About the Role
We are looking for Research Scientists to join the Autonomous Agents team, and produce research in the development of next-generation technologies to power increasingly open-ended autonomous agents which strive to assist, support, and supplement humans in their daily personal and professional lives. Specifically, we are looking for researchers interested in topics along the lines of: Enhancing the fast online adaptation of state-of-the-art models, enabling them to efficiently compound knowledge at inference in agentic tasks, and learning from experience; and/or the continual consolidation of experience into model parameters, using methods such as distillation and reinforcement learning; and/or the development of self-improvement methods which are robust to degeneracies such as model collapse or reward hacking. Within the team, Research Scientists are encouraged to lead/support a research agenda aimed at producing practically applicable technological advances in the ability of increasingly autonomous agents to assist, support, and empower humans. The expectation is that research scientists will conduct novel research according to ambitious long-term agendas, while maintaining a strong focus on methods and tools offering practical benefits in the short term as a form of pragmatic grounding. Central to this process is the idea that rapid iteration over and refinement of solutions catering to real-world use-cases provides a strong basis for better understanding the research boundary in a fast paced field.
Responsibilities
- Participate in the ideation and development of new use-cases and desired capabilities of human-oriented agents of any form, advised by the current state of research.
- Partner with research engineers to develop ambitious prototypes pertaining to desired or anticipated agent use-cases, and design and implement evaluation protocols around these prototypes.
- Identify, motivated by empirical study of existing methods' failures or limitations on use-case-based evaluations, the roadblocks and research challenges, and...
- ...develop novel technical or methodological solutions to overcome such obstacles and limitations.
- Identify sources of data, design and implement data collection processes (supported by research engineering partners), and conduct human annotation and evaluation campaigns for the production and evaluation of strong baselines for each use-case.
- Help identify, within Google DeepMind’s broad portfolio of research projects, methods which could be adapted or tried against our evaluations, as well as teams and individuals with which the team could partner to overcome challenges whilst providing grounding and evaluation for that collaborator's research agenda.
Requirements
- A PhD in a technical field or equivalent practical experience. This specific role is targeting recent graduates, and the ideal candidate will be willing to work closely with one or more senior researchers on established high-value projects.
- Experience in a research domain connected to the production of increasingly autonomous human-oriented agents, (e.g. LLM-powered agents, RL/IL, applications in NLP, evaluation design).
- A desire to produce the next generation of agentic systems capable of learning from and efficiently adapting to deployment in real-world scenarios.
- A strong technical background in RL, Imitation Learning, Distillation, and working with/designing environments.
- Experience with In-Context Learning, Continual Learning (either in the context of RL or LLM)
Qualifications
- Strong end-to-end system building and prototyping skills.
- Experience with one or more of: fine-tuning LLMs, running human data collection/annotation campaigns, self-play, multi-agent systems, meta-learning, meta-RL, and/or skill-discovery.
- Experience with open-ended learning, RL, and frontier methods for training LLMs (RLVR, RLHF, RLAIF, multi-turn RL, multi-agent interactions, reward function design and modelling, etc.).
- A curiosity about, or experience with research topics surrounding personalization, memory, reasoning, self-improvement, and safety.
- Experience with designing and evaluating agentic tasks.
Benefits
bonus + equity + benefits