arXiv2026
Mengsay Loem
Engineer at Preferred Networks · formerly Researcher at Sansan · Tokyo, Japan
I build and study language and vision-language models — with a focus on how we evaluate them, and how they hold up in low-resource settings.
News
- Released Samdei Type, open models that turn romanized Khmer into Khmer script.
- Guest speech at ACET 2026 in Phnom Penh, on judge-free evaluation of generation in low-resource languages.
- New preprint, Khmer PFGen-Bench, a judge-free benchmark for open-ended generation in Khmer.
- Started Notes, a public notebook of paper-reading memos and research thoughts.
- Joined Preferred Networks as an engineer.
Selected publications
All 26 →JNLP2025
Journal of Natural Language Processing, Vol. 32, Issue 2, pp. 480–496
★ Best Journal Paper Awardpaper
COLM2024
Continual Pre-Training for Cross-Lingual LLM Adaptation: Enhancing Japanese Language Capabilities
Conference on Language Modeling (COLM), Univ. of Pennsylvania, USA
ECAI2024
27th European Conference on Artificial Intelligence (ECAI-2024), Santiago de Compostela, Spain
BEA2023
18th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2023), pp. 205–219, Toronto, Canada
Projects
All projects →
Samdei Type
Open models that turn romanized Khmer, the Latin spelling many Cambodians use in chat, into Khmer script. Small enough for a phone keyboard, and released with all their training data and a new benchmark of everyday messages.
Recent notes
All notes →A short note on repetition mismatch in pre-training data-mixture experiments: why the best mixture at a small budget can change at the target budget, and how downsampling to preserve repetitions makes small experiments predictive again.
A short note on romanized multilingual pretraining, why it improves cross-script transfer, and why the mechanism may be more complicated than lexical overlap alone.
A short note on translationese in multilingual LLM generation, and a question that may be more useful than asking whether a model "thinks in English."