Researcher · Tencent Hy Team
Haoran Sun
I am Haoran Sun, a researcher at Tencent Hy RL Team, working on scientific intelligence and large language models. My work focuses on training, alignment, and evaluation for STEM agents and STEM reasoning. I believe that agents will become deeply involved in scientific and engineering workflows—from chip design and aerospace engineering to the search for extraterrestrial civilizations—as we move toward artificial superintelligence (ASI).
What I do
I was a master's student at Tianjin University from 2022 to 2025, working in the TJUNLP Lab under the supervision of Prof. Deyi Xiong. My graduate research focused on reinforcement learning from human feedback (RLHF) for large language models, fine-tuning large language models, multilingual neural machine translation, and end-to-end speech translation.
I am now a researcher at Tencent Hy RL Team, studying the capabilities of large language models in STEM, especially STEM agents and STEM reasoning. My work centers on reinforcement learning (RL), synthetic data generation, and supervised fine-tuning (SFT), with major contributions to capability development for Hy-series models (including Hy3 and Hy4), particularly the capabilities evaluated by Terminal-Bench-Science, Terminal-Bench-4, HLE, SUPERChem, etc.
Hy-series blog posts
Paper work
2025 → 2022* Shared co-first authorship.
First-author and co-first-author work
2025Curiosity-Driven Reinforcement Learning from Human Feedback 2025MA-RLHF: Reinforcement Learning from Human Feedback with Macro Actions 2024Conifer: Improving Complex Constrained Instruction-Following Ability of Large Language Models 2024FuxiTranyu: A Multilingual Large Language Model Trained with Balanced Data 2023Towards a Deep Understanding of Multilingual End-to-End Speech Translation 2022Language Branch Gated Multilingual Neural Machine Translation 2022Multilingual Neural Machine Transliteration with Adaptive Segmentation SchemesOther papers
2025TaP: A Taxonomy-Guided Framework for Automated and Scalable Preference Data Generation 2025Hunyuan-TurboS: Advancing Large Language Models through Mamba-Transformer Synergy and Adaptive Chain-of-Thought 2024Multilingual Large Language Models: A Systematic Survey 2024Regularizing Cross-Attention Learning for End-to-End Speech Translation with ASR and MT Attention Matrices 2023CCSRD: Content-Centric Speech Representation Disentanglement Learning for End-to-End Speech Translation 2023CKDST: Comprehensively and Effectively Distill Knowledge from Machine Translation to End-to-End Speech Translation