Research Scientist: Agentic User Signals & Reward Modeling

Google LLC; Zürich; On-site; CHF 180,000 - 260,000; Full time

Job summary

Google DeepMind's Gemini User Signals team is seeking a Research Scientist to drive post-training strategy for our advanced models. You will convert complex user interactions into training signals, focusing on reward modeling, preference optimization, and data-centric strategies to align model behavior with user needs.

You will help define scalable, cross‑modality recipes that apply across model sizes, supporting safe, ethical deployment and measurable improvements in online user satisfaction.

Qualifications

• PhD in Machine Learning, NLP, or related field (or equivalent practical experience).
• Experience with Python and one or more deep learning frameworks (PyTorch or JAX).
• Research experience in post-training, RL, agentic RL, preference learning, or reward modeling.

Responsibilities

• Design and implement mechanisms to harvest, filter, and utilize high-quality user signals to iteratively improve Gemini models.
• Develop and refine reward models and critique models that capture user intent and satisfaction for RLHF and alignment.
• Create and validate horizontal post-training recipes across model sizes and modalities (text and audio).
• Investigate methods for weighting, synthesizing, and selecting data to maximize signal impact on performance.
• Establish evaluation frameworks linking offline metrics to online user satisfaction.

Skills

Python programming

Deep learning

Research agility

Education

PhD in ML/NLP or related field

Tools

PyTorch

JAX

Job description

Google DeepMind's Gemini User Signals team is seeking a Research Scientist to drive post-training strategy for our advanced models. You will convert complex user interactions into training signals, focusing on reward modeling, preference optimization, and data-centric strategies to align model behavior with user needs.

You will help define scalable, cross‑modality recipes that apply across model sizes, supporting safe, ethical deployment and measurable improvements in online user satisfaction.