Ruobing Xie is a senior researcher of Tencent Hy. He received his BEng degree in 2014 and his master degree in 2017 from the Department of Computer Science and Technology, Tsinghua University. His research interests include Large Language Model, Pre-training, and Recommender System. He has published 100+ papers in top-tier conferences and journals including ACL, ICML, NeurIPS, KDD, AAAI, with 17000+ citations according to Google Scholar. He has received the first prize of the Natural Science Award (Science and Technology) of the Ministry of Education, RecSys 2024 Best Full Paper, RecSys 2023 Best Short Paper, the Paper Award Nominations of TheWebConf 2023 (spotlight), EMNLP 2025 SAC highlight, and CCIR 2024 best paper candidate. He is supported by Scholar of Young Talent Promoting Project of CAST, and is the World's Top 2 % Scientists. He is a member of Chinese Information Processing Society of China, Youth Working Committee & Social Media Processing Committee & Information Retrieval Committee. He serves as AC of ACL, AAAI, WWW, EMNLP, NAACL, etc.
谢若冰，腾讯混元高级研究员，2017年毕业于清华大学自然语言处理与社会人文计算实验室，人工智能高级工程师，主要研究方向为大语言模型、预训练、推荐系统等。他于ACL、ICML、NeurIPS、KDD、AAAI等顶会期刊共发表相关论文100余篇，Google Scholar统计引用超过17000次，h index超过50。他曾获教育部自然科学一等奖，世界互联网领先科技成果奖，RecSys 2024最佳论文，RecSys 2023最佳短文，WWW 2023亮点论文（最佳论文提名），EMNLP 2025 SAC highlight，CCIR 2024最佳论文提名等。他入选了中国科协青年人才托举工程，全球前2%顶尖科学家，西贝尔学者，担任中国中文信息学会青年工作委员会委员、社会媒体处理专委会委员、信息检索专委会通讯委员、ACL/AAAI/WWW/EMNLP/NAACL领域主席等。
Got one paper accepted at NeurIPS about LLM quantization.
Release Hunyuan research blog -- From LR to ELR: A Better Heuristic for Pretraining Dynamics Link.
Got five papers accepted at EMNLP about MoE scaling law, high-dimensional RoPE, quantization, sparse SFT, and RL4RAG.
Attended ICML 2026 at Seoul, South Korea.
Got two papers accepted at TOIS and TKDE about LLM4Rec.
Got four papers accepted at ACL and ICML about MoE routing strategy, multi-teacher distillation, joint SFT&OPD training, and mechanistic interpretability of LLM.
Got four papers accepted at ICLR about more generalized SFT, dynamic data mixing, and RL self-rewarding of LLM.
Got three papers accepted at AAAI and TOIS about hybrid LLM and negative sampling in recommendation.
Received EMNLP 2025 SAC highlight award!
Attended EMNLP 2025 at Suzhou, China.
Got one paper accepted at AAAI about Full/Linear attention hybrid model.
This paper proposes Hybrid Gated Attention (HyGA), which combines three complementary multi-source gating strategies with low-rank decomposition and a learnable attention sink. HyGA reports consistent improvements over standard gated attention in training loss and downstream performance across backbones and compute budgets, reaching the Pareto frontier of effectiveness and efficiency. arXiv
This paper proposes Manifold Power Iteration (MPI), which uses a "Power-then-Retract" update to align each MoE router row with the principal singular direction of its expert, building direct connections between router weights and their corresponding experts' capabilities. MPI provides convergence guarantees and shows gains in pretraining MoE models from 1B to 11B parameters. arXiv
* equal contribution · † corresponding author