Patrick Amadeus Irawan

About Me

I am a PhD researcher at MBZUAI, advised by Alham Fikri Aji. I build adaptive multimodal systems that can acquire new capabilities while preserving what they already know.

Previously, I was a Research Engineer at Singapore Management University under Chong-Wah Ngo, where I worked on multilingual and multimodal learning. During my undergraduate studies at Institut Teknologi Bandung under Ayu Purwarianti, I developed efficient methods for generating explainable visual question-answering data.

My research goal is to develop self-improving multimodal systems that able to continuously learn from feedback and adapt how they act and reason to keep up with an everchanging world.

Research

How can we design multimodal systems that can efficiently perceive evolving world, acquire new skills, and continually improve themselves?

Efficient learning — efficient multimodal learning, VLM post-training, capability preservation
Self-improvement — test-time learning, test-time adaptation, self-refinement, continual learning, self evolving agents
Reasoning and evaluation — multimodal reasoning, multimodal robustness evaluation, multilingual, multicultural, and low-resource evaluation

Publications

My work spans efficient multimodal learning and VLM post-training, inference-time adaptation and self-improving systems, and multilingual, multicultural, and robustness evaluation.

Google Scholar ↗Full publication list →

Experience

PhD Researcher — MBZUAI
Research Engineer — Singapore Management University
Software Engineer — IT Bauschmiede
Data Scientist Intern — Supertype
Software Engineer Intern — Blibli
Software Engineer Intern — Ruangguru

Updates

[Apr. 2026] LinguDistill is out on arXiv, studying how selective cross-modal distillation can recover linguistic ability in VLMs while preserving multimodal competence 🧠
[Feb. 2026] Two papers accepted to CVPR 2026: M4-RAG and Vision Language Models are Confused Tourists 🎉
[Nov. 2025] Our study on culturally-conflicting visual reasoning, Vision Language Models are Confused Tourists, is now on arXiv 🧳
[Oct. 2025] Entropy2Vec got accepted into MRL Workshop @ EMNLP 2025 🌐🇨🇳
[Jul. 2025] Seeing Culture Benchmark is accepted to EMNLP 2025 🇨🇳
[Apr. 2025] WorldCuisines receives Best Theme Paper at NAACL 2025 🎉🌏🍽️
[Mar. 2025] Admitted to the Fall 2025 cohort of the MBZUAI PhD program in NLP 📚
[Jan. 2025] WorldCuisines and ProxyLM are accepted to NAACL 2025 🇺🇸🎖️