Research Scientist, User Signals from Agentic Logs, DeepMind

As a part of the Gemini User Signals team, a Research Scientist will help drive the post-training strategy of our most advanced models, dedicated to establishing the "flywheel" of alignment, converting complex user interactions and feedback into training signals that fundamentally improve model behavior. In this role, you will lead research into reward modeling, preference optimization, and data-centric strategies to close the loop between user deployment and model training. You will help define the horizontal recipes that allow us to scale these insights across different modalities and model families.Artificial intelligence will be one of humanity’s most transformative inventions. At Google DeepMind, we are a pioneering AI lab with exceptional interdisciplinary teams focused on advancing AI development to solve complex global challenges and accelerate high-quality product innovation for billions of users. We use our technologies for widespread public benefit and scientific discovery, ensuring safety and ethics are always our highest priority.

We are pushing the boundaries across multiple domains. Our global teams offer varied learning opportunities and career pathways for those driven to achieve exceptional results through collective effort.

Design and implement mechanisms to harvest, filter, and utilize high-quality user signals (implicit and explicit) to iteratively improve Gemini models.

Develop and refine reward models and critique models that accurately capture user intent and satisfaction, serving as the foundation for Reinforcement Learning from Human Feedback (RLHF) and other alignment techniques.

Create and validate "horizontal" post-training recipes that can be robustly applied across different model sizes and modalities (including text and audio).

Investigate novel methods for weighting, synthesizing, and selecting data to maximize the impact of user signals on model performance.

Establish rigorous evaluation frameworks to measure the correlation between offline metrics (reward model accuracy) and online user satisfaction.

Minimum qualifications:

PhD in Machine Learning, natural language processing, or a related field (or equivalent practical experience).

Experience with software development using Python and one or more deep learning frameworks (e.g., PyTorch or JAX).

Research experience in post-training, reinforcement learning (RL), agentic RL, preference learning, or reward modeling.

First-author (or co-author) publication in one or more peer-reviewed machine learning or natural language processing (NLP) conferences (e.g., NeurIPS, ICML, ICLR, ACL, EMNLP).

Preferred qualifications:

First-author publications at top ML/NLP venues (NeurIPS, ICML, ICLR, ACL, EMNLP).

Experience with modern alignment and post-training methods (e.g., RLHF, DPO, PPO, rejection sampling, or self-play).

Experience modeling multi-turn interaction data, agent trajectories/logs, or user feedback (implicit and explicit signals).

Experience training or evaluating reward models, critique models, or automated evaluators (e.g., LLM-as-a-judge).

Experience in agentic post-training.