Yang Deng

📢 #SSNLP2026 – Singapore Symposium on Natural Language Processing 🇸🇬 SSNLP is returning on Jan 21, 2026 hosted at SUTD! As the annual gathering of Singapore’s NLP community, SSNLP brings together students, faculty, and industry researchers to share ongoing work, exchange ideas, and spark new collaborations. This year, we are especially excited to welcome leading researchers in the field as this year’s keynote speakers: Yuki Arase @Yuki_arase (Tokyo Institute of Technology), Tanmoy Chakraborty @Tanmoy_Chak (IIT, Delhi), Hung-yi Lee @HungyiLee2 (National Taiwan University), and Derek Wong @Keredwong (University of Macau), offering valuable insights into the latest advances in NLP. Free registration is available (thanks to our generous sponsors!) — simply scan the QR code on the poster. Registration is open until January 4, 2026. Held right before AAAI, SSNLP 2026 warmly welcomes everyone to join us in Singapore! 🔗 More information & registration: ssnlp2026.github.io/

🙋🏻‍♂️ I'm actively recruiting for multiple fully-funded positions in my group at SUTD! Topic: NLP / LLM / Multimodal Openings: 1 x Postdoc: start date is flexible 1 x PhD: earliest batch is Spring 2026 2-3 x Visiting students: fully funded for 6-12 months. Open to Bachelor / Master / PhD students, priority will be given to students with relevant research / publication experience. More details: Potential projects: isakzhang.github.io/index.ht… Expectations of different roles: isakzhang.github.io/getinvol… Feel free to drop me an email if you are interested! If you are attending #SSNLP2025 or #ICLR2025 next week, it would be great to chat in person :)

SeaLLMs-Audio: Large Audio-Language Models for Southeast Asia 🎧 📢 Excited to share SeaLLMs-Audio, the multimodal (audio) extension of the SeaLLMs family. Following the release of SeaLLMs-v3 last year, we've focused on expanding to audio/speech capabilities, addressing the unique challenges of developing multimodal multilingual models for Southeast Asian languages. ✨ Key features: 1️⃣Multi-lingual: Covers 5 languages (for now): 🇮🇩Indonesian, 🇹🇭Thai, 🇻🇳Vietnamese 🇬🇧English, 🇨🇳Chinese. 2️⃣Multi-modal: Works with audio-only, text-only and audio-text inputs. 3️⃣Multi-task: Handles various tasks including audio analysis (e.g., audio captioning, ASR, S2TT, speech emotion recognition, and speech summarizatio) and general voice chat interactions. Since this is a relatively unexplored space, we also developed a new SeaBench-Audio benchmark for eval. Therefore, we are super welcome for your feedback as we continue improving these models! Try it yourself and join us in advancing LLMs for SEA👇: Homepage: damo-nlp-sg.github.io/SeaLLM… Model: huggingface.co/SeaLLMs/SeaLL… Demo: huggingface.co/spaces/SeaLLM… Github: github.com/DAMO-NLP-SG/SeaLL…

Thrilled to share Babel - a new family of open-source multilingual LLMs covering 90% of the world's population! I've talked to many people about Babel over the past few years, and many know what this project means to me. So let me share my personal thoughts instead of writing a short and eye-catching promo post, since you can find all the information from the project page (babel-llm.github.io/babel-ll…). Multilingual NLP research has always faced a chicken-and-egg problem: without good multilingual models, it's difficult to do more research on them, then this makes it even harder to develop better ones. Bloom used to be a cure for this, but was soon left behind by the surprising capabilities of more modern LLMs from 2023. After releasing the first version of SeaLLMs, I was quite encouraged (by the positive feedback from the SEA region) to expand the language coverage -- or develop a more “general multilingual” LLM in other words. The idea of project Babel emerged at that time. But it was put on hold for some reasons. Then things began to shift in mid-2024, as more LLMs started focusing on multilingual capability like Qwen-2, Gemma-2, and Llama-3.1. Though not specifically claiming to be multilingual LLMs, they indeed supported more languages. However, we soon found that the supported languages were heavily biased - languages spoken in more well-developed communities/countries were served first (e.g., Europe), while languages with lots of speakers were neglected (e.g., many many other regions). So when people (and my ex-boss, and my ex-ex-boss) asked me why I always wanted to do this, I told them there's really a need for it. However, due to various reasons, this model was delayed for a while. But I'm glad to see it finally here. We humbly release the Babel models and deeply acknowledge there's still much room for improvement. We're quite confident in the base version of two sizes, but I have to admit the chat version is not fully satisfactory to us (given the lack of multilingual alignment data). We extremely welcome all kinds of feedback and community contributions to further improve it. Lastly, almost all my interns at Alibaba were (hopefully happily) involved in this project. Especially the leading author Yiran, who poured tremendous energy into this project. All credits go to them.