Research Scientist, Gemini Agent Safety Post Training, DeepMind

Google•Mountain View, CA

3d•$207,000 - $300,000

About The Position

The Gemini Safety team is accountable for the standards of Google DeepMind’s flagship releases. In this role, you will develop algorithmic solutions to advance user-facing architectures and bring Large Language Model post-training expertise to ensure outputs are aligned and production-ready. You will take end-to-end ownership as a builder and shipper, working with an interdisciplinary team focused on advancing AI, solving complex global challenges, and accelerating product innovation. Safety and ethics are central priorities.

Requirements

PhD in Computer Science, a related field, or equivalent practical experience.
3 years of experience initiating, scoping, and executing research agendas or leading academic projects.
One or more peer-reviewed scientific publication submissions (e.g., NeurIPS, ICML, ICLR, CVPR, ACL, CoRL) or public preprint repositories (e.g., arXiv).

Nice To Haves

Experience in AI safety and model alignment.
Experience developing AI agents, safety-critical systems, or Machine Learning Infrastructure.
Expertise in reward modeling, Reinforcement Learning (RL) for LLMs, instruction tuning, and long-range RL.
Ability to drive research concepts to product realization, with a strong track record of engineering abilities and experimental work.
Track record of publications at AI/ML venues (e.g., NeurIPS, ICLR, ICML, EMNLP, AAAI, UAI).

Responsibilities

Design, implement, and maintain high-quality protocols to identify safety gaps in model behavior across Gemini and GenAI models, and feed insights back into post-training to improve in-model and out-of-model protection.
Explore, adapt, and apply modeling and alignment techniques to solve real-world safety and alignment issues under tight launch timelines.
Drive innovation in model optimization, advancing the application of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) techniques at massive scale.
Partner with post-training, evaluation, and product teams to align incentives, close safety gaps, and influence safety decisions without direct authority.