New: upload your resume and get matched to AI jobs. Free to start →

Research Scientist, User Signals from Agentic Logs

Zürich, Switzerland • Full-time

Posted 1mo ago

Remote Work Policy

On-site

Employment Type

Full-time

Categories

LLM Engineer

About the job

As a Research Scientist on the Gemini User Signals team, you will play a crucial role in shaping the post-training strategy for advanced models. Your work will focus on creating a feedback loop where complex user interactions and feedback are transformed into training signals to enhance model behavior. This involves leading research in reward modeling, preference optimization, and data-centric approaches to connect user deployment with model training. You will also contribute to defining scalable methods for applying these insights across various modalities and model families. Google DeepMind is a leading AI lab dedicated to advancing AI for global challenges and product innovation, prioritizing safety and ethics.

Responsibilities

  • Design and implement mechanisms to collect, filter, and use high-quality user signals to improve Gemini models.
  • Develop and refine reward and critique models to accurately reflect user intent and satisfaction for alignment techniques like RLHF.
  • Create and validate 'horizontal' post-training methods applicable across different model sizes and modalities.
  • Investigate novel methods for data weighting, synthesis, and selection to maximize user signal impact on model performance.
  • Establish rigorous evaluation frameworks to measure the correlation between offline metrics and online user satisfaction.

Requirements

  • PhD in Machine Learning, NLP, or a related field, or equivalent practical experience.
  • Experience with Python and deep learning frameworks like PyTorch or JAX.
  • Research experience in post-training, reinforcement learning, agentic RL, preference learning, or reward modeling.
  • First-author or co-author publication in a top ML/NLP conference (e.g., NeurIPS, ICML, ICLR, ACL, EMNLP).

About Google DeepMind