# Research Scientist, Gemini Safety

**Company**: Google DeepMind
**Location**: Mountain View, California, US
**Work arrangement**: onsite
**Experience**: senior
**Job type**: full-time
**Category**: Engineering
**Industry**: Technology
**Wikidata**: https://www.wikidata.org/wiki/Q15733006

**Apply**: https://job-boards.greenhouse.io/deepmind/jobs/7421111?utm_source=yubhub.co&utm_medium=jobs_feed&utm_campaign=apply
**Canonical**: https://yubhub.co/jobs/job_1e0f3b52-1ae

## Description

We're seeking a versatile Research Scientist to join our Gemini Safety team, responsible for advancing the safety and fairness behaviour of state-of-the-art AI models. As a key member of our team, you will apply and develop cutting-edge data and algorithmic solutions to ensure Gemini models are safe, maximally helpful, and work for everyone.

Key responsibilities include:

* Post-training/instruction tuning state-of-the-art language models, focusing on text-to-text, image/video/audio-to-text modalities and agentic capabilities
* Exploring data, reasoning, and algorithmic solutions to ensure Gemini models are safe and work for everyone
* Improving Gemini's adversarial robustness, with a focus on high-stakes abuse risks
* Designing and maintaining high-quality evaluation protocols to assess model behaviour gaps and headroom related to safety and fairness
* Developing and executing experimental plans to address known gaps or construct entirely new capabilities

To succeed in this role, you should have a PhD in Computer Science or a related field, significant LLM post-training experience, and a track record of publications at top conferences. Experience in reward modelling and reinforcement learning for LLMs instruction tuning, long-range reinforcement learning, safety, fairness, and alignment is an advantage.

## Skills

### Required
- PhD in Computer Science or a related field
- Significant LLM post-training experience
- Post-training/instruction tuning state-of-the-art language models
- Exploring data, reasoning, and algorithmic solutions
- Improving Gemini's adversarial robustness

### Nice to have
- Reward modelling and reinforcement learning for LLMs instruction tuning
- Long-range reinforcement learning
- Safety, fairness, and alignment

---

Source: [Apply at job-boards.greenhouse.io](https://job-boards.greenhouse.io/deepmind/jobs/7421111?utm_source=yubhub.co&utm_medium=jobs_feed&utm_campaign=apply)