Research Scientist, Gemini Safety

About the role

Snapshot

Artificial Intelligence represents one of humanity's most promising inventions. At Google DeepMind, we unite scientists, engineers, and machine learning specialists to advance artificial intelligence. Our mission uses technology for public benefit and discovery, guided by strict safety and ethical standards. The Gemini Safety team owns the safety and fairness of Google DeepMind models. The Research Scientist or Research Engineer applies advanced data and algorithmic methods to user-facing systems. The environment moves quickly and depends on close cooperation. The team culture emphasizes mutual support, commitment, and collaboration.

About Us

Artificial Intelligence represents one of humanity's most promising inventions. At Google DeepMind, we unite scientists, engineers, and machine learning specialists to advance artificial intelligence. Our mission uses technology for public benefit and discovery, guided by strict safety and ethical standards. We work with partners on vital challenges, placing safety and ethics above all else.

The Role

We seek a versatile Research Scientist comfortable defining approaches and handling technical work for Gemini Safety. The team drives the development of foundational technology used across Gemini App, Cloud API, and Search. Key duties center on post-training and instruction tuning for large language models. The focus covers text-to-text, image and video-to-text, and audio-to-text modalities, including agentic abilities. The role advances adversarial robustness, especially for high-risk misuse scenarios. The scientist will design evaluation protocols that detect model behavior gaps related to safety and fairness. They will plan experiments to close those gaps or create new capabilities. The role also pushes innovation in large-scale Supervised Fine Tuning and Reinforcement Learning.

About You

To succeed as a Research Scientist on the Gemini Safety team, you must hold a PhD in Computer Science or a related field. Equivalent hands-on experience may substitute for the degree. You need significant post-training experience with large language models in real systems. Strong written communication skills are essential for documenting methods and results. You must handle complex, ambiguous issues that impact real users. You will work closely with cross-functional teams under fast-moving timelines.

Additional experience strengthens your application. This includes reward modeling and reinforcement learning for language model instruction tuning. Background in long-range reasoning and agentic task solving is valuable. Published work at top venues such as NeurIPS, ICLR, ICML, EMNLP, AAAI, or UAI is an advantage. A record of moving research concepts to shipped products matters. Experience in applied research within high-stakes domains helps. Familiarity with the JAX ecosystem is also beneficial.

Google DeepMind values diverse backgrounds, knowledge, and perspectives. We harness these qualities to create significant impact. We offer equal opportunity regardless of sex, race, religion or belief, ethnic or national origin, disability, age, citizenship, marital status, domestic partnership, civil partnership, sexual orientation, gender identity, pregnancy, or related conditions. If you require accommodation for a disability or additional need, please inform us.

Key Facts

Location: USA

Engagement: Full-time employment.

Base

compensation: $191,231.00 annually.

Signing bonus: $20,000.00 one-time payment.

Total expected

compensation: $328,800.00 over four years.

Vesting schedule spans four years.

What You'll Do

You will design evaluation suites that measure safety and fairness for text, image, video, and audio model outputs. You will gather and structure data to expose failure modes in large models. This work will drive targeted post-training experiments. You will strengthen adversarial robustness for Gemini models, focusing on high-stakes misuse scenarios. You will create and maintain protocol definitions that clarify safe and helpful behavior for Gemini products. You will plan and run experimental campaigns that close identified gaps or enable new model capabilities. You will guide how Supervised Fine Tuning and Reinforcement Learning methods scale across Gemini App, Cloud API, and Search. You will partner with product teams and engineers to turn research findings into reliable production features. You will define metrics and dashboards that make model behavior gaps visible to researchers and stakeholders.

Requirements

You hold a PhD in Computer Science or a related field, or matching practical experience through your work history. You have significant hands-on experience post-training large language models in production settings. You demonstrate strong written communication for documenting methods and results. You show comfort working with complex, ambiguous problems that affect real users. You commit to deep collaboration with cross-functional teams under fast-moving timelines.

Nice to Have

Experience with reward modeling and reinforcement learning for language model instruction tuning. Background in long-range reasoning and agentic task solving. Published work at major venues such as NeurIPS, ICLR, ICML, EMNLP, AAAI, or UAI. A track record of taking research concepts into shipped products. Experience collaborating on applied research in high-stakes domains. Familiarity with the JAX ecosystem.

Skills and Tools

JAX, Gemini, TensorFlow, PyTorch.