Research Scientist/Engineer, GenMedia Safety Post-Training TLM, DeepMind

Google•Mountain View, CA

1d•$262,000 - $364,000

About The Position

Google DeepMind is a pioneering AI lab with interdisciplinary teams advancing AI to address complex global challenges and accelerate product innovation. The organization uses its technologies for public benefit and scientific discovery, with safety and ethics as priorities. This role leads research and engineering on safety post-training for multimodal foundation models, shaping model behavior and advancing approaches that generalize across media types. It also involves managing a team of 10 or more researchers and engineers and fostering scientific excellence and collaboration. The listed US compensation is $262,000–$364,000, plus a 25% bonus target, equity, and benefits. Individual pay depends on job-related skills, experience, and relevant education or training.

Requirements

PhD or equivalent practical research experience.
5 years of experience in a technical leadership role setting direction and guiding project workstreams.
5 years of experience in people management, supervision, or team leadership.
5 years of experience in machine learning, specifically with large language models (LLMs), multimodal foundation models, or reinforcement learning (RL), along with a safety and alignment specialization such as model alignment, AI safety, or post-training techniques including Supervised Fine-Tuning and reward modeling.
One or more peer-reviewed scientific publications (e.g., NeurIPS, ICML, ICLR, CVPR, ACL, CoRL) or public preprints (arXiv).

Nice To Haves

6 years of demonstrated experience initiating, scoping, and executing independent research agendas or leading collaborative academic projects.
Experience in core strategic focus areas: LLMs, post training, reasoning and reinforcement learning, agents, multimodal perception/vision, robotics, audio/speech, safety and alignment, or scientific AI (health, materials, fusion).
Track record of bridging research breakthroughs with product engineering or external scientific/policy stakeholders.
Sustained record of first-author and senior-author scientific publications in premier peer-reviewed conferences (e.g., NeurIPS, ICML, ICLR, CVPR, SIGGRAPH, CoRL) or journals (e.g., Science, Nature, PNAS).

Responsibilities

Define the architectural strategy for placing safety interventions across multimodal foundation models, balancing shared approaches with bespoke solutions.
Advance generalizable post-training methods across media types so models remain aligned as prompting and conditioning capabilities emerge.
Shape intrinsic model behavior to refuse severe harms while fulfilling borderline requests with high generation quality and controllability.
Partner with cross-functional teams to distill serving-time classifiers into training rewards and expand the safety and quality trade-off frontier.
Manage a team of 10 or more researchers and engineers, modeling scientific excellence and building a collaborative culture.