Introducing WPO: Enhancing RLHF with Weighted Preference Optimization 🌟 Our new preference optimization method reweights preference data to simulate on-policy preference optimization using off-policy data, combining efficiency with high performance. ✅ up to 5.6% better than

🚨 Excited to share our new preprint with Shujian Zhang, Wenxuan Zhou, Marzyeh, and Sanqiang Zhao. 🚀 TLDR: We propose SFTMix, a novel recipe that elevates language model instruction tuning without relying on expensive, well-curated datasets. arxiv.org/abs/2410.05248

How can LLM architecture recognize Instruction Hierarchy? 🚀 Excited to share our latest work on Instructional Segment Embedding (ISE)! A technique embeds Instruction Hierarchy directly into LLM architecture, significantly boosting LLM safety. 🧵[1/n]

📝 For more details and experiments, please check our arxiv: arxiv.org/abs/2410.09102 Thanks to all of my amazing collaborators 🙌 Wenxuan Zhou Shujian Zhang KaiqiangSong Silei Xu xcc Prof. Prateek Mittal and more colleagues from @zoom. 🧵[7/n]

I am hiring a Student Researcher at Google DeepMind for 2025! 👩🔬 Interested in improving multi-turn optimization and reasoning capabilities of LLMs? 🧑‍🎓 Currently studying for a Bachelor's/Master's/PhD? 🧑‍💻 Have solid engineering and research skills? 🌟We want to hear from you!

Our Gemini 2.5 tech report is out on Arxiv (arxiv.org/pdf/2507.06261)! Nice work from the team! 🌟