Publications

Sparse Autoencoders Can Capture Language-Specific Concepts Across Diverse Languages

Published in , 2025

This paper is about utilizing sparse autoencoder to find language-specific concepts that could be important to multilingual capability. We also introduced new interpretability method called SAE-LAPE.

Download here

Are Knowledge and Reference in Multilingual Language Models Cross-Lingually Consistent?

Published in , 2025

In this paper we analyzed about the multilingual knowledge factual consistency and relate it with the training.

Download here

Crowdsource, Crawl, or Generate? Creating SEA-VL, a Multicultural Vision-Language Dataset for Southeast Asia

Published in ACL 2025, 2025

First south-east asian multimodal VQA cultural benchmark dataset

Download here

Published in , 1900