Research and selected work
My work spans navigation, Southeast Asian datasets, and satellite image analysis. Across these settings, I examine how models support decisions and where their assumptions fail. The projects below show my contributions and the questions I want to pursue next.
Past work and research direction
Filled shapes link to past papers and projects. Dashed shapes show open questions and link to their starting points. Hover over a shape or focus it to read more.
Navigation Vision-Language Models Multilingual NLP Benchmarks; Datasets Cultural; Regional AI Safety; Security Monitoring Remote Sensing Semantic Segmentation Urban Policy ENCP SEA-VL CommonLID Anthropogenic
Regional Adaptation ProCANet Mining
Footprints Retention Ponds Plastic Bag Ban Cross-modal
Abstention Goal Hijack Probe Fraud Detection Routes after
asking for help ENCP gives an agent a set of suggested actions at each step. Its guarantee limits the chance of missing a correct action anywhere along the route. This guarantee requires routes used to set the cut-off and future routes to be exchangeable. Reordering those routes must leave the probability of the whole collection unchanged. However, asking for help can change which routes the agent follows. I want to test whether the guarantee still holds after that change. Withholding answers
as error costs change The same error can carry different costs in different settings. I want rules that adjust when a model answers or asks for review. Those rules should also work when the model was trained with different error costs. Reviewing model signals
when inputs change A monitor reads values inside a model to flag answers that may need review. These values, called internal model signals, can change with the language, inputs, or model. Therefore, I want to test when the monitor’s own judgement needs review.
- Past paper or project
- Topic
- Research direction
Selected work
In navigation, I developed a method for choosing which actions an agent should consider. Alongside this work, I contributed regional data systems and designed a model for flood mapping.
ENCP
Keeping the correct action available throughout a route
At each step, a navigation agent chooses from a set of suggested actions. I studied how to keep the correct action available throughout the route.
- My contribution
- I developed ENCP, which uses completed routes to set the cut-off for including an action. Moreover, I led the paper.
- Evidence
- Tests covered 4 navigation policies, 3 scores, and 2 benchmarks. They met the reported targets for including correct actions at individual steps.
- Scope
- The guarantee requires routes used to set the cut-off and future routes to be exchangeable. Reordering those routes must leave the probability of the whole collection unchanged.
SEA-VL
Images and language grounded in Southeast Asia
Models need tests that reflect the languages and cultures of their users. SEA-VL, a dataset of Southeast Asian images and text, provides material for those tests.
- My contribution
- I built systems to collect regional data and check benchmark quality.
- Evidence
- The dataset contains 1.28 million images across 11 regional languages.
- Scope
- The dataset supports regional evaluation. A reliability claim still needs tests for the model and task using it.
ProCANet
Flood maps from colour and near-infrared images
Colour images can confuse flood water with shadows and dark soil. Near-infrared light lies just beyond the visible colour range. Images using that light help distinguish water, so I studied how to combine both inputs.
- My contribution
- I designed ProCANet to combine features from both inputs, then trained and evaluated it.
- Evidence
- Predicted and reference flood regions reached an overlap score of 0.815 on Sen1Floods11.
- Scope
- I also tested images from a different satellite without retraining. The reference labels came from a water index, which uses reflected light to identify water.
Publication archive
2026
ENCP: Episode-Normalized Conformal Prediction for Vision-and-Language Navigation
ENCP sets thresholds using whole routes and gives guarantees on including correct actions throughout a route. Tests cover 4 navigation policies on R2R and REVERIE. However, the guarantee requires calibration and test routes whose probabilities stay unchanged when reordered.
Details
Authors Vicky Feliren, A. Taufiq Asyhari, Muhamad Risqi U. Saputra
Abstract Vision-and-language navigation requires uncertainty estimates for dependent, variable-length routes. ENCP rescales a nonconformity score using a policy’s residual confidence and calibrates the maximum score per episode. Under exchangeable calibration and evaluation episodes, the method covers the correct action throughout a route with probability at least the chosen coverage level. It allows dependence between steps. Evaluations across 4 navigation policies and 3 scores on R2R and REVERIE met the reported empirical step-coverage targets, including on unseen buildings. That empirical result does not extend the formal guarantee to a changed distribution of buildings.
DOI 10.48550/arXiv.2609.17499
CommonLID: Re-evaluating State-of-the-Art Language Identification Performance on Web Data
A community-built, human-annotated web benchmark covering 109 languages.
