Batch 2027 Projects

These are the research topics for the SEACrowd 2027 Apprentice Program (February – June 2027). Each topic lists its mentors and the time zones of the mentor group, so you can check that the team's schedule fits your week. Apply as a mentee between Nov 20 and Dec 13, 2026.

Projects

Burak Satar Hitesh Laxmichand Patel

Do VLMs Follow a Tradition as It Unfolds?

Burak Satar · Singapore Management University Hitesh Laxmichand Patel · Oracle

Do VLMs know, see, and locate where a Southeast Asian tradition is in a video? A diagnostic question set on top of the Cultural Moment Benchmark.

Multimodal Culture Evaluation & benchmarks UTC-5 / UTC+8
Samuel Cahyawijaya Alham Fikri Aji Jonibek Mansurov

Multilingual Jailbreaking in Agentic AI

Samuel Cahyawijaya · Cohere Alham Fikri Aji · MBZUAI Jonibek Mansurov · MBZUAI

Jailbreaking beyond the user prompt: injecting harmful instructions through tool outputs, documents, and skills in multilingual agentic AI.

Safety & security Agents Multilingual NLP UTC+0 / UTC-4
Fajri Koto Muhammad Dehan Al-Kautsar

Predicting LLM Performance for Cost-Aware Model Selection

Fajri Koto · MBZUAI Muhammad Dehan Al-Kautsar · MBZUAI

Predicting quality, latency, and cost across a model-by-task grid to route each request to the cheapest LLM that meets a quality bar.

Efficiency Evaluation & benchmarks Multilingual NLP UTC+4
Derry Wijaya Burak Satar David Anugraha

What is your language model thinking right before it says the wrong thing?

Derry Wijaya · Monash University Indonesia Burak Satar · Singapore Management University David Anugraha · Stanford University

Finding and intervening on the internal circuits behind safety, honorific, and honesty behaviors with probes, steering, and activation patching.

Mechanistic interpretability Safety & security Multilingual NLP UTC-8 / UTC+7 / UTC+8
Samuel Cahyawijaya Amit Agarwal

What Really Matters? Multilinguality in the Era of Agentic AI

Samuel Cahyawijaya · Cohere Amit Agarwal · Oracle

Is multilinguality best built into the agent, or added through a harness? Comparing both paradigms, and hybrids, in noisy agentic environments.

Agents Multilingual NLP Evaluation & benchmarks UTC-8 / UTC+0