<p>Snapshot</p> <p>We are looking for a research scientist who will drive research in Gemini for information tasks. The candidate will primarily work on post-training, but could potentially also work on model-external interventions.</p> <p>About Us</p> <p>Our team works on improving Gemini on tasks where users interact with the model to complete information journeys; this includes improving helpfulness and factuality of Gemini models. To this end, we have developed novel post-training innovations to improve the quality, groundedness and factuality of Gemini models in search on mode. Our work impacts product surfaces including AI Mode, Gemini App, AI Studio and Vertex AI.</p> <p>The Role</p> <p>In this role, we expect the candidate to work on improving Gemini for information tasks, focusing on quality of information-seeking responses (helpfulness, factuality, grounding, and other aspects). It is an opportunity to explore fundamental issues in modeling and data interventions for information-seeking scenarios, with very significant opportunities in shaping Google's products in this space.</p> <p>Key responsibilities:</p> <ul> <li>Research on post-training (e.g., RL and SFT) for information-seeking scenarios in Gemini </li><li>Research on novel evaluation methods for improving model quality, grounding and factuality </li><li>Research on orchestration of tool calls, and improved retrieval methods, for information-seeking scenarios </li></ul> <p>About You</p> <p>In this role you will lead working on factuality, investigate novel recipes for post training involving retrieval, particularly Google search.</p> <p>In order to set you up for success as a at Google DeepMind, we look for the following skills and experience:</p> <ul> <li>PhD in a relevant area, or an equivalent research/publication record </li><li>Number of years experience: anything from recent PhD onwards </li><li>Strong software-engineering skills in addition to a research background </li></ul> <p>In addition, the following would be an advantage: (require maximum of 5 and minimum of 3 items)</p> <ul> <li>Experience in reinforcement learning </li><li>Experience in post-training methods </li><li>Experience in LLMs for information-seeking scenarios </li></ul>