Details
Authors Pedro Ortiz Suarez, Laurie Burchell, Catherine Arnett, Rafael Mosquera-Gómez, Sara Hincapie-Monsalve, Thom Vaughan, Damian Stewart, Malte Ostendorff, Idris Abdulmumin, Vukosi Marivate, Shamsuddeen Hassan Muhammad, Atnafu Lambebo Tonja, Hend Al-Khalifa, Nadia Ghezaiel Hammouda, Verrah Otiende, Tack Hwa Wong, Jakhongir Saydaliev, Melika Nobakhtian, Muhammad Ravi Shulthan Habibi, Chalamalasetti Kranti, Carol Muchemi, Khang Nguyen, Faisal Muhammad Adam, Luis Frentzen Salim, Reem Alqifari, Cynthia Amol, Joseph Marvin Imperial, Ilker Kesen, Ahmad Mustafid, Pavel Stepachev, Leshem Choshen, David Anugraha, Hamada Nayel, Seid Muhie Yimam, Vallerie Alexandra Putra, My Chiffon Nguyen, Azmine Toushik Wasi, Gouthami Vadithya, Rob van der Goot, Lanwenn ar C'horr, Karan Dua, Andrew Yates, Mithil Bangera, Yeshil Bangera, Hitesh Laxmichand Patel, Shu Okabe, Fenal Ashokbhai Ilasariya, Dmitry Gaynullin, Genta Indra Winata, Yiyuan Li, Juan Pablo Martínez, Amit Agarwal, Ikhlasul Akmal Hanif, Raia Abu Ahmad, Esther Adenuga, Filbert Aurelian Tjiaranata, Weerayut Buaphet, Michael Anugraha, Sowmya Vajjala, Benjamin Rice, Azril Hafizi Amirudin, Jesujoba O. Alabi, Srikant Panda, Yassine Toughrai, Bruhan Kyomuhendo, Daniel Ruffinelli, Akshata A, Manuel Goulão, Ej Zhou, Ingrid Gabriela Franco Ramirez, Cristina Aggazzotti, Konstantin Dobler, Jun Kevin, Quentin Pagès, Nicholas Andrews, Nuhu Ibrahim, Mattes Ruckdeschel, Amr Keleg, Mike Zhang, Casper Muziri, Saron Samuel, Sotaro Takeshita, Kun Kerdthaisong, Luca Foppiano, Rasul Dent, Tommaso Green, Ahmad Mustapha Wali, Kamohelo Makaaka, Vicky Feliren, Inshirah Idris, Hande Celikkanat, Abdulhamid Abubakar, Jean Maillard, Benoît Sagot, Thibault Clérice, Kenton Murray, Sarah Luger
Abstract CommonLID is a community-built, human-annotated benchmark for identifying 109 languages in web data.
DOI 10.18653/v1/2026.acl-long.1527
Anthropogenic Regional Adaptation in Multimodal Vision-Language Model
GG-EZ improved cultural relevance for Southeast Asia by 5–15% while retaining more than 98% of global benchmark performance.
Details
Authors Samuel Cahyawijaya, Peerat Limkonchotiwat, Tack Hwa Wong, Hitesh Laxmichand Patel, Amit Agarwal, Manuel Antonio Rufino, Carlos Rafael Catalan, Muhammad Reza Qorib, Vicky Feliren, Holy Lovenia, Aye Hninn Khine, Frederikus Hudi, David Anugraha, Alham Fikri Aji, Romrawin Chumpu, Viet-Thanh Pham, Minghan Wang, Mohamed Fazli Mohamed Imam, Ruochen Zhang, Joseph Marvin Imperial, Khumaisa Nur'aini, Do Xuan Long, Musa Izzanardi Wijanarko, Joel Ruben Antony Moniz, Patrick Amadeus Irawan, Hanif Muhammad Zhafran, Isaiah Flores, Salsabila Zahirah Pranida, Jun Kevin, Jostin Jerico Rosal, Patricia Nicole Monderin, Kun Kerdthaisong, Ahmad Mustafid, My Chiffon Nguyen, Natchapon Jongwiriyanurak, Siva Worajitwannakul, Haochen Li, Adrian Xuan Wei Lim, Bin Wang, Muhammad Ravi Shulthan Habibi, Lynnette Hui Xian Ng, Mithil Bangera, Yeshil Bangera, Priyaranjan Pattnayak, Dun Li Chan, Sherissa Caren Djuniwar, Cho Chan Myei Oo, Hee Ming Shan
Abstract Anthropogenic Regional Adaptation improves a model’s relevance to a specific region while preserving its ability to generalise globally.
DOI 10.48550/arXiv.2604.11490
2025
Multi-modal deep learning approaches to semantic segmentation of mining footprints with multispectral satellite imagery
Mapped mining areas at 37 sites worldwide using deep learning and multispectral satellite images.
Details
Authors Muhamad Risqi U. Saputra, Irfan Dwiki Bhaswara, Bahrul Ilmi Nasution, Michelle Ang Li Ern, Nur Laily Romadhotul Husna, Tahjudil Witra, Vicky Feliren, John R. Owen, Deanna Kemp, Alex M. Lechner
Abstract Earlier remote sensing studies often covered one mine, one type of mining feature, or only the outer boundary of several mines. We studied 37 mining sites worldwide, covering commodities from gold to coal. We built a multispectral dataset with refined boundaries, shapes, and class labels, then trained models to separate mining and non-mining land. We tested U-Net, DeepLabV3+, FPN, SegFormer, and the IBM-NASA Prithvi model across several band combinations and transfer-learning setups. An ImageNet-pretrained FPN with a DenseNet-121 backbone performed best when trained on RGB and near-infrared bands. We also assessed performance across climate groups and mining commodities.
DOI 10.1016/j.rse.2024.114584
Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia
A Southeast Asian vision-language benchmark with 1.28 million culturally relevant images across 11 regional languages.
Details
Authors Samuel Cahyawijaya, Holy Lovenia, Joel Ruben Antony Moniz, Tack Hwa Wong, Mohammad Rifqi Farhansyah, Thant Thiri Maung, Frederikus Hudi, David Anugraha, Muhammad Ravi Shulthan Habibi, Muhammad Reza Qorib, Amit Agarwal, Joseph Marvin Imperial, Hitesh Laxmichand Patel, Vicky Feliren, Bahrul Ilmi Nasution, Manuel Antonio Rufino, Genta Indra Winata, Rian Adam Rajagede, Carlos Rafael Catalan, Mohamed Fazli Mohamed Imam, Priyaranjan Pattnayak, Salsabila Zahirah Pranida, Kevin Pratama, Yeshil Bangera, Adisai Na-Thalang, Patricia Nicole Monderin, Yueqi Song, Christian Simon, Lynnette Hui Xian Ng, Richardy Lobo Sapan, Taki Hasan Rafi, Bin Wang, Supryadi, Kanyakorn Veerakanjana, Piyalitt Ittichaiwong, Matthew Theodore Roque, Karissa Vincentio, Takdanai Kreangphet, Phakphum Artkaew, Kadek Hendrawan Palgunadi, Yanzhi Yu, Rochana Prih Hastuti, William Nixon, Mithil Bangera, Adrian Xuan Wei Lim, Aye Hninn Khine, Hanif Muhammad Zhafran, Teddy Ferdinan, Audra Aurora Izzani, Ayushman Singh, Evan Evan, Jauza Akbar Krito, Michael Anugraha, Fenal Ashokbhai Ilasariya, Haochen Li, John Amadeo Daniswara, Filbert Aurelian Tjiaranata, Eryawan Presma Yulianrifat, Can Udomcharoenchaikit, Fadil Risdian Ansori, Mahardika Krisna Ihsani, Giang Nguyen, Anab Maulana Barik, Dan John Velasco, Rifo Ahmad Genadi, Saptarshi Saha, Chengwei Wei, Isaiah Edri W. Flores, Kenneth Chen Ko Han, Anjela Gail D. Santos, Wan Shen Lim, Kaung Si Phyo, Tim Santos, Meisyarah Dwiastuti, Jiayun Luo, Jan Christian Blaise Cruz, Ming Shan Hee, Ikhlasul Akmal Hanif, M. Alif Al Hakim, Muhammad Rizky Sya'ban, Kun Kerdthaisong, Lester James Validad Miranda, Fajri Koto, Tirana Noor Fatyanosa, Alham Fikri Aji, Jostin Jerico Rosal, Jun Kevin, Robert Wijaya, Onno P. Kampman, Ruochen Zhang, Börje F. Karlsson, Peerat Limkonchotiwat
Abstract Southeast Asia remains underrepresented in vision-language research despite its linguistic and cultural diversity. SEA-VL addresses this gap through open, high-quality datasets grounded in regional languages and cultures.
DOI 10.18653/v1/2025.acl-long.916
Progressive Cross-Attention Network for Flood Segmentation Using Multispectral Satellite Imagery
ProCANet reached a state-of-the-art IoU score of 0.815 on the Sen1Floods11 flood benchmark.
Details
Authors Vicky Feliren, Fithrothul Khikmah, Irfan Dwiki Bhaswara, Bahrul Ilmi Nasution, Alex M. Lechner, Muhamad Risqi U. Saputra
Abstract We present ProCANet, which applies self-attention and cross-attention in stages to combine multispectral features for flood segmentation.
DOI 10.1109/LGRS.2024.3495974
Enhancing urban resilience through integrated flood policy and planning
A mixed-method study of how retention ponds reduce urban flood risk.
Details
Authors Eka Permanasari, Marco Wijaya, Vicky Feliren, Fithrotul Khikmah, Alex M. Lechner, Altaf Virani, Muhamad Risqi U. Saputra
Abstract This study examines how retention ponds in South Bandung can support long-term urban flood management.
DOI 10.2166/aqua.2025.292
2021
The Effect of Plastic Bag Ban Policy Towards Waste Complaints in Jakarta
Measured the policy’s effect on waste complaints submitted through JAKI and Qlue.
Details
Authors Vicky Feliren, Yudhistira Nugraha, Bahrul Ilmi Nasution, Clarissa Febria Finola, J.I. Kanggrawan, A. L. Suherman
Abstract Jakarta introduced a plastic bag ban on 1 July 2020. This study measures its effect on waste complaints.
DOI 10.1109/ICISS53185.2021.9533236