About
The three-minute version
Show a computer a photo of a wedding in Ohio and it will tell you about 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 confident text looks like, not what it is looking at.
This happens because the models are trained on whatever was on the internet, and the internet is not the world. Roughly speaking, the machine has seen a million Ohio weddings and a handful of Barong dances.
My work is to prove that precisely enough that it can be fixed. We build tests. Our first benchmark, Seeing Culture (EMNLP 2025), does something most tests do not: it makes the model show its work. First answer the question about the artifact; then point to that artifact in the picture. A model that names the right thing while circling the wrong part of the image did not know the answer; it guessed well. Ordinary accuracy scores cannot tell those two apart. Ours can.
The result is not subtle. On our benchmark GPT-o3 answers 91% of across-culture questions correctly, and 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 same principle into video, where culture lives in time as well as in pixels. A model must name a Southeast Asian concept from its symbolic description, recognize it among unlabeled video moments, and then locate its sub-events inside a different video of the same concept. We score the three abilities separately, and they do not compose: even the strongest closed-source models clear all three stages for fewer than 30% of concepts, and a 14-rater human study shows the knowledge required is country-specific, not regional. You can try all three stages yourself on the project page.
How I got here
I grew up in Bursa, in northwest Türkiye, and studied electronics engineering at Uludağ University, with exchange semesters in Siena and Naples. After that a software engineering traineeship in Valencia, a master’s back in Bursa on vehicle detection, a year as a machine learning engineer at Turkish Technic in İstanbul, and mentoring at a developer festival in London.
Since 2020 I have been in Singapore: five years at NTU and A*STAR’s Institute for Infocomm Research on a SINGA scholarship, with three months at the University of Bristol in Dima Damen’s group, and since March 2025 a Research Scientist at SMU.
Bursa → Siena and Naples → Valencia → Bursa → İstanbul → London → Singapore → Bristol → Singapore.
I am Turkish, I live in Singapore, and I work on Southeast Asia. I am an outsider to every culture in the Seeing Culture and Cultural Moment benchmarks, which is why they were built with people who are not: the co-authors, students and annotators from the countries in them. Cultural AI built only by people who visited is the problem, not the fix.
What I think the field should do
This is the checklist that I would argue for:
- 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 lucky guessing.
- 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 are trained on web data, and the web over-represents Western scenes, so the models have seen far more Western weddings than Balinese Barong dances. On the Seeing Culture benchmark (EMNLP 2025), a two-stage test built across seven Southeast Asia 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. The gap is systematic, not noise.
How do you measure whether a vision-language model actually understands a cultural scene?
Ask it to show its work. Accuracy alone cannot separate understanding from a confident guess, because a model can name the right festival while looking at the wrong part of the image. Seeing Culture (EMNLP 2025) pairs every culturally grounded question with a segmentation mask, so the model must answer and then ground its answer in the pixels. Cultural Moment (EMNLP 2026, oral) extends the same 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 1,065 images of 138 cultural artifacts from seven Southeast Asia countries with 3,178 questions. Cultural Moment (EMNLP 2026, oral) covers 306 expert-curated concepts across 624 videos in the same seven countries. SEA-VL (ACL 2025) contributes 1.28M culturally relevant images, and SEACrowd (EMNLP 2024) consolidates corpora across roughly 1,000 Southeast Asia languages. Seeing Culture and Cultural Moment are both public: the data is on Hugging Face under Multimedia-SMU and each has its own project site.
Can a vision-language model be accurate and still not understand?
Yes. This is the single clearest result from the Seeing Culture benchmark: a model can answer 91% of questions correctly while its visual grounding stays under 32.5 mean IoU, which means it names the right cultural artifact and then highlights the wrong region of the image. High accuracy with low grounding is the signature of a model that has learned what a confident answer sounds like rather than what it is looking at.
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 Asia country is not a proxy for the others. Cultural benchmarks need annotators from each culture represented, which is why Seeing Culture and Cultural Moment were built with co-authors, students and annotators from the countries in them.
Who works on culturally-aware vision-language models in Singapore?
Burak Satar is a Research Scientist at Singapore Management University who works on culturally-aware vision-language models with a focus on Southeast Asia, alongside Prof Chong-Wah Ngo. He is the lead author of Cultural Moment (EMNLP 2026, oral) and co-first author of Seeing Culture (EMNLP 2025), and he is open to research collaborations, student internships at SMU, and invited talks on cultural AI evaluation.
Elsewhere
I 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.