Research Interests

My research goal is to build robust, self-improving multimodal agents that perceive, reason, and act reliably across diverse environments. I develop evaluation and post-training methods for vision-language and video models, with a current focus on interactive environments. My prior work spans VLM post-training and knowledge distillation, multimodal reasoning and robustness, and multilingual and multicultural benchmarks.

Research Areas: multimodal reasoning and evaluation, multimodal post-training, interactive environments and world models, embodied AI

Education

Advisor: Alham Fikri Aji

computational neuroscience, tech management

visual question answering; synthetic data generation; reasoning LLM

Publications (* (Co-)first author † Second author)

• Ego2Act: Evaluating Goal-Directed Manipulation in Egocentric Video Generation^*
  video generation, world model, evaluation, multi agent system, embodied agentsPreprint 2026
• LinguDistill: Recovering Linguistic Ability in Vision Language Models via Selective Cross-Modal Distillation^*
  VLM, SFT, post-training, knowledge distillation, architecturePreprint 2026
• Anthropogenic Regional Adaptation in Multimodal Vision-Language Model
  VLM, SFT, post-training, data curation, evaluation, multilingual, multimodalAACL 2026
• Vision Language Models are Confused Tourists^*
  VLM, visual reasoning, robustness, hallucination, interpretabilityCVPR 2026
• A Massive-Scale Multilingual Multi-Cultural Multimodal RAG^†
  RAG, multilingual, multimodal, evaluation, test-time scalingCVPR 2026
• Counting to Four is Still a Chore for VLMs^*
  VLM, numerical reasoning, interpretability, RLCVPRW 2026
• Crosslingual Language Modeling Entropy as End-to-End Learnable Language Representations^*
  LLM, multilingual, representation learning, unsupervised learningEMNLPW 2025
• WorldCuisines: A Massive-Scale Benchmark for Multilingual and Multicultural VQA on Global Cuisines^*
  VQA, multilingual, multicultural, benchmark, evaluation, data curationNAACL 2025 (Best Theme Paper)
• Seeing Culture: A Benchmark for Visual Reasoning and Grounding^†
  visual grounding, evaluation, object detection, multiculturalEMNLP 2025
• Predicting Language Model Performance on Multilingual Tasks via Proxy Models
  multilingual, evaluation, efficient, regression, SFTNAACL 2025
• Towards Efficient and Robust VQA-NLE Data Generation with Large Vision-Language Models^*
  VQA, VLM, NLP, synthetic data generation, visual reasoningCOLING 2025
• SEACrowd: A Multilingual Multimodal Data Hub and Benchmark Suite for Southeast Asian Languages
  multilingual, multicultural, video understanding, speech understanding, evaluationEMNLP 2024

Work Experience

• Discovered up to a 30% performance degradation in frontier VLMs on a multicultural visual grounding & VQA benchmark, resulting in an EMNLP benchmark publication and an analysis-focused preprint.
• Reduced data validation timeline by days for 50K+ spatial annotations with an automated VLM-based evaluation pipeline that automatically flagged poor polygon segmentation via foreground-color ratios.
• Minimized model guessing shortcuts and ensured benchmark robustness by engineering a heuristic-based question-generation engine to synthesize hard adversarial distractors.

• Automated manual enterprise operations by integrating real-time scheduling, invoicing, and geolocation systems serving 3+ scaleup companies.
• Cut system latency by ~20% while maximizing uptime and scalability by architecting a microservices migration from a legacy monolithic MVC infrastructure.

• Scaled command-line autograder software to support 1k+ concurrent users with 99% uptime.
• Contributed to an 8% increase in absolute approved test-case volume within 6 months at major corporation scale.
• Optimized a topic extraction pipeline, resulting in a 2x efficiency gain and 20% classification improvement.
• Led the development of real-time AR and RAG multimodal systems for a high-stakes demo of 20k+ audiences.

Achievement & Awards

Advising & Mentorship

References