vision-language models biased against non-Western cultures?
Yes, and it is measurable. Vision-language models learn from web data, which over-represents Western scenes. On the Seeing Culture benchmark (EMNLP 2025), built across seven Southeast Asian countries, GPT-o3 answers 91% of across-culture questions correctly but never exceeds 32.5 mean IoU when asked to point to the artifact it just reasoned about: accurate in words, wrong in what it looks at.
How do you measure whether a vision-language model actually understands a cultural scene?
Ask it to show its work. Seeing Culture (EMNLP 2025) pairs every culturally grounded question with a segmentation mask, so the model must answer and then ground that answer in the image. Cultural Moment (EMNLP 2026, oral) extends the idea to video and scores naming, recognition and temporal localization as three separate abilities.
What benchmarks exist for cultural understanding in Southeast Asia?
Seeing Culture (EMNLP 2025) covers 138 cultural artifacts from seven Southeast Asian countries; Cultural Moment (EMNLP 2026, oral) covers 306 expert-curated concepts across 624 videos in the same countries. Both are public on Hugging Face under Multimedia-SMU. SEA-VL (ACL 2025) and SEACrowd (EMNLP 2024) add large-scale regional image and language data.
Is one annotator from Southeast Asia enough to build a cultural benchmark?
No. The 14-rater human study in Cultural Moment (EMNLP 2026, oral) shows that the knowledge these tasks require is country-specific rather than regional, so a rater from one Southeast Asian country is not a proxy for the others. Cultural benchmarks need annotators from each culture they represent.
Who is Burak Satar?
I am a Research Scientist at Singapore Management University, working with Prof Chong-Wah Ngo on culturally-aware vision-language models for Southeast Asia. I lead Cultural Moment (EMNLP 2026, oral) and am co-first author of Seeing Culture (EMNLP 2025). I am open to research collaborations, student internships at SMU, and talks on cultural AI evaluation.
Elsewhere
I co-run Turquoise Dot, a meetup for Turkish and Turkic researchers and tech professionals in Singapore, and I keep a list of cultural VLM resources current as the field moves.
If any of this is useful to you, email me or book a 30-minute chat